1957 12_#12 12 #12

1957-12_#12 1957-12_%2312

User Manual: 1957-12_#12

Open the PDF directly: View PDF .
Page Count: 260

Download
Open PDF In Browser	View PDF

Proceedings of the

EASTERN JOINT COMPUTER CONFERENCE

December 9-i3, i957

Washington, D.C.

THEME: COMPUTERS WITH DEADLINES TO MEET

SPONSORS:
THE INSTITUTE OF RADIO ENGINEERS
Professional Group on Electronic Computers

THE ASSOCIATION FOR COMPUTING MACHINERY
THE AMERICAN INSTITUTE OF ELECTRICAL ENGINEERS
Committee on Computing Devices

Printed in the United States of America

Price

$3.oe

PROCEEDINGS OF THE
EASTERN JOINT COMPUTER CONFERENCE

PAPERS AND DISCUSSIONS PRESENTED AT THE
JOINT IRE-ACM-AIEE COMPUTER CONFERENCE
WASHINGTON, D.C.

DECEMBER 9-13, 1957

THEME: COMPUTERS WITH DEADLINES TO MEET

SPONSORS
The Institute of Radio Engineers Professional Group on Electronic Computers
The Association for Computing Machinery
The American Institute of Electrical Engineers Committee on Computing Devices

Published by

THE INSTITUTE OF RADIO ENGINEERS,
1 East 79 Street, New York 21, N.Y.

~NC.

ADDITIONAL COPIES '
Additional copies may be purchased from the sponsoring
societies listed below at $3.00 per copy. Checks should be
made payable to anyone of the following:
INSTITUTE OF RADIO ENGINEERS
I East 79 Street, New York 21, N.Y.
ASSOCIATION FOR COMPUTING MACHINERY
2 East 63 Street, New York 21, N.Y.
AMERICAN INSTITUTE OF ELECTRICAL ENGINEERS
33 West 39 Street, New York 18, N.Y.

LIBRARY OF CONGRESS CATALOG CARD NUMBER: 54.9071
Copyright © 1958

THE INSTITUTE OF RADIO ENGINEERS, INC.

NATIONAL JOINT COMPUTER COMMITTEE
M. M. ASTRAHAN, Chairman
IBM Corporation

N. H. TAYLOR, Vice-Chairman
Lincoln Laboratories, M.I.T.

IRE Representatives
DANIEL HAAGENS, Underwood Corporation

W. S. SPEER, Norden-Ketay Corporation

L.. NOFREY, Marchant Research, Inc.

N. H. TAYLOR, Lincoln Laboratories, M.I.T.
WERNER BUCHHOLZ, Ex-Officio
IBM Corporation

L. G. CUMMING, Headquarters
The Institute of Radio Engineers

ACM Representatives
S. FERNBACH, University of California

ALAN PERLlS, Carnegie Institute of Technology

GILBERT W. KING, International Telemeter Corporation

F. M. VERZUH, Massachusetts Institute of Technology
I

J. W. CARR III, Ex-Officio
University of Michigan
J. MOSHMAN, Headquarters
Council for Economic and
Industry Research, Inc.

AlEE Representatives

J. G. BRAINERD, University of Pennsylvania

H. F. MITCHELL, JR., Sperry Rand Corporation

FRED KALBACH, Burroughs Corporation

DAN C. ROSS, IBM Corporation

E. L. HARDER; Ex-Officio
Westinghouse Electric Corporation
R. S. GARDNER, Headquarters
American Institute of Electrical Engineers

:":aison with National Simulation Council
R. M. HOWE
University of Michigan

EASTERN JOINT COMPUTER CONFERENCE COMMITTEE
1957
S. N. ALEXANDER, Chairman
National Bureau of Standards

TECHNICAL PROGRAM COMMITTEE
HARRY H. GOODE, Chairman
University of Michigan
I. L. AUERBACH
Auerbach Electronics Corporation

R. M. HOWE
University of Michigan

A. A. COHEN
Remington Rand UNIVAC

E. C. JOHNSON
Bendix Aviation Corporation

J. A. HADDAD
IBM Corporation

R. E. SPRAGUE
Tel·eregister Corporation

LOCAL ARRANGEMENTS COMMITTEE
JOHN R. PROVAN, Chairman
U.S. Bureau of the Budget
W. HOWARD GAMMON, Assistant Chairman
Office of the Secretary of Defense

MARGARET R. FOX, Secretary
National Bureau of Standards
Advisory Staff

S. N. ALEXANDER
WALTER L. ANDERSON
DAVID S. BENDER
RALPH I. COLE

HOWARD T. ENGSTROM
EZRA GLASER
LOWELL H. HATIERY

EDWARD J. MAHONEY
GEORGE W. PETRIE
HAROLD K. SKRAMSTAD
MARK SWANSON

Vice-Chairmen
DELMER C. PORTS, Finance
Jansky and Bailey, Inc.

CLARKE RISLER, Hotel Arrangements
Remington Rand Division, Sperry Rand Corporation

RICHARD T. BURROUGHS, Registration
IBM Corporation

L. DAVID WHITELOCK, Exhibits Liaison

MALCOLM B. CATLIN, Publicity
Council for Economic and Industry Research, Inc.

ETHEL C. MARDEN, Women's Activities
National Bureau of Standards

U.S. Navy Department

PUBLICATIONS COMMITTEE
MORRIS RUBINOFF, Chairman
Philco Corporation

R. J. KONEFAL
Philco Corporation

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

TABLE OF CONTENTS
The Numericord Machine-Tool Director ................................................................... Gerald T. Moore
Design of a Numerical Milling Machine System .................................................. Y. C. Ho and E. C. Johnson

Logical Organization of the DIGIMATIC Computer ........ , ................................................. Jack Rosenberg

The Master Terrain Model System ~ ....................................................................... Joseph A. Stieber

A Coordinated Data-Processing System and Analog Computer to Determine Refinery-Process Operating Guides .... C. H. Taylor, Jr.

System Characteristics of a Computer Controller for Use in the Process Industries .................... W. E. Frady and M. Phister

Optimized Control through Digital Equipment .................................................................. . E. J. Otis

Real-Time Presentation of Reduced Wind-Tunnel Data ................................... M. Seamons, M. Bain, and W. Hoover

The Mechanization of Letter Mail Sorting.. . .................................................................... I. Rotkin

Preparations for Tracking Artificial Earth-Satellites at the Vanguard Computing Center ....................... . D. A. Quarles, Jr.

Use of a Digital Computer for Airborne Guidance and Navigation ...................................... S. Zadoff and J. Rattner

Some Experimentation on the Tie-In of the Human Operator to the Control Loop of an Airborne Navigational Digital Computer System
............. ' ......... :' .......................................................................... Corwin A. Bennett

Multiweapon Automatic Target and Battery Evaluator .......................... . D. E. Eisenberg, A. E. Miller, andA. B. Shajritz

Control of Automobile Traffic-A Probl,em in Real-Time Computation ............................. ~ '" ........ . D. L. Gerlough

Physical Simulation of Nuclear Reactor Power Plant Systems ......... , ............ .J. J. Stone, Jr., B. B. Gordon, and R. S. Boyd

Robert H. Kohr

An Analog-Digital Simulator for the Design and Improvement of Man-Machine Systems .. H. K. Skramstad, A. A. Ernst, and J. P. Nigro

Facilities and Instrumentation Required for Real-Time Simulation Involving System Hardware ........ " ........ . A. J. Thiberville

Application of Computers to Automobile Control and Stability Problems .................

••••••••••••••••••••••

Problems in Flight System Simulation ........................................................... '" ........ . E. J. McGlinn

100

Analog, Digital, and Combined Analog-Digital Computers for Real-Time Simulation ................... C. G. Blanyer and H. Mori

104

The Place of Self-Repairing Facilities in Computers with Deadlines to Meet ....................................... , . Louis Fein

111

Organizing a Network of Computers to Meet Deadlines ................ . A. L. Leiner, W. A. Notz, J. L. Smith, and A. Weinberger

115

A Program-Controlled Program Interruption System ........................................................ F. P. Brooks, Jr.

128

A Transistor-Circuit Chassis for High Reliability in Missile-Guidance Systems ................................... G. A. Raymond

132

A Method of Coupling a Small Computer to Input-Output Devices without Extensive Buffers .................. James H. Randall

136

The Synthesis of Computer-Limited Sampled-Data Simulation and Filtering Systems ......................... Arthur S. Robinson

139

SAGE--:.-A Data-Processing System for Air Defense ............................. . R. R. Everett, C. A. Zraket, and H. D. Benington

148

AN/FST-2 Radar-Processing Equipment for SAGE ..................... W. A. Ogletree, H. W. Taylor, E. W. Veitch, and J. Wylen

156

Operation of the SAGE Duplex Computers ................................ '" ....... . P. R. Vance, L. G. Dooley, and C. E. Diss

160

A Digital System for Position Determination ................................................................... Dan C. Ross

164

Real-Time Data Processing for CAA Air-Traffic Control ...................................................... G. E. Fenimore

.169

Design Techniques for Multiple Interconnected On-Line Data Processors ............................. F. J. Gaffney and S. Levine

172

Reservations Communications Utilizing a General Purpose Digital Computer .................................... . R. A. McAvoy

178

Stock Transaction Records on the Datatron 205 ............................................................... . A. H. Payne

183

A Small, Low-Cost Business Computer ................................................................... . Alex B. Churchill

187

A Self-Checking System for High-Speed Transmission of Magnetic-Tape Digital Data ................................ E. J. Casey

190

Communication between Remotely Located Digital Computers ................................. G. F. Grondin and F. P. Forbath

194

Communication Switching Systems as Real-Time Computers ..................................................... . A. E. Joel

197

An Introduction to the Bell System's First Electronic Switching Office ........................................ R. W. Ketchledge

204

Traffic Aspects of Communications Switching Systems .......................................... " ............. Joseph A. Bader

208

The Use of the IBM 704 in the Simulation of Speech-Recognition Systems ........................................ G. L. Shultz

214

An Automatic Voice Readout System ......... , ..................................... '" ......... C. W. Poppe and P. J. Suhr

219

Experiments in Processing Pictorial Information with a Digital Computer .......... R. A. Kirsch, L. Cahn, C. Ray, and G. H. Urban

221

Optical Display for Data-Handling System Output ............................................................... James Ogle

230

Devices for: Reading Handwritten Characters ................................................................. T. L. Dimond

232

Automatic Registration in High-Speed Character Sensing Equipment ....................................... . Abraham I. Tersoff

238

The National Cash Register High-Speed Magnetic Printer .......................................... , ...................... .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . J. Seehoj, M. Armstrong, G. Farley, M. Leinberger, M. Markakis, and S. Smithberg

243

On-Line Sales Recording System .............................' ............................ . J. S. Baer, A. S. Rettig, and I. Cohen

251

Organization of Simulation Councils, Inc .......................... : . . . . . . . . . . . . . . . . . . . . . . . . .. ................................

257

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

The Numericord Machine- Tool Director
GERALD T. MOOREt

HAT is numerical control of machine tools? I
would answer by saying that it is a system of
.
machine-tool control in which the machining operation is guided by instructions in the form of coded
numbers. These instructions may be inserted via punched
cards, punched tape, magnetic tape, or other suitable
means. A complete sequence of operations is predetermined and programmed in a coded form which is understandable to the controller or the director, as it is called.
As divided into their broad classifications; the two types
of numerical machine-tool controls are:

1) Positioning controls. A sequence of positions of a
tool is controlled, some operation occurring at each position before the tool continues to the next position. Here,
it is generally unimportant by which route and at which
speed the tool progresses from one position to the next.
The tool is not in contact with the workpiece when moving between positions.
2) Path controls. The tool is made to follow a prescribed path over the surface of the workpiece at a prescribed, but not necessarily constant, velocity. Depending
upon the particular control system, the path may be in
two or three dimensions.
The numericord machine-tool director, about which I
am going to talk, is a path control system. While it differs
in some respects from other path control systems, a study
of its functioning will serve to demonstrate the processes in'volved in path control.
When I speak of the machine-tool director system, I am
not including the machine tool itself with its power servomechanisms and error-detecting and amplifying circuits.
That is separate equipment. I am talking about the dataprocessing and digital-to-analog conversion equipment
which is necessary to provide real-time continuous-control signals in response to the numerical instructions inserted into the director. In the Numericord system there
is no physical interconnection between the director system
and the machine-tool controls. The continuous-control signals are recorded on magnetic tape and are subsequently
played back at the machine tool. The interposition of the
recording and playback functions in the sequence of control makes possible the divorcing of the director system
from the machine tool. Therefore, a magnetic tape may
be repeatedly used to produce several identical parts on
the machine tool. Meanwhile, the director system is recording tapes for other machine tools. Fig. 1 shows the director system.
t Concord Control, Inc., Boston, Mass.

Fig. 1-The director system.

The Numericord director has punched paper tape as its
input. Coded numbers on this tape prescribe the path of
the cutting tool center in five axes. Thus, for example,
a milling machine on which the milling head has two rotationaldegrees of freedom as well as orthogonal X, Y,
and Z degrees of freedom may be controlled. The punched
paper tape does not specify the path at all points, therefore
it is necessary for the director to interpolate between
specified points. That is, the continuous-control signals
for the five axes must direct the cutting tool along some
path between the points specified on tape.
The amount of data that is required on the input tape
for any numerical system depends upon the interpolation
method used in that system, and, in general, the amount
of data decreases with increasing complexity of the interpolator. If you were to define positions on paper tape successively at one-thousandth intervals on the workpiece, no
interpolation between defined points would be necessary
at all in order to attain a reasonable degree of accuracy.
On the other hand, if the director will interpolate linearly
between defined points, that is, if the director directs the
machine tool to cut a straight line between defined points,
then the paper-tape input need define only the end points
of all straight-line cuts. Here, however, a curve must be
defined as a series of straight-line segments, the number
of the segments depending upon the prescribed accuracy.
If, for instance, you were required to cut half of an inside
circle of six-inch diameter with a two-inch diameter cutter
maintaining an accuracy of one thousandth,: linear interpolation would require that the paper tape input specify
78 straight-line cuts. More elaborate interpolation schemes
are possible, which pass higher degree curves through a
number of specified points. Depending upon the type of
cutting to be done, these systems may reduce considerably
the amount of data. required at the director input at the
expense of a greater amount of equipment within the
director. All considerations being taken, the Numericord
designers were led to the choice of a linearly interpolating
system.

Moore: The Numericord Machine-Tool Director
The interpolator has five output lines, one for each
axis. On each of the output lines discrete "command"
pulses appear. One pulse represents a fixed increment of
displacement at the machine tool, the amount of displacement being referred to as the quantization level of the
system. The N umericord director has a quantization level
of one eighth of a thousandth of an inch. Thus, the occurrence of 8000 command pulses in succession on the
X -axis output line would drive the tool one inch over
the workpiece in the X direction. Since the power servomechanisms at the machine tool respond to analog signals
and not to pulses, a pulse-to-analog conversion must take
place. This occurs in the "decoder." The output from the
decoder consists of five command synchro signals. These
are recorded on magnetic tape. When played back at the
machine tool, the signals provide the command positions
for five servodrives.
A block diagram of the director system is shown in
Fig. 2.

serially, line by line, each line being translated into a fourdigit binary code and stepped into four magnetic-core
stepping registers. The first digit of the four-digit code is
weighted five, the second is weighted two, and the third
and fourth are each weighted one. The binary code for
seven, therefore, is the binary 1100. Each of the four
stepping registers is associated with one of the four
binary digit columns in the translated number. One register is designated the "five" register. Into it goes the most
significant binary digit. The next register is designated
the "two" register, and it receives the second most significant digit. The third and fourth registers are the "one-A"
and "one-B" registers respectively.
Table I demonstrates the coding.
TABLE I
CODING

OXS
OXS

PUNCHED
PAPER
TAPE

TAPE

5 CHAINS OF
COMMAND PULSES

5 COMMAND
SYNCHRO SIGNALS

Fig. 2-BIock diagram of a Numericord system.

DATA INPUT

The terminal point of each straight-line segment of the
programmed tool path is specified on the paper tape by
coding the distance in each axis from the terminal point of
the previous straight-line segment. In addition to the incremental distances for the five axes, a time-of-cut, or command time, is specified for each straight-line segment,
and a direction-of-cut or sign code is inserted at the beginning of each dimension. One line of tape is required
for each coded decimal digit or sign. The arrangement of
data is such that the command time appears first. Three
lines of tape are allocated to command time so that the
command time is three decimal digits long. The command
time is followed by a sign and seven decimal digit codes
for the X axis, then by a sign and seven decimal digit
codes for the Y axis, and so on for each of the five axes.
The seven decimal digits indicate hundreds of inches, tens
of inches, units, tenths, hundredths, thousandths, and
tenths of thousandths of inches. The seventh digit is either
a zero or a five so that distances are programmed in multiples of a half of a thousandth. The director will handle a
maximum distance of 399.9995 inches in all axes. Of
course, when the fourth and fifth axes are used to control rotations, a conversion must be made from angular
degrees to linear inches so that the programming can· be
done in inches.
The command-time and command-distance information
for a single straight-line cut comprise one "block" of
paper-tape information. At the director, each block is read

OXS

lXS

1X5

lX5

+
+
+
+
+
+
+
+
+
+

OX2
OX2
1X2
IX2
1X2
OX2
OX2
1X2
1X2
1X2

+
+
+
+
+
+
+
+
+
+

OX1
OX1
OX1
OX1
1X1
OX1
1X1
OX1
OX1
1X1

+
+
+
+
+
+
+
+
+
+

OX1

IX1
1X1

1X1

OX1

OXI

1X1

IX1

As each character is read, the appropriate code is set
into the shift registers and advanced one position into the
registers. When one block of tape has been read, the registers are full, and the tape reader stops. The numbers read
first are stored in the magnetic cores farthest down the
stepping registers. The stepping registers have between
them a group of four cores (one core per register) to store
each coded decimal digit of the three command-time digits;
they have a group of four cores to store each coded decimal digit of the seven command-distance. digits for each
axis, and they have a group of four cores (some of which
are redundant) to store each of the signs. So there is a
total of 12 command-time cores, 140 command-distance
cores, and 20 sign cores.
Fig. 3 shows the stepping register with the command
time, 200 seconds, and with the command distance,
-250,9645, stored as an example. Note that the arbitrary
choice was made to use the same code for minus as for
two. A limitation on the command-time code is that not
more than one command-time core contains a binary "one."
The reason for this will be seen as we progress. The allowable command times are 200, 100, 50, 20, 10, 5, 2, 1,
and 0.5 seconds.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

8
COMMAND
TIME
Is REGISTER

ITIJ.-lID..!1JCYUlIIII.J.-OIIIT]-i~1

IA REGISTER

2 REGISTER

DATA IN

5 REGISTER

o
IZI

BLOCK REPRESENTS A CORE
X IN BLOCK REPRESENTS A STORED BINARY ·ONE"

Fig. 3-Magnetic core shift register.

250.9465 INCHES

STORED AS

EXAMPLE

Fig. 4-Arrangement of interpolation counter in relation
to storage cores of a typical axis.
INTERPOLATION

Upon the occurrence of an internally generated signal,
the contents of the dimersion-storage cores of the four
stepping registers are transferred in parallel to a secpnd
set of storage cores. Simultaneously, the sign and command time cores are reset, causing pulse outputs from
cores which had been storing ones. The stepping registers
are thus freed to receive another block of information
while the coded dimension numbers of the previous block
are available to be operated upon in the second set of cores.
Let us consider of what this operation must consist. We
are attempting to convert a coded number into a corresponding number of discrete pulses at a rate of one pulse
per eighth of a thousandth, or a rate of four pulses for our
least programmabl,e distance, one half of a thousandth.
Thus, let us interrogate nondestructively the core in which
the half-thousandth bit is stored~ If this core contains a
binary "one," a pulse output will occur each time we interrogate it; if we interrogate it four times during the
processing of the block, four output pulses will occur. We
see that we can weight the binary digit stored in any core
by fixing the number of times that that core is interrogated during the programmed command time. The cores
storing binary "ones" weighted at 0.001 inch will be interrogated eight times during any command time; the cores
storing binary "ones" weighted at 0.002 inch will be interrogated 16 times during a command time, a 0.005-inch
core 40 times, and so forth. The outputs of the cores for
one axis are buffered onto a common output line so that

the command pulses on that line are a result of contributions from all the cores which have "ones" stored in them
in that axis.
It would appear that the weighting functions for the
core-stored dimension could be generated by a counting
chain, and this is just what is done. A counting chain composed of cascaded binary and decade scalers is used. Fig. 4
shows how this counting chain, called the interpolator
counter, is arranged with respect to the core storage for
one axis. The decade circuits consist of four flip-flops
connected so that, for 10 pulses entering a decade, there
occur 5 carry and 5 noncarry transitions of the first flipflop, 2 carry and 2 noncarry transitions of the second flipflop, and 1 carry and 1 non carry transition from each of
the third and fourth flip-flops. The noncarry transitions of
each flip-flop trigger an interrogate pulse which results in
a command pulse out, if a binary "one" is contained in the
magnetic core being interrogated. Although Fig. 4 shows
the cores of only one axis, each flip-flop in the interpolator
counter interrogates the five corresponding cores of the
five axes. With a divide-by-four circuit beyond the halfthousandth flip-flop, the half-thousandth core is interrogated four times for every end carry. The end carry signals the end of the straight-line motion.
It is interesting to note two properties of the interpolation counter without which this system of linear interpolation would not work.
1) No two cores of any axis are interrogated simultaneously, and hence, the command-pulse contributions
of the various cores appear as separate discrete pulses on
the command-output line. This is because each oscillator
pulse propogates down the chain as carry transitions until
the first flip-flop ready for a noncarry transition is reached.
The noncarry transition of that flip-flop does not result in
any action farther down the chain. Hence, only one noncarry transition can occur anywhere in the chain for each
input pulse to the chain.
'
2) It can be shown that, regardless of what pattern of
"ones" and "zeros" exists in the storage system, that is,
regardless of what number has been stored, the resulting
pulse distribution is such that the displacement vs time
for any axis never varies from a perfect ramp by more
than one quanta.
Our command-pulse clock-oscillator frequency in the
Numericord system is 16 kc. Referring again to Fig. 4,
we see that 3.2-million oscillator pulses are required for
each end carry. At an input rate of 16 kc, it would require
200 seconds to cycle through or cause an end carry. However, if we feed in our 16-kc clock pulses farther down the
chain, it will require less time for the counter to cycle
through. The N umericord system feeds clock pulses to nine
gates, only one of which is open at a time. So pulses are
fed to one of nine input points along the interpolator
counter. The nine command times available are, as I pre-

Moore: The Numericord Machine-Tool Director
viously mentioned, 200, 100, 50, 20, 10, 5, 2, 1, and 0.5
seconds. These gates are controlled by flip-flops, one of
which is set by the pulse output of the appropriate command-time storage core.
The end carry from the interpolation counter clears the
storage register and resets the command-time flip-flop so
that no more oscillator pulses are admitted to the counter.
When the clearing action is complete, the next command
dimension is dumped in the storage register from the
stepping register. Simultaneously, the nine command-time
storage cores are reset resulting in an output pulse from
the one core holding a "one." This pulse sets the appropriate command-time flip-flop. Clock pulses are now entered
into the counter at a point in accordance with the new command time. Comrp.and pulses continue to appear on the
output line after the end carry, but at a rate determined by
the new command time and the new command distance.
So there may be a discontinuity in command-pulse rate
and, therefore, a discontinuity in command velocity at the
end of one straight-line cut and the beginning of the next.
In programming an excessively large velocity step in
any axis, the programmer may use a special code which
will automatically reduce the clock rate as the end of the
cut leading into the velocity step is approached. So the
velocity step occurs at a much lower clock-pulse level and
consequently results in a much lower velocity step at the
machine tool. After the velocity step has passed, the clock
rate rises to normal.
You can see in Fig. 5 that when clock pulses are entered
into the counter at a point other than through the 200-second gate, not all cores will be interrogated. For instance, a
50-second command time results in the 100-inch and 200inch cores not being interrogated. Therefore, a restriction
must be placed on the programmer so that the programmed
distance for 50 seconds is not more than 99.9995 inches in
any axis. In fact, the restriction for any command time is
such. that the maximum vector component of feedrate in
each axis is two inches per second.
DECODING

Since a synchro signal is well suited to the control of position, the N umericord system was designed to produce
command synchro signals from the command-pulse outputs
of the interpolator. This is done in the electronic phaseshift decoder. This decoder produces six 200-cps square
wave outputs, one for each axis, plus one for a :.eference.
The axis signals are phase shifted with respect to the reference by an amount proportional to the command distance. The decoder output is similar to the output of a
rotary command synchro where the stator windings are
excited by two or three phase reference voltages, and the
phase of the rotor signal with respect to anyone of the
stator-phase voltages· is proportional to the mechanical
angle of the rotor.
The mechanism by which the phase shift is produced in

COMMAND
PULSE CLOCK

Fig. 5-Interpolation counter with command-time gates and
flip-flops.

CLOCK
OSCILLATOR

REFERENCE SQUARE
. WAVE

PERIOD= T

COMMAND PULSE
INPUT
'(OUTPUT OF INTERPOLATOR)
~MMAND PULSE
I

REFERENCE

OUTPUT~
I

ITT

AXIS

OUTPUT

~
Fig. 6-Sample decoder.

response to a command pulse is best understood through
reference to Fig. 6. Here, two binary-counting chains of
equal length are shown. Both have inputs from a common
high-frequency oscillator, the carrier-clock oscillator. One
of these counting chains has an additional input which is
the command-pulse line. I f pulses only appear on the clockoscillator input line, and if both counting chains initially
start with all flip-flops reset to a common state, then the
square-wave signals appearing on the plates of the last flipflop in each chain will be of the same frequency and will
be in phase with each other. Now, if a command pulse
appears on the command line at a time between the occurrence of two clock pulses, the first flip-flop of this claim will
become 180 degrees out of phase with the first flip-flop of the
other chain. One less clock pulse will be required for this
chain to cycle through a complete count than is required
for the other chain. Therefore, the transition of the last
flip-flop occurs sooner for this chain, by the amount of
time between clock pulses. If no more command pulses
appear, the two chains continue counting, and the phase of
the square-wave signal of the last flip-flop in the second
chain remains advanced by T microseconds with respect
to the first chain. ( T = clock-pulse period.) The first
chain is the reference chain, and the square-wave signal
from the plate of its last flip-flop is the reference-output

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

signal and is essentially the signal which excites the stator
of the feedback synchro. The second chain is the axis
chain and its output is the phase-shifted signal to be compared with the phase of the rotor signal on the feedback
synchro. The reference signal must, of course, be filtered
to a sine wave and converted to 2 or 3 phase.
When the direction of machine-tool travel is negative,
the command pulses are made to delete incoming clock
pulses, one for each command pulse. So the phase of the
axis counter lags the reference by further fixed increments
with. every command pulse.
In the N umericord system, there are five counting
chains in addition to the reference chain so that five motions may be simultaneously controlled. Each chain is a
combination of cascaded binary and decade scalers so that
the total reduction of frequency is by a factor of 800 in
each axis. The clock frequency is 160 kc and, therefore,
the nominal output frequency is 200 cps. Since command
pulses may be entered (added or subtracted) at the highfrequency end of the scalers at a rate of up to 16,000 pps,
it is possible to modulate the output phase at a rate of
7200 degrees per second, or, in other words, to modulate the
frequency at a rate of 20 cps. Each command pulse shifts the
phase by one 800th of a cycle or by 0.45 degrees. Eighthundred 'pulses or 0.1 inch of command causes the phase to
:shift one cycle.
.The carrier-clock oscillator and command-clock oscillator are not synchronized. It could happen that a command
pulse and a carrier-clock pulse could appear simultaneously at the input to a decoder axis counting chain if the
precaution were not taken to avoid this. A circuit which
we call the chronizer circuit prevents this from happening.
In Fig. 7, we see that a command pulse sets a flip-flop to
the "one" state, which after a short delay, opens a gate.
The next carrier-clock pulse that occurs passes through
this gate and resets the flip-flop. A pulse is produced at a
fixed interval after the reset transition of the flip-flop. The
time that this pulse can occur with respect to the time that

carrier-clock pulses occur is determined by the setting of
the delay. The delay is adjusted so that the pulse occurs between two carrier-clock pulses. The pulse will either be
added to the carrier-pulse chain entering the axis counter,
or it will generate a gating potential of sufficient length to
prevent the next carrier pulse from entering the counter,
depending on the state of the add-subtract flip-flop.
TO

r==---..,--------~~;Ti:GN~~IN

TO AXIS
COUNTING CHAIN

Fig.7-Chronizer.

I have attempted to explain the operation of the essential
feed-forward elements of the Numericord system. There
are many auxiliary features, an explanation of which
time does not permit. There is an indication scheme by
which a continuous decimal-digit display of the actual
phase between any axis and the reference is presented to
the operator. There are area alarms which point to specific
areas in the equipment when a fault occurs.
You can appreciate that the director is a very' special
purpose type of computer, and I think you will appreciate
that much computation may be necessary in the initial
paper-tape preparation. These computations include determining tool-center offsets, since it is the contour of the
point of tangency between the tool and the workpiece that
is of interest, whereas it is the path of the. center of the
tool that must be programmed. The computations also include determining the straight-line segments necessary to
approximate a specified curve with a given degree of accuracy. A general-purpose computer lends itself to these
computations, while the real-time problem of interpolating
and rate generating is the special province of the director.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Design of a Numerical Milling Machine System
Y.

Hot

AND

GENERAL CONSIDERATIONS

A. Basic Concept of Numerical Control

A
n.

DV ANCED data-processing and cantral techniques
can be used in many ways to. effect substantial
impravements in present manufacturing pracesses. A gaad example is the applicatian af numerical cantral to. machine taals. As currently used in this cannectian,
numerical cantral describes laasely the cancept af aperating machine taals fram infarmatian recarded an punched
cards, punched tape, ar magnetic tape. The recarded informatian may ar may nat be in digital farm. If it is nat,
hawever, the cantral record is generally praduced from
numerical data by equipment which is cansidered to. be
part af the system. Hence the use af the ward "numerical."
The underlying abjective in applying numerical cantral
to. machine taals is impravement in the aver-all pracess af
praducing finished parts fram basic design infarmatian.
Impravements cammanly saught include greater accuracy
and reproducibility of the part, increased machine praductivity, reductian in taaling casts, reductian in skilled manpawer required, and aver-all shartening af the manufacturing cycle. Sizable gains can be realized thraugh attentian to any af several specific prablem areas. Hawever,
maximum benefits are to. be expected anly if the entire
manufacturing pracess is cans ide red as an integrated
system.
The praper starting paint far such an appraach may
well be in the design stages which immediately precede
manufacturing. Actually, seriaus thaught has been given
to. the use af madern data-pracessing techniques in mechanizing the design pracess itself.l Hawever, little pragress
has been reparted to. date except where the design pracess
is at least partially analytical already, as with certain types
af cams. For the mast part, therefare, in cantemporary numerical machine-taol systems it is presumed that a mare
or less canventianal engineering drawing of the part to
be made is available.
A further stage af manual effart, referred to. here as
pracess planning, likewise appears in present systems.
This is cancerned with the develapment af data pertaining
to. the metal-cutting aspects af the jab: the manner in
which the part is to. be maunted an the machine, the sequence af cuts to. be made, the am aunt af metal to. be remaved in each cut, the cutter size and shape, the feed rates,
etc. Since mast af this infarmatian is derived fram the
t Bendix Aviation Corp., Detroit, Mich.
G. R. Price, "How to speed up invention," Fortune) p. 150;
November, 1956.
1

E. C. JOHNSONt

part drawing, utilizing past experience, it might be expected that it wauld be pas sible to mechanize this task.
Practically, hawever, mechanizatian has nat yet been
faund feasible because af the large number af decisians to.
be made and the difficulty af defining suitable criteria. The
input to. a numerical manufacturing system praper therefare cansists af twa basic kinds af infarmatian: geametrical data and machining instructions.
At the output end af the system is a machine for physically praducing the part. The result mast ideally is a part
an which all machining aperations are camplete. It may
even be desirable to. cansider that automatic inspectian af
the part is included as well. As a practical matter, hawever,
it may be necessary to. accept far less, recagnizing that same
finishing aperatians may be required either by hand ar by
machines essentially unrelated to. the numerical pracess.
Likewise, inspection may be entirely separate. With the
latter reservatians in mind, a numerical manufacturing
system af the type being discussed can be represented symbalically as in Fig. 1.

NUMER1CALMA~UFACTU"'I~SYSTEM

r---------T---------r---------,
I
I
I
I

I
I
I
I
I

"mM

moc

~
I

I
I
I
I

________

I _________ L
t

________

PART

Fig. I-Symbolic representation of a numerical
manufacturing system.

B. Path Control of a Milling Machine
In same machining techniques such as drilling and
braaching, the shape af the cutter determines the shape af
the wark piece. Cantral af such machines is largely a matter af pasitianing the cutter to. the praper lacatian between
operatians. Milling and turning, hawever, are techniques
in which the shape of the part surface may have little or
nathing to. do with the shape af the cutter. The surface is
determined rather by the relative motian between cutter
and wark piece.
The amount af cantral infarmatian necessary in such a
pracess depends nat anly an required accuracy in the usual
sense, but also on surface finish. Generally speaking, surface finish has to do. with unifarmity in relatively small
areas; accuracy has to. do. with absalute lacatian. Irregularities permitted by the surface-finish requirement are frequently a factor af ten ar mare smaller than the errars
which are talerated fram accuracy cansideratians alane.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Hence basic resolution provided in the motions of the
machine must usually be many times better than implied
by the accuracy requirement.
Coarse control information would be filtered to some extent by unavoidable lags in the machine drives. Such
DIRECTION
smoothing might be completely satisfactory at high velocities, or feed rates, but completely ineffective at low rates.
In addition, the dynamic performance of the machine
drives should be high for other reasons-to maintain low
following errors on complex contours at reasonable speeds,
and to resist load forces generated by the cutter over a
wide-frequency range. Consequently dynamic filtering of
coarse control information by the machine drives cannot
generally be relied upon to provide the necessary smoothFig. 2-Scanning nature of three-dimensional contour milling.
ness of motion.
Another factor tending to increase the amount of information required in the control of a milling machine is
the relative inefficiency of the scanning procedure used
conventionally to generate such shapes as the one illustrated in Fig. 2. Little if any advantage can be taken of
the similarity of cutter motions. on successive passes. To
(a)
(b)
make matters worse, the spacing of the passes may be dicFig. 3-Illustration of two major computational problems. (a) Cuttated by the need to keep the scallops produced by the cutter-center offset. (b) Interpolation or path generation.
ter within the limits allowed by a tight surface-finish retechniques, practically has not yet been found to be so.
quirement.
The problem is somewhat analogous to that encountered However, several other problems of a computational nain the production of a television image which is pleasing ture definitely are subject to mechanization.
One of the most significant of these is the cutter-offset
to the eye. The large amount of redundancy between adjacent lines in a frame is recognized. Theoretical consid- problem. As- previously noted, complex surfaces are generations clearly indicate that substantial improvement is erated in milling by relative motion between cutter and
work piece. In the usual case, it is the cutter axis or center
possible.
However, techniques which would avoid this redundancy that is directly controlled. The surface of the work piece,
and still be practical on an economic basis are yet to be however, is produced by the periphery of the cutter, Fig.
3 (a). A translation from part-design information is theredemonstrated.
Thus the amount of information required to control a fore necessary to determine a cutter-center path which will
milling machine, using conventional cutters and procedures 'produce the desired surface.
Another major problem is that of path generation, or
and with no built-in smoothing other than that provided by
the usual dynamic lags on the drives, is quite high. The interpolation. Fig. 3 (b) illustrates a very common situarate of information flow depends on machine feed rates. tion in which a portion of the part to be made is a circulp,r
These may range from fractions of an inch per minute to arc. End points are located and the radius is specified.
several hundred inches per minute. A theoretical upper Somewhere in the manufacturing process there must be
limit to the usable information rate might be established in the ability to establish from this kind of information an
terms of the dynamic
capability of the machine drives. essentially continuous sequence of cutter-center positions
\
There would seem to be little point, for example, in chang- sufficient to generate the curve.
Although other problems such as optimum control of
ing the position command at a rate appreciably higher
than that at which the machine drive servos can respond. feed rate can be considered, the cutter-offset and interpolaLarger steps at a less frequent rate may give just as satis- tion problems seem to be the most fundamental. In genfactory results. Taking full advantage of this principl~ eral, either of the problems can be difficult and tedious to
may involve more expense in terms of additional hard- handle manually. The economic success of numerical maware than could ordinarily be justified. The "pulse~multi chining appears to hinge on the development of effective
plication" scheme described in Section III approximates techniques for dealing with these problems automatically.
this, however.

~~::=§:~===rl'

D. System Organization

C. Computational Problems
As mentioned earlier, the process-planning stage, while
perhaps theoretically amenable to modern data-processing

A numerical manufacturing system can be divided into
any number of physically distinct functional elements,
provided suitable means are available for communication

Ho and Johnson: Design of a Numerical Milling Machine System
between elements. The last element in the system, as presently conceived, is a machine tool of more or less conventional configuration (assuming that there is no radical
departure in metal-removal techniques). A control element of some description will necessarily be direct-coupled
to the machine. This element may be a full-scale computercontrol device for performing all data-processing operations at the machine, or it may be little more than a datareceiving device. Between the machine control unit and
the input end of the system, many variations are possible.
Although computing techniques .are common at present,
a basic incompatibility exists between the instantaneous
time scales required for some calculations, such as cutter
offset, and the rate of control information required by the
machine to achieve uniform and continuous operation. A
computer which is fast enough to meet the peak information rate on a real-time basis would be far more powerful
than necessary on the average. Isolation of the difficult
computational tasks from machining proper is therefore
advisable. This implies that there should be at least one
element in addition to the machine and its control unit.
Other principles can be formulated from the data-handling viewpo[nt which under some circumstances can serve
as useful guides. For example, consider the two computational problems discussed in the preceding part. That
of cutter offset generally does not produce additional information; the cutter-center surface is not usually significantly. more complex than the part surface. The process of
interpolation, on the other hand, is basically one of generating additional data. Other things being equal, therefore,
cutter offset should be performed before interpolation in
order to avoid having to transmit and operate on an unnecessarily large volume of data.
Other considerations lead to basic conflicts. It may be
feasible to perform the purely data-processing operations
at a rate which is significantly faster on the average than
the rate at which metal can be removed. A single dataprocessing facility could then service a number of machines. In this case, all possible data processing should be
done away from the machine and its directly connected
control· unit. This approach is attractive from the viewpoint of simplifying the equipment required at the machine. However, as emphasized previously, the control of
milling-machine motions requires a very large amount of
data. In the approach just described, all of this informatiOI~ would be transmitted to the machine, none of it being
generated in the machine control unit. The result may be
extremely bulky control records and expensive and complex equipment for their generation.

E. A Solution to the Systems Problem
From a functional point of view, present numericalcontrol systems differ mainly in the manner in which they
are subdivided physically, the operations assigned to each
element, and the data-transmission links between elements.
Because of the numerous conflicting considerations, obvi-

Fig. 4-Pictorial diagram of the Bendix numerical milling
machine system.
z

/
STRAIGHT -LINE
APPROXIMATION
CALCULATED
POINTS

DESIRED
CUTTER-CENTER
PATH

~------------------------------~ y

x
Fig. 5-Straight-line approximation to a curve.

ously no system can be optimum in an absolute sense. The
system illustrated pictorially in Fig. 4, however, meets to
a high degree the over-all objectives outlined in Section
I-A.
This system comprises two major groups of equipment.
One consists of the machine and its directly connected
control unit. The other involves a small, general-purpose,
digital computer for data preparation. The basic conflict
between the required large volume of control information
and extensive computation at the machine is resolved by
building into the machine control unit the ability to perform an elementary straight-line curve-fitting or interpolating operation. With coordinate differences provided
along an arbitrary cutter-center path, as indicated in Fig.
5, the machine control unit in effect generates the connecting straight lines.
The remaining computation--compensation for cutter
geometry and gross interpolation of curves and surfaces-is performed by the data-preparation equipment. Linear
interpolation at the machine is relatively economical, is
compatible with the peak rates of control information de-

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Fig. 7-Flexowriter-Verifier for preparation of process tapes.
Fig. 6-Computer, magnetic-tape unit, and input-output unit.

manded by the machine, and, most important, very greatly
reduces the amount of information which must be transmitted to the machine control unit by way of the control
record. The volume of information is reduced to the point
where punched tape of conventional configuration is feasible as the control record. The system is digital throughout up to the point of producing operating signals for the
machine drives.

II.

DATA PREPARATION

The heart of the data-preparation portion of the system
is the Bendix G-15D general-purpose computer. An automatic programming system called COMPAC (COMprehensive Program for Automatic Control) enables the computer to accept raw dimensional data and machining instructions in a language familiar to the process planner.
This section gives a brief description of the equipment
and discusses the design of the automatic program.

A. Equipment
1) Computer: The Bendix G-15D (Fig. 6) is a mediumspeed computer, operating serially with a magnetic drum
of 2160-words capacity.2 Two noteworthy features make it
particularly suitable for this application. First, the G-15D
instruction has a micro-programming structure consisting
of seven independent parts. The effectively two-address
nature of the instruction facilitates minimum-access coding to achieve maximum computing speed. Second, input
and output can proceed simultaneously with computatiOn
through built-in buffer-storage registers. Both ~eatures
were exploited in the design of the program.
2) Au%iliary Equipment: Three pieces of auxiliary
equipment supplement the computer in the data-prepara
tion system. They are a Friden Flexowriter-Verifier, a
special input-output unit, and a magnetic-tape unit.
2 H. D. Huskey and D. C. Evans, "The Bendix G-15 general
purpose computer," Proc. WESCON Computer Sessions, pp. 87-91;
August, 1954.

Information describing the part and its machining
process is introduced into the computer by way of a
punched tape (the process tape). The production of this
tape is accomplished by the Flexowriter, an electric typewriter equipped with a tape reader and a tape punch (Fig.
7). Input data as furnished by the process planner on a
handwritten manuscript, or process sheet is copied on the
Flexowriter. As a by-product of this typing, a punched
tape is produced which contains in coded form the process
information. An additional tape reader, shown in the
background of Fig. 7, acts as a verifier to check the correctness of the tape in a separate typing operation. The
tape then serves as direct input to the computing system.
The Flexowriter may also be used to duplicate process or
control tapes.
The input-output unit (Bendix AN-2) consists of a
punched-tape reader, a tape punch, and a control desk
housing some electronic circuitry. The reader accepts process tape in standard Flexowriter code and translates it to
straight binary. The punch produces control tape directly
as output from the computer. The special format used on
the control tape may also be read back into the computer
by means of the input-output unit.
The magnetic-tape unit (Bendix MTA-2) supplements
the internal memory of the computer and provides permanent storage for all computer programs. Both the inputoutput unit and the magnetic-tape unit are shown in Fig. 6
with the computer.

B. Program Design
The basic objective of COMPAC is the translation of
engineering information which can be interpreted by human beings into information of a form which can be
recognized by the machine control unit. Two essential
functions involved in this process are coding and computing. With respect to the former, the problem in
program design is to determine the type of Ilumericalcontrol instructions to be built into the program and the
format of the input data. Here, convenience of use from

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

IN.

-00/.5000

seTUP POINT

CLEARANCE
Pl..lNE

IN.

000.0000

IN.

001.2500

000.'500

END POINT OF SECTION

INITIAL
CLEARANCE

IN.

DATE
TOLERANCE

9-6-57

IN.

000.0002
CIRCLE
RADIUS
IN.

TAPE NO
SHEET

PLANNER

Y.C. HO

;!
~

E. I.A. PART

PAin NO.

:!i"

FEED
RATE
IN.lMIN.

/
TOOL
DIM!.
IN.

I
FINAL
OJT

iS~

~~
41

r--o,-

PRINT
20 (TAB)

of' (cr)
030.0000

(2 ) 000.0000
(S) 011.0000
(4} 011.0000

(5'.')

oos.oooo

(6) 004.0000

003.0000
003.0000
00'1.5000
004.5000

000.7!S00
00('.1400
000.7500
000.7500

003.0000

000.7600

-004.0000

-z

0
0
0

,1
/

2
2

/
0

004.5000

40
05

00
00
00

ao
ao
030.0000

-001.0000

40
000.2'00

000.1:500
OO"'.~OOO 000.7500
(9) '00'1.0000 00".5000 000.1500
(/0!J 00"'.5000 004.0000 000. 7~OO
(II) -001.$'000 004.$'000 00/.2500
f2~ -001. 5000 000.0000 000.70$00
(7) 009.5000

00".0000

(8~ 010.0000

001.6000
000.5000 -000.5000
001.6000 -003.0000
000.5000 -000.5000

0 -0 -2
a 2
1
I -0 ,2
1
0
2
0
a 0
0
0
0

03 04

00
00
00

00
00

END

,P ...." • •

~ ~

Fig. lO-Format of the Bendix COMP AC-I process sheet.

(a)

(b)
Fig. ll-Part to which the process sheet shown in Fig. 10 applies.
(a) Profiling outside of D. (b) Pocket milling inside of D.

discussion of the processing for this part has been presented elsewhere. 3
The design of this program provides a highly systematic method of introducing the part design data. The
process planner needs to know little of actual computer
programming beyond the functions performed by each of
the eight operation codes. A somewhat different approach
includes a greater number of more basic operation codes.
Although potentially more flexible, such an approach appears to demand more of. the process planner in an area
at present unfamiliar to him. At the expense of some
flexibility, the COl\1PAC programming system tends to
offer easier and more inclusive instruction codes in preference to codes of a microprogramming nature.
The choice of entering data together with operation
codes is prompted by somewhat similar considerations.
Valuable storage space is saved when only a small amount
of information is stored in the computer at one time.
Although this necessitates the occasional repetition of
identical data points in the same program, it eliminates the
possibilities of error by the process planner in attempting
to identify specific points with numerical-control instructions not entered simultaneously.
An over-all flow diagram of the COMPAC system is
shown in Fig. 12.
2) Computation Techniques: The preceding describes
the functional aspects of the program. In this part the
mechanization and execution of these functions are discussed. The four major steps in this computation process
are outlined below.
3 E. C. Johnson, "Bendix tape preparation system," Proc. EIA
S3!mp. on Nume1,ical Control, p. 63; September, 1957.

lio and Johnson: Design of a Numerical Milling Machine System
the process planner's viewpoint is of prime importance.
In computing, on the other hand, the problem is to make
most effective use of the computer, considering such characteristics as storage capacity and computing speed. These
two problems are invariably conflicting and require certain
compromises. Over-all design of the program is discussed
from both viewpoints.
1) Instruction Codes and the Process Sheet: The form
of the instruction code is shown in Fig. 8. It consists of a
variable-length data field and a fixed-length operation
code. The data field may contain up to 600 decimal digits
with signs. The operation code is a two-digit number with
sign. A negative sign indicates that the following operation
requires use of new routines stored on magnetic tape. The
computer then proceeds to load the appropriate program
as specified by the code. Once the program is loaded, operation codes again are given a positive sign until another
search on magnetic tape is required. This scheme allows
for practically unlimited expansion of the program. The
present description, however, is limited to the eight operation codes comprising COMPAC-1, which are handled by
routines stored exclusively on the drum' of the computer.
The meaning of these codes is explained below.

DATA

OPERAnON

~-----VARIABLE LENGTH
FIXED LENGTH
UP TO 600 DECIMAL DIGITS
WITH SIGNS

2 DECIMAL DIGITS
AND SIGN

Fig. 8-Form of COM PAC instructions.

Pro file milling-aDO": Data preceding the 00 code
describe a circular or straight section of the profile of a
part. The computer calculates the cutter-center path required to produce this profile, and punches the results in
the form of control tape. Computation of an appropriate
intersection with the cutter-center path of the following
section, and automatic deceleration at a corner are included.
Pocket milling-U01": Data preceding this code specify the number of subsequent profile-milling codes that define the boundary of the enclosed pocket or profile, the
number of roughing cuts required to clean out the pocket,
and the amount of material to be removed in the final cut.
The computer produces a c01).trol tape that completelymachines the pocket on the basis of information defining its
boundaries only. All computational features provided in
profile milling' are carried over to pocket milling as well.
. Feed rate and tool diameter- u 40": A new feed rate
or tool size to be used in subsequent computations is indicated by information preceding the 40 code.
Auxiliary function- u 4l": Data preceding this code
define special control-tape characters that control, on-off
functions at the machine. Two examples are shown in
Fig. 9.

,13
:i
>-

a:l

uu
ow
..IX

01-

a:lU

we>::
e>::o

01-

~E~

11.>-

oii
>-><:z<
<(,
\

'I-

\ EQUATION OF CIRCLE

I .... - ' "

CHORD LENGTH
EQUATION OF CIRCLE

Fig. I7-Computation of intermediate points on a circle.

Approximately 300 instructions are required in this portion of the program.
Output routine: - After a new point is determined on
the cutter-center path, further processing is necessary before a block of output tape can be punched. The following
principal steps are' involved.
1) The differences ~%J ~y, and ~z, between the current
tool location and the desired new end point are computed.
2) On the basis of the specified feed rate and the length
of the cut b.s = V (b.%) 2
(b.yF
(AZ)2, total cutting

STOPS TAPE ON "END OF BLOCK" OR
REFERENCE IN REVERSE SEARCH

MODE

ADVANCE LINE (READ NEXT
BLOCK OF TAPE)

DETECTS FAILURE TO STOP ON "END OF BLOCK"
LINE
,.... DETECTS TAPE READING ERROR
[

~MPLETION SIGNAL

r---~--~)(~~/-)~(------------------------~r

YO EMERGENCY

STOP RELAY
REMOVES HYDRAULIC

CONTROL

TRANSFER LINE (TRANSFER
LTEMPORARY INFORMATION
INTO ACTIVE STORE.}

. . / POWER TO MACHINE

AXES

1""""--.1......::..----.,

MAKES HIGH TAPE CONTROl-LEO
FEED RATES POSSIBLE
WITHOUT REOUIRING A CLOCK
RATE UP TO 8 11 MES GREATER
THAN PRESENTLY USED (110 kcl

EXCESS ERROR

¥"""" SELECTOR

SWITCH

.050"

X-AXIS

<:::>

rSERVO AMPLIFIER

E~~~;~~~R

I A~Up:7~:R

MULTIPLICATION FACTOR

~----j~

PHOTO-

PU~~~~~R~~PE I R~~~~R
READER

fz, 7,1,4,15,8: 30 IPM

I
I
I

r---------- -- __ .J

IS SELECTED BY SWITCH
ON MACHINE PENDANT

'-i
<:::>
;;:s-.
~
c..,
<:::>
~

b
('::)

c..,

LOCATED. AT

I
AXIS TO BE CONTROLLED

~
~

I
I

TACHOMETER

.( IF NEEDED)

l~'''-

c.C5.
~
<:::>

'--h
~

~
HYDRAULIC
M('\TOR

SERVO

VALVE

TORQUE
MOTOR

('::)

SH I FT PULSE TO
COR E LINES FOR
READING IN ANO
SHIFTING INFORMA1'fON
...

" - QUANTIZER
FEED BACK
UNIT

[iliill.~

CLOCK (TRIGGER
SUPPLY FOR BISTABLE
MULTIVIBRATORS)

[

POSITION INDICATOR

~
~.

-[MONiTORS COMMAND POSITION
OF ANY SELECTED AXIS FOR
SET UP. TOOL CHANGE, ETC
TO THE NEAREST .0002 INCH

<.Q

~
~
~

;;:s-.
~.

TEST IN
PROGRESS LI GHT

ERROR

LIGHT

1.6 Me

CLOCK CONTROL

('::)

~
"....
('::)

SHIFT PULSE

CONTROL.
A BUILT- IN SELF CHECK FEATURE WHICH
MARGINALLY CHECKS ALL LOGIC FROM THE
TAPE READER AMPLIFIER THROUGH TO THE

ERROR REGISTERS WITH SIMULATED TAPE
SIGNALS ANO CONTROLLED CLOCK ANa
SHIFT PULSE AMPLlTUOES.
IF THERE IS ANY MALFU1'IICTIONG AN ERROR
LIGHT WI LL GIVE WARN I NG.

Fig. 18--Block diagram of machine control unit.

I--'

. PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

rat ely controlled pulse rate. Consider the serial counter
labeled command scaler in Fig. ·18. A count up to overflow
starting with zero will determine an interval directly proportional to the clock frequency. If the number of stages
in the command scaler is changed, the time to overflow will
vary by pow:ers of two. If, in addition, the rate of counting
is changed, a further variation in the time to overflow can
be achieved. If the time to overflow of the command scaler
is defined as the cutting interval, then by combining the
two above mentioned techniques, a very large range of
cutting intervals is possible.
In the Bendix system, the length of the counter is arranged to be equal to the length of the binary number
representing the largest of the three movements toebe produced. For example, a movement in space of ~% = 6.5336
inches (2 15 pulses), ~y = 0.8192 inches (212 pulses), and
AZ = 0.0000 inches (zero pulses) determines a command
scaler of 16 stages. The rate of counting in the command
scaler is controlled according to the feed-rate number.
This is calculated from:

NON-

CARRY

PULSE

CLEAR

FROM BUFFER
STORAGE

MAGNETIC-CORE SHIFT REGISTER
FEEO PULSE
TO COMMAND
SCALER

'FiEDRArENUMBER SrORAG~-

Fig. 19-Binary-operational multiplier using magnetic-core
shift registers.

over a complete overflow period (2N input pulses) will be
and ~z pulses respectively. The pulses generated
for each axis are essentially uniformly distributed
throughout the counting interval due to the linear interas described in Section II-B, 2. The reason for this parpolating properties of f-:e binary op~rational multiplier.
ticular form is now clear: the feed-rate number merely
The control of distances is thus accomplished.
controls the input pulse rate to the command scaler so
3) Pulse Multiplication: The previous two parts dethat it overflows in ~t seconds.
scribe the use of two cascaded binary operational multiThe\ control of pulse rate is accomplished by a circuit
pliers for control of the number and rate of pulses gencommonly known as a binary operational multiplier. This
erated in any given interval. In this scheme a maximum
device has two inputs, one of which is a pulse train and
of one command pulse is produced on each circulation of
the other a binary number. The output is another pulse
the command scaler. Actually, since any binary combinatrain, scaled down in the rate proportional to the fraction
tion is possible in the distance commands, the output pulse
represented by the binary number. The basic principle of
rate
may be as low as half of the recirculation rate.
the multiplier is well known. 5 However, the circuit used
The recirculating registers used in this system are 22
in this system is novel in that it utilizes serial computing
bits long. A complete circulation therefore requires 22
techniques.
clock pulses. In order to achieve the high command-pulse
Fig. 19 shows schematically the multiplier which varies
rate of 20,000 per second previously mentioned, without
the counting rate of the command scaler according to the
an unduly high clock rate, a means for pulse multiplication
feed-:rate number specified. Two delay-line type recirculatis provided in the control unit. As described below multiing registers are used. One is operated as a serial counter
plication becomes effective only at relatively high pulse
(that called feed-rate scaler in Fig. 18) by recirculating its
rates.
contents through a half adder into which one timing pulse
The action of the pulse multiplier can best be described
is added each cycle. The first "one" produced by the adder
by way of an example. Consider a case in which the feedduring each circulation is termed the noncarry pulse. This
rate number is 1023 (2 10 _1) and ~% is 1.6384 inches
pulse is used to gate the contents of the other recirculating
(10,000,000,000,000 or 213 pulses). The input pulse rate
register which contains the feed-rate number, most sigto the command scaler will be 5000 per second at a clock
nificant bit first. The resulting pulse train then serves as
frequency of 110 kc. The output pulse rate from the cominput to the command scaler.
ma?d scaler wi!l be scaled down by two due to the configu2) Cutting-Distance Control: The command scaler beratIon of the dIstance command as a binary number. Thus
ing another serial counter, also generates one non;arry
the maximum rate that could be achieved would be 2500
pulse each time an input pulse is added. If these noncarry
pulses per second, corresponding to a feed rate of 30 inches
pulses are used to gate the output of the ~%, ~y, and ~Z
per minute.
recirculating active-storage registers, the gated outputs
If the feed-rate number turns out to be greater than
1023, C!S a result of a request for a higher feed rate the
,
G~. A. ~eyer, "Digital techniques in analog systems" IRE
1 RA N .". ON ELUILUIC L~L~.'lJL'ERS, vol. EC-3, pp. 23-29; J u~e, 1954. pulse multiplier becomes operative. Suppose, for exa~ple,
~%, ~y,

Ho and Johnson: Design of a Numerical Jl;Iilling Machine System
the feed-rate number turns out to be 2047 (2 11 -1). Overflow detection is then carried out one stage earlier in the
command scaler, thus terminating the process when only
half as many pulses are generated. The cutting interval,
or time to overflow, is thereby halved.
The pulses now are all multiplied by two before being
transmitted to the servosection. The resultant number
of pulses therefore remains the same as in the previous
example. Since the cutting interval is halved, the feed rate
is effectively doubled.
The feed-rate number is still treated as a 10-bit number
with the binary point automatically shifted one bit to the
left. For feed-rate numbers of 12 or 13 bits, the outputs
from the command scaler are multiplied respectively by
four or eight. With this technique, command pulse rates
as high as 20,000 per second can be produced with a clock
rate of only 110 kc. A smooth transition to these high
feed rates is automatically provided in the scheme.
When the desired feed rate is sufficiently high as to
require pulse multiplication, the number of pulses generated must be a multiple of two, four, or eight. Consequently, the distance commands must be rounded off to
0.0004, 0.0008, or 0.0016 inch, respectively. To prevent
errors of this type from accumulating from one block of
control tape to the next, the computer program described
in Section II keeps track of such round-offs. It continuously applies corrections from one block to the next so
that at anyone point the computed distance commands
cannot be off by more than the current round-off quantity.
Situations in which this round-off procedure would tend
to produce errors in the part surface also result in transient servo errors. At the high feed rates involved, the
servo errors can be expected to be many times larger than
the round-off errors. Likewise, under steady-state conditions, the pulse-multiplication technique does not generally
produce a significant effect. Few servos, for example, have
dynamic characteristics capable of detecting the difference
between a uniform command pulse rate of 20,000 per
second and a pulse train consisting of 2500 groups per
second of eight pulses each.
C. The Digital Servo
Once the distance commands have been transformed
into discrete pulses, they must be further converted into
machine motion which is essentially analog in nature.
,,Several approaches are possible by which the required
transformation can be made. In one method the discrete
pulse trains are first converted to analog commands. These
then are compared with feedback signals produced by
analog devices as in conventional servosystems. Such a
scheme is described by Moore. 6 The approach employed in
this system, as well as numerous other applications of this
type, utilizes digital feedback instrumentation. For each
unit distance traveled by the machine, a pulse is generated
6

G. T. Moore, "The Numericord machine-tool director" this
,

~su~p.~

by the feedback instrument. A running difference is kept
between the number of command pulses and the number
of feedback pulses received by the servosystem. This
difference is then used as the error signal to drive the servomotor.
1) Feedback Instrument: The feedback instrument used
in the system is a rotary electromagnetic device called a
quantizer. It has one rotor winding and two stator windings spaced 45 electrical degrees apart. The windings of
the rotor and stator, having the equivalent of 250 poles,
are formed by etching conductors and copper-clad plastic
disks. The rotor is excited from a high-frequency source
of about 1.6 mc. The signals induced in the stator windings
depend on their position with respect to the rotor winding.
They are thus amplitude modulated by movement of the
rotor. The two stator outputs are fed to demodulators and
wave-shaping circuits from which they emerge as rectangular, or two-level, waves.
2) Synchronizer: The synchronizer, which receives the
rectangular signals from the demodulators and waveshaping circuits, converts them into pulses representing
unit motions. The zero crossings of one signal are used
to define unit rotation of the quantitizer shaft. The relationship between the direction of zero crossing and the
level of the other signal indicates the direction of rotation.
Command pulses produced by the digital interpolator
are also fed to the synchronizer and combined in the opposite sense with quantizer pulses. Simultaneous occurrence of command and feedback pulses is detected, and if
necessary one is delayed by a clock period. The synchronizer outputs then consist of two trains of pulses, one train
representing error in the positive direction and one in the
negative direction.
3) Error Register: The error register is an ll-stage
reversible binary counter, composed of flip-flops, which
counts continuously the pulses supplied by the synchronizer. The count in the register therefore represents at all
times the accumulated difference between the number of
command pulses and the number of feedback pulses. The
first five stages of the register are decoded by means of a
binary-weighted resistive summing network to produce
a proportional voltage. The last stage of the register indicates the sign of the error and is similarly included in the
network with appropriate weight. Large counts are not
decoded but clamped Cat a fixed level.
Provision is .made to detect overflow of the counter and
thereby cause an emergency shut down of the system.
Excess-error detection at adjustable levels of 0.010, 0.025,
and 0.050 inch is also provided.
4) S ervodrive: The analog error voltage from the resistor decoding network serves as input to the servoamplifier. In all present applications of the system, the machine
slides are powered by high-performance electrohydraulic
drives. The drives consist of piston-type rotary hydraulic
motors controlled by servo valves f.rom a constant-pressure
hydraulic supply. The valves in turn are stroked by elec-

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

trically operated "torque motors." As a combination, the
specially developed valve and torque motor have a bandwidth in excess of 250 cps.

D. Supervisory Control and Other Features
1) Control Logic: Control tapes are read one block at
a time into buffer storage by means of a photoelectric tape
reader operating at the rate of 150 lines per second. Four
magnetic-core shift registers serve as temporary storage
for the three distance commands and the feed-rate number, and another four registers serve as active storage.
Under normal circumstances, overflow of the command
scaler causes clearing of all counters and storage registers
in the interpolator and the transfer of input information
from temporary storage into active storage. Reading of
the next block of control tape into temporary storage then
commences.
As information is read in from the tape, it is automatically checked for parity. Detection of an error causes an
appropriate indication, and switches the machine to manual
control at the conclusion of the cutting specified by the
previous block. Other control circuits are provided to detect excess errors in the servo systems, failure of the
clock source, and malfunctioning of the tape reader, the
power supplies, and other components. These errors are
indicated on the control panel and automatically cause
suspension or complete shutdown of the machine system
until the errors are corrected. On-off operations at the machine can be initiated by the special auxiliary-function tape
codes at any time.
2) Manual and Tape-Control Synchronization: The interpolator is designed so that at any time it is possible to
switch from tape to manual control and vice versa without
loss of synchronization. This is accomplished by terminating the flow of input pulses to the command scaler under
manual control, thus preventing the further generation of
non carry pulses in the command scaler. Consequently, no
output pulses will be delivered to the servos. The command scaler, as well as all storage registers for the distance
numbers and the feed~rate number, retain their contents
through recirculation until automatic operation is resumed.
The feed-rate scaler associated with the feed-rate register
is kept running at all times. Command pulses are obtained
for manual control by detecting its overflow at different
stages. The clock frequency is adjustable from 20 to 100
per cent of nominal without disturbing the operation of
the unit. Since a change in clock frequency changes the
over-all speed of operation proportionally, a wide range
of manual feed-rate override is thereby provided.
3) Position Indicator Counter: Command pulses produced by the interpolator canbe accumulated in a separate
reversible decimal counter. The contents of this counter,
having a capacity of 999.9998 inches, is displayed directly
in inches, using decimal indicator tubes. The counter can
be switched to any desired axis for monitoring purposes or
for manual positioning.

4) Marginal-Checking Circuitry: The control unit contains built-in circuitry by which tape-controlled operation
of the unit can be simulated for test purposes. Simulated
operation is obtained by furnishing dummy inputs by way
of the tape-reader amplifiers to the temporary storage
registers and causing a transfer into active storage. The
interpolator begins production of command pulses ·for
the servosystems on each axis. The simultaneous overflow of all error registers with the command scaler in the
proper state will reinitiate the test cycle as long as the
test switch is on. Otherwise the process halts, and an error
signal appears. This built-in circuitry tests 80 to 90 per
cent of the logic in the control unit. A more exhaustive
test can be made with test tapes which in addition check
the tape reader and the remaining portion of the servosystems. Supply-voltage levels, the clock-wave shape, and
the shift current for the magnetic-core circuitry can be
varied to provide for marginal checking under the above
test conditions.
E. Components and Construction
The machine control unit uses vacuum-tube flip-flops,
triggers and cathode followers, together with germanium
diodes, in a de gating system. Storage registers and scalers
are made up of magnetic-core shift registers with associated vacuum-tube read-write circuitry. The unit contains
approximately 400 tubes, 2800 diodes, and 200 magnetic
cores. Plug-in type construction and etched circuitry are
used throughout. A view of the control unit is shown In
Fig. 20 with the cabinet doors removed.

IV.

MACHINE

Fig. 21 illustrates an application of the control equipment just discussed to a contour milling machine of the
moving-column type. 7 Position of the cutter is controlled
along three mutually perpendicular axes with respect to a
stationary work piece mounted on the vertical angle plate.
Motion perpendicular to the plane of the work (along the
cutter axis) is provided by the spindle head assembly. The
spindle in turn is carried by a slide which moves vertically
on· the column. Longitudinal motion is provided by the
column itself, carrying the operator's platform with it.
All three axes of the machine are equipped with highperformance hydraulic servodrives of the type described
in Section III -C, 4. The longitudinal axis, having the longest stroke of 172 inches, is driven through a rack-andpinion arrangement. Backlash is effectively eliminated by
means of a dual-drive system in which two motors, connected hydraulically to the same valve, work against each
other through separate gearing to the rack. The vertical and
transverse axes, having strokes of 52 and 18 inches respectively, are driven by single motors acting through precision
ball-nut lead screws. A preloaded double-nut arrangement
7 Machine designed and built by the Kearney and Trecker Corp.,
Milwaukee, Wis.

Ho and Johnson: Design of a Numerical Milling Machine System

Fig. 22-Aircraft fitting machined under numerical control.

Fig. 20-Photograph of machine control unit.

mechanical structure. For small signals, the bandwidth of
the longitudinal system is in the range of 7 to 10 cps in
spite of a moving structure which weighs over 35,000
pounds. A sudden· stop from 30 inches per minute (without programmed slowdown) results in an overshoot on
the order of 0.010 inch. The bandwidth of the other two
axes is considerably higher, with correspondingly faster
response to severe transients.
Maximum speed of the machine slides is nominally 180
inches per minute, although feed rates are usually limited
to 100 inches per minute during actual cutting. Initial tests
indicate that the machine is capable of repeatability within
0.001 inch over the entire range of strokes.

Fig. 21-Application of control equipment to a large three-axis
contour milling machine.

is designed to eliminate backlash between the machine slide
and the lead screw.
Feedback information is provided to the machine control
unit by means of quantizers, according to the system outlined in Section III-C. In the case of the longitudinal axis,
the quantizer is coupled to the slide by means of a highprecision rack and pinion. For the two shorter axes, the
instruments are geared directly to the drive screws. Gear
ratios between quantizers and machine motion are such as
to produce 5000 pulses per inch of travel.
Dynamic response is unusually good for a machine of
this size as a result of the hydraulic drives and the care
taken during design to achieve a high-stiffness, low-mass

ILLUSTRATION OF SYSTEM PERFORMANCE

The part illustrated in Fig. 22 is an aircraft fitting
produced from a forged blank by a machine functionally
similar to that just described. The process consisted of
both roughing and finishing cuts on the outside of the
flange, on the inside of the flange on both sides of the
web, and in an area between the flange and the center
circle on each side of the web. Single cuts were taken in
order to machine mounting pads on one edge of the flange,
and to finish the inside of the large hole near the center.
Preparation of the process sheets for this part by an
experienced process engineer would require approximately
10 hours, starting with a suitable working drawing. When
advantage· is taken of the symmetrical-cutting feature of
the machine, six sheets of the COMPAC-1 format are
necessary for full definition of the machining process.
These result in a process tape 46 feet long, which requires
2.4 hours to punch and verify with the Flexowriter. Computer time is approximately 1.2 hours. This latter time
includes the reading of· the process tape, all necessary
computing, the punching of the control tape, and the printout of coordinates and feed rates at frequent intervals

PROCEEDINGS OF T.HE EASTERN COMPUTER CONFERENCE

along the cutter-center path. The print-out operation accounts for about 25 per cent of the total computing time.
The control tape for the part is 70 feet long.
TABLE I
PRODUCTION PROCESS TIMES FOR TRACER-CONTROLLED AND
N UMERICALL Y CONTROLLED MACHINES

Tracer
Control
Fixture design and manufacture
Template design and manufacture
Process~sheet preparation
Process-tape punching and verifying
Computing time
Total tooling time

126 hours
84 hours

Numerical
Control
71 hours
-

10 hours
2.4 hours
1.2 hours

210 hours

85 hours

3.75 hours
1.75 hours

1 hour
1 hour

5.5 hours

2 hours

(which incidentally was not produced by numerical
control).
Incorporation of linear interpolation with the machine,
permitting final control-tape preparation directly by a small
general-purpose computer, represents a compromise having many desirable operating features. Sufficient data is
not yet available to permit a full scale economic evaluation of the system. However, the above example illustrates
the reasonableness of processing time, computing time,
and tape lengths characteristic of this basic approach.
Productivity of the controlled machine in terms of speed,
accuracy, and surface finish has been well demonstrated
in actual field use.

VI.

Set-up time for first part
Machining time
Total time machine occupied

The times involved in the production process are summarized in Table I. Corresponding data for the conventional method utilizing a tracer-controlled machine are also
given. This comparison shows a substantial reduction in
both tooling and machining times for the numerical approach. It also emphasizes the fact that the operations
directly concerned with numerical processing now require
a relatively small fraction of the time consumed by the
more conventional steps sti11left in the production process,
such as the design and manufacture of the holding fixture

Discussion
Question: Doesn't pulse multiplication
effectively change the quantizing level of
the system?
Mr. Ho: The answer to that is a
qualified yes. If you are multiplying by
2, 4, or 8, then the distance you can specify
is rounded off to the nearest 2, 4, or 8
pulses. In other words, if you are cutting
at 240 inches per minute, you can specify
distance only as close as 1.6 mils. We say
this isn't too big a handicap. If you are
cutting at 'such a high speed, the servo
error is going to be many times greater
than that. Also, in the automatic tape
preparation program, we keep track of
these errors and correct from one block
to the next so that at any time we are never
off more than the 2, 4, or 8 pulses. When
we finally come to stop after cutting, the
automatic program applies a correction and
returns the tool to the accurate position
again. As far as the tool engineer is concerned, he doesn't have to worry about this.
In fact, he doesn't know it exists.
Question: What are the checking and
printing features of the automatic program?
Mr. Ho: There are several loading
checks on the automatic program. For ex-

ACKNOWLEDGMENT

Development and production of the equipment described
in this paper would not have been possible without the
joint efforts of a large number of people. Over-all responsibility for the system, as well as development of computer
programs, the machine control unit, and servodrive components has been carried by the Research Laboratories Division of Bendix Aviation Corporation. Bendix Computer
Division has contributed materially in the development of
magnetic-core circuitry, and in production of the control
units. Design and production of the machine mentioned in
this paper was entirely the responsibility of the Kearney
and Trecker Corporation. The authors gratefully acknowledge the contributions of numerous personnel in all of
these organizations.

ample, if a tool engineer describes a cutter-center path which would result in a
situation impossible for the machine to
cut, an error stop would be indicated by
the automatic program.
The print feature allows the position of
the cutter to be checked as the tape is
being prepared. Now, this print feature can
be suppressed by including another instruction on the process sheet. On the circular
arcs or other curves' points, we can print
out intermediate points at any frequency
specified, for example, every 15 points. But
the end point of every section, straight line
or curve, is always printed out if print-out
is requested. Print-out slows down the
computer somewhat. Every print-out takes
approximately ten seconds.
The computer output tape is a sandwich
construction utilizing aluminum foil between two mylar layers. The process tape
made by the Flexowriter can be verified
by the verifying attachment to the Flexowriter, essentially by typing the same information for the second time and comparing with the first tape.
Question: Is there any compensation
made for variation of cutter size due to
wear?
Mr. Ho: No. Maximum travel of the
machine is 170 inches, but the control unit

can accommodate straight-line cuts up to
approximately 409 inches on each axis.
Question: What is the drum storage
capacity in the G-15 computer?
Mr. Ho: This machine has a storage
capacity of approximately 2000 words.
Question: What are the components in
the computer system which prevent the
computer from running in real time?
Mr. Ho: The most difficult problem in
numerical control is cutter-center offset. Especially in three dimensions, the cuttercenter offset problem generally is very
complicated and leads to all sorts of exceptions. A completely general tape-preparation system which could keep up 'with the
machine and supply information as demanded would require a very large and
very fast computer. In general, we find
a physical separation for buffering purposes is almost essential between the cuttercenter offset problem and the machine control problems. We feel that if you have
a computer that is fast enough to supply
information to the machine in real time
under peak demand conditions, then most
of the time the computer will be sitting
there doing nothing. Only when really complicated situations come up that require full
use of the' computer, would it be utilized
efficiently.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Logical Organization of the DIGIMATIC Computer
JACK ROSENBERGt

SYSTEM SPECIFICATIONS

RECEDING the design of the original DIGIMATIC computer and control system, Electronic
Control Systems Inc., (ECS) performed technical
and economic surveys to determine the requirements for
successful entry into the commercial automatic machinetool market. The results may be summarized as follows.
1) A three-axis milling machine, with any two of the
three slides simultaneously controlled, was the most
logical tool to be adapted.
2) Cost and complexity of the control apparatus directly
associated with the machine tool must be minimized,
and reliability maximized. This dictated the elimination of the computing or interpolation function from
the tool, with only the control function remaining.
3) From the standpoint of flexibility, accuracy, reliability, cost, and bandwidth, the only form of precomputed memory consistent with the above was magnetic tape recorded in digital incremental form.
4) The programming process by which part-drawing
data is converted to magnetic-tape commands in a
special-purpose computer must be rapid, accurate,
economical, and easily accomplished by regular machine-shop personnel.
5) To satisfy the bulk of shop needs (about 95 per
cent of commercial parts), the computer must generate (interpolate) linear or true circular paths with
an accuracy of +0.001 inch, without accumulation
of error, from data already present on drawings or
from those readily obtainable from drawings.

MATHEMATICAL REQUIREMENTS-STRAIGHT LINES

As in most engineering projects, the most difficult task
was defining the problem; once this was accomplished, the
technical solutions were evolved fairly readily. Since this
paper deals mainly with the computer portion of the system, it is possible to derive the input data, its code, and
the input device from 4) and 5) above. Part drawings
normally define a piece by the decimal coordinates of
the start and end of each segment of the contour, and for
circular arcs, usually include the radius and decimal coordinates of the center. Thus, the computer should accept
contour breakpoints in decimal form; if code conversion is
necessary prior. to interpolation, it should be accomplished
automatically.
"
The input mechanism, for reliable use by shop people,
should be extremely simple and provide means for verificat Electronic Control Systems, Inc., Los Angeles, Calif.

\R~

tion. A decimal keyboard device (such as that on an adding
machine) with print-out tape is familiar to nearly all such
personnel.
Let us begin by examining the case of a straight-line
segment in the X - Y plane. The general equation of such a
line
(1)
Y = mx + b,
must be solved by the computer, and given the start and
end points of the desired segment (which lies on this
line) .
Since it is fair to assume that a workpiece will contain
continuous contours, the constant b may be eliminated
from (1) due to Jhe fact that the cutting tool has been
brought to the proper start point for this path as a result
of the completion of the previous segment.
Now m represents the slope, or tangent of the angle be:..
tween the desired line and the X axis. By definition, the
tangent is equal to the change in Y from the start point
(P 1) to the end point (P 2) divided by the change in X
over the same interval. It is exactly expressed as follows:

Y2 - Yl
m=---'
X2 -

(2)

Eq. (1) has therefore been reduced to the form

Y =

Y2 - Yl

---x+ b,

(3)

X2 -

and for the reason indicated above, b need not be computed. This expression is illustrated graphically in Fig. 1.
The line to be traversed has now been expressed in terms
of the data normally furnished on a part drawing.
y

m= tan 0 =&
Llx

Fig. 1.
ENGINEERING SOLUTION-STRAIGHT LINE

Quantities Y1, Y2J %1, and %2 are constants over the entire
line, and are known before the interpolation begins. Since
differences Y2 - Yl and %2 - %1 need be calculated only
once for the entire segment between P 1 and P 2, these subtractions may be performed on a relatively slow mechanical adding machine without any undue time penalty, inasmuch as subtraction time is less than keyboard entry time.

P. CAS TAN IAS

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Some keyboard device to receive data is unavoidable, so it
may as well be an adding machine.
Therefore we will insert quantities Y2 and Y1 (in that
order) .into an adding machine, then order it to find the
difference, and automatically store Y2 and Y1 in an electrical
register (relays). Next, X2 and Xl will be entered, and
X2 - Xl will be stored in another relay register. Let us call
these differences Ay and AX, respectively.
An electronic counter can be used to divide input pulses
by an integral quantity which may range from one up to
the counter capacitance; in such service it is usually termed
a predetermined counter. Most commercial predetermined counters using a binary or modified binary code,
are preset mechanically and require a considerable interval after recognition to be reset to an initial condition and
to be ready to accept further input pulses.
However, a high-speed all-decimal predetermined
counter can be designed around the magnetron beamswitch tube. Each target may be connected to a coincidence
gate. When the total count in the multistage counter has
proceeded from 0 to the integer by which the input pulses
are to be divided, recognition will occur, an output pulse
will be generated, and the entire counter rapidly reset to
o via the spades. One microsecond is sufficient for reset;
thus, the input pulse rate may be as high as 1 megacycle
and still permit reliable counting and resetting. We term
the process dividing, and such a counter, a divide counter.
Fig. 2 shows the logic of a divide counter.
r-------- ------------------,

RESET PULSE (TO 0)

Fig. 2-Divide counter, register containing quantity by which
the clock pulses will be divided.

Suppose we feed pulses from a common clock into two
divide counters, one dividing the clock rate by Ax, the other
byLiy, as pictured in Fig. 3. If the clock frequency f is
constant, a train of uniformly spaced pulses will emerge
from each divide counter; the output frequency of the
first counter will be
be

L,
that from the second counter will
Ax

~,
and the ratio between these rates will be
Liy
~y

---=-,

TAPE
RECORDER

Fig.

4~Straight-line

generator.

continue until the Y-axis drive receives a pulse total of Ay,
the X-axis drive a total of Ax. Two additional counters and
three recognition gates as shown in Fig. 4 monitor this
process and turn off the clock when recognition signifies
P 2 has been reached. Simultaneously the pulses from the
divide counters are recorded on appropriate tracks of an
eight-track magnetic-tape recorder for later use in controlling the machine-tool table.
Since each of the two output pulse trains will have uniform pulse spacing, they will be periodic, and therefore
optimum for driving servomechanisms at constant velocities. The description of the DIGIMATIC computer as a
linear interpolator requires but one further embellishment
to be complete. For vector machine table feed rates to be
controlled automatically, the two output pulse trains are
sampled by analog pulse rate discriminators, combined in
quadrature, and the resultant voltage compared to that
commanded by a feed-rate potentiometer calibrated in,
inches per minute. The difference voltage operates a clock
frequency control system in a closed loop, so that the desired feed rate results.
MATHEMATICAL REQUIREMENTS-CIRCLES

f
~x

Fig. 3-Straight-line generator.

(4)

which is exactly the slope of the line connecting P 1 and P 2.
Assuming each output pulse represents a motion of 0.001
inch by a machine-tool slide, the interpolation process must

The engineering solution given above describes a means
for generating a line of constant slope by producing two
pulse trains whose frequencies are constant, and are related
so that
(5)

N ow let us examine the mathematical characteristics of a

Rosenberg: Logical Organization of the DIGIMATIC Computer

circle. The general equation for a circle with its center at
point (a,b) in the XY plane is

(x - a) 2

+ (y -

b) 2 = r2.

(6)

We can derive the exact expression for the instantaneous
slope at any point by first taking differentials
2(x - a)dx

2(y - b)dy =

(7)

Separating variables we reduce it to
(y - b)dy = - (x - a)dx,

(8)

and finally the slope is expressed by
dy

x - a

y - b

(9)

ENGINEERING SOLUTION-CIRCLES

In the solution for first-degree equations, predetermined
dividing counters were utilized to divide the common
clock-pulse source by two constant quantities. Referring
to the expression for slope of the circle given in (9), it
can be seen that if quantities oX" - a and y - b can be made
available at all times, the predetermined counters mentioned earlier can be used to divide the clock-pulse rate f
by these varying quantities, and a circle can be generated
by continuously changing the slope to fit the above equation.
Inasmuch as the slope changes in sign four times during
the generation of a complete circle, it becomes necessary to
keep track of the magnitude and sign of the quantities
% a and y - b. This can be accomplished by the use of
reversible counters which can add and subtract pulses, instead of simple monodirectional counters which would have
sufficed for performing the pulse-summing operation in
Fig. 4. We call such counters "sum" counters, and we have
obtained economy of components and sufficient counting
speed by using decimal glow-transfer counter tubes of the
Erikson type. Since it is necessary to present the quantities
oX" - a and y - b in parallel decimal form to the inputs of
the divide counters, reversible gas-tube GSI0C, which
has all of its ten cathode electrodes brought out to the tube
socket, was chosen.
In the case of the circle, the input commands to the divide counters have to be switched from the output terminals of relay registers, as described for a straight line, to
the output terminals of.sum counters. In addition, direction
gates must be added between the output of the sum counters and the tape-recording channels, and must be commanded by.the sign of the appropriate sum counter, to
make the logic completely consistent with the mathematical
requirements of a circle. The logic of a circle generator,
which receives as input information the coordinates of the
start point, end point, center, and instructions as to
whether the circle is to be generated in a clockwise or
counter-clockwise direction, is given in Fig. 5.
To convert the straight-line generator of Fig. 4 to a
circle generator of Fig. 5, it is necessary to provide a 100-

pole double-throw relay, which switches the 50 input terminals of each 5-decade divide counter from a relay register
to a sum counter. Some changes in the method of handling
input data must also be incorporated, to permit the sum
counters to be initially preset to the actual quantities
oX" 1 - a, Yl - b at the start of the circular arc.
Mention should be made of the singularities which occur
four times during the generation of a complete circle. The
instantaneous slope of the circle twice goes to zero and
twice becomes infinite. The former cases correspond to
the quantity.oX" - a becoming equal to zero, and the latter
cases occur when y - b becomes zero. If the clock frequency is divided by zero at these times, an infinite output
rate should be produced by the predetermined counters.
However, as soon as one additional pulse at this infinite rate
is emitted and added into the previous total of zero in the
reversible counter, the appropriate total changes from zero
to either plus one or minus one, and the predetermined
counter is then asked to divide by a finite integer. The
problem can be resolved by setting up a system of logic
which causes the predetermined counter to emit one pulse
without receiving a pulse from the clock, if it is ready to
be preset by the total in the reversible counter and finds
this total to be zero. By thus generating a pulse with no
input from the clock, we can simulate an infinite ratio of
output to input at the point of singularity. Fig. 5 assumes
the use of this type of predetermined counter.

Fig. 5-Circle generator. Note i-In sign block of reversible counters, the letter s refers to the true sign, the letter c refers to
the complement· of s. Note 2-With switches on the sign outputs of the above counters set as shown by the solid lines, the
circle will be generated in a clockwise direction. If they are set
as shown by the dashed lines, the circle will be generated in a
counter-clockwise direction.

To illustrate the operation of this all-decimal interpolator, Fig. 6 shows the path described on graph paper by the
occurrence of pulse outputs in the case where the center of
the circle is at (0,0) and the radius is 10 increments. Although the appearance of the outline is not smooth, in practice a smooth contour is machined because of the smoothing action of the servomechanisms which drive the slides
of the machine-tool table. Furthermore, the example was
chosen for simplicity, since a circle with such a small radius
is somewhat academic. Standard milling tools are available
only in diameters of 1/16 inch (0.0625 inch) and higher.

PROCEEDINGS OF THE EASTERN COlvfPUTER CONFERENCE

S Af T
~IT

I'-

-era 0
/

.....

r'\
Fig. 8-Model 120 DIGIMA TIC computer.

Fig. 6-Interpolated circle. (Radius 10 increments.)

Fig. 7-Model 120 DIGIMATIC computer.

Thus additional smoothing will be provided by the large size
of the periphery of the cutting tool compared to the value
of an increment, which in our case is 0.001 inch.
The DIGIMATIC Model 120 computer, whic11 follows
the logical principles described above, is pictured in Fig. 7.
The relay register, clock, divide counters, and sum counters are housed in the cabinet at the left. The desk contains
the magnetic-tape handler and some input distribution circuits, while the control console (including adding machine)
may be seen on the desk top. The cabinet on the right contains only power supplies.
A close-up of the control console is shown in Fig. 8.
Fig. 9 is a· photograph of a triangle, circle, and parabola
which was machined from tape prepared by the 120 computer.
GENERATION OF OTHER CURVES

The technique of using two divide counters to generate
a curve of continually changing slope can also be applied
to other second-degree curves. As an example, the equation
for a parabola (principal axis parallel to the X axis) is
(y - k)2 = 2p(x - h).

(10)

In this case k, p, and h are constants. The instantaneous
slope is
dy

y - k

Fig. 9-Geometric contours produced by model 120
DIGIMATIC computer.

~
AXiS

+(y-k)

CLOCK

~
+p
.

x Axis ~

Fig. lO-Parabola generator.

It may be seen that it differs from the slope of a circle
only in the respect that the numerator is a constant instead
of a variable. Fig. 10 indicates the logic of an interpolator
for this kind of parabola. In practice, we have generated
parabolas by entering the information as if a circle was
desired, and preventing x-axis command pulses from
reaching the X sum counter (by removing a driver tube).
The case of an ellipse can be analyzed as follows. The
general equation is
(x - h) 2

(y - k)

---+---=
a

(11)

y-k

From this the slope is derived

(12)

Rosenberg: Logical Organization of the DIGIMATIC' C0111puter
yAxis

yAxis

xAxis

Fig. ll-Ellipse generator.

(x - h)b 2

(y - k)a 2

Fig. 12-Higher order curve generator.

(13)

Fig. 11 shows how this could be implemented. Since two
additional constants occur, a and b, additional divide counters would be necessary, as well as memory registers to
retain the quantities for presentation to the extra divide
counters.
This computing philosophy can be extended to accommodate higher order equations. As an example, consider this
fourth-degree curve

y = S(x - 10)4.

(14)

It has the instantaneous slope
dy

20(x - 10)3,

Discussion
Question: What is the reliability of
the Electronic Control Systems Director
and mean time to failure?
Answer: Well, the most unreliable component of our Director is the human operator and unfortunately he breaks down
pretty often. The entry of numerical data
from a planning sheet is a fairly tedious
process and we find that monotony is the
chief producer of errors. It is a little hard
to answer this accurately, since we have
been operating a laboratory system, not an
industrial one.
I will say that the mean free time is
beyond one day. We have some peculiarities
in the Los Angeles climate which control
this. We use standard telephone-type relays
and whenever there is a strong dust storm
from the desert, it plays hob with the relays
and we have to clean them out. In general,
the mean free time is more influenced by
electromechanical than electronic components.
Question: What is the least count,
least programmable increment of the ECS
system? What is the maximum feed rate of
any axis?

(15)

which can be generated by the computer logic given
Fig. 12.

III

CONCLUSION

A special-purpose, high-speed, digital computer has been
described which operates completely in the decimal system,
and contains built-in programs for straight lines and circles,
selection being made by the turn of a switch. The input
data required is inserted very simply, and is either already
present or may be simply derived from information on part
drawings. This computer could be described as a function
generator, which generates analytically expressed relationships in a digital manner. Similar computing principles are
being applied to control problems outside the machine-tool
field.

Answer: In our present system, one
pulse equals one thousandth of an inch, so
this is the least programmable increment.
Our computer controls the vector feed rate,
so that the maximum feed rate in one axis
is not necessarily the actual cutting feed
rate. To put it another way, our maximum
clock-pulse frequency is 500 kc and if you
will go through the mathematics and logic
of our system, it works out that for most
typical cuts on our prototype Bridgeport
Mill we must limit the feed to 15 inches
per minute. The manufacturer of the original machine provided power feeds up to
this rate.
We designed our DIGIMATIC computer to produce feed rates up to 15 inches
per minute. There are some cuts that we
could make much faster. However, the machine-tool servos could not follow them. To
put it another way, the feed rate must depend on the type of material and depth of
cut. Our new Director, now near completion, will interpolate at eight times real time
for feed rates up to 50 or 100 inches per
minute.
Question: Do you verify the tape before you use it for direction of the machine tool, and if "yes," by what means?

Answer: At the present time, we make
a pulse count. Our magnetic tape contains
six control channels, a plus and minus channel for each of the feeds, X, Y, and Z,
and for each segment of the cut we perform a pulse count for each axis. This is a
fairly tedious process, so what we usually
do, unless the part. must be made to unusually close tolerances, is simply machine
a part. It would be possible to play the
tape back into the sum counters of our
computer. We did not provide for it in the
present system. We will in the forthcoming ones.
Question: Do you have a stored program for the interpolation process, or is
it made by hardware?
Answer: We have only two programs
for this computer, both of them built in.
The selection of linear or circular interpolation is controlled by a number of relays
which change the input lines to our divide
counters. In the case of straight lines, the
relays connect the divide counter inputs to
numerical relay storage. In the case of
circles, these inputs are connected to the
sum counter outputs.
This is about the only meaning that I
can place on the question.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

The Master Terrain Model System
JOSEPH A. STIEBERt

;
T

SYNOPSIS

HE master terrain model system is an automatic
data reduction system which will extract threedimensional contour information from maps of various proj ections, aerial stereo photographs, or existing
master models, and store these data in a universal format.
From these stored commands, the system will automatically drive a fabrication unit and produce three-dimensional terrain models.
The master terrain model system was conceived to fill
the serious need of the armed f~rces for a faster and less
costly method of constructing terrain models. The proposed system will replace the present manual process of
construction with a completely automatized system. The
system will further provide a valuable means for the permanent storage of master model data and thus supplant
present unwieldy storage methods and eliminate the loss
of valuable models due to natural deterioration and handling.
The system, which is now in the stage - of prototype
construction, will consist of a map scanning unit, a recording and playback unit, and a contour cutting mechanism,
all units being programmed and controlled by digital computers. The digital computer control system will interject
positioning corrections into the scan data for the various
map projections, so that the scanned three-dimensional
information can be recorded and stored in a universal
spherical format. The computer system will additionally
provide a feedback error correction medium to the scanning
drive and also supply an interpolation of pulse analog signals to produce faired curved surfaces.
The Naval Training Device Center has for many years
been concerned with the development of better methods of
producing master terrain models, reproduction models', and
the surfacing of these models with mapping or photographic intelligence.
A long range program, covering many projects, has
produced .new techniques and equipments which have
contributed greatly to the present success of many operational and training devices used by our military forces.
The development of an automatic system for the production of terrain models has been evolved through many
years of research by engineers of the Center. Many methods were tried to translate contour elevation data into digestible information for recording and machine consumption. Among these have been systems using color coding,
magnetic, electrochemical, photographic, and line counting
t U. S. Naval Training Device Center, Port Washington, N.Y.

techniques. Of those investigated, the photographic processes coupled with copper etching techniques appeared to
offer the best possibility of meeting the desired performance characteristics from a standpoint of simplicity, speed
of scanning, and accuracy.
When the determination of prime feasibility was accomplished, a contract was let to Technitrol Engineering
Company to undertake a design study of one-year duration
for further development of an over-all completely automatic system. As a result of the success of this design
study, a second contract was let to Technitrol Engineering
Company for the finalized design and construction of a
prototype model of the system. This prototype is presently
nearing completion and preliminary tests indicate that all
specification requirements will be met successfully.
SYSTEM ApPROACH

The magnitude and complication of the initial system
designed required a design study of one-year duration to
adequately prove preliminary engineering concepts. The
speed, accuracy, and flexibility requirements of the system
pres'ented many complicating factors which had to be
solved one at a time.
A basic problem was the development of a method of
translating two-dimensional positions plus a third-dimensional code into computer words.
Let us first study the physical aspects of the problem
and the terms to be used. We will be dealing with maps of
various projections, scales, and miscellaneous dimensional
terminology all of which must be reduced to a common
denominator which can be assimilated by a data reduction
system. A three-dimensional model is in effect a scale
map which has been deformed into the third dimension to
simulate the exact contours of the earth which it represents. The problem here is to convert two-dimensional
maps into three-dimensional models. A typical master
model, as shown in Fig. 1, presently is produced bymanual
methods. Thus, a model of this type requires many months
of arduous labor to produce. The objective of the master
terrain model system is to produce these same models in a
matter of hours with greater accuracy and with consequent
savings in cost.
MAP CODING METHODS

The concept of an automatic terrain model system starts
with the preparation of data from a flat map plate for
machine acceptance. If we consider the familiar multi color
map, it has been printed from as many as a dozen separate
color plates to consolidate the colors and mapping intelli-

Stieber: The Master Terrain Model System
PROJECTION
CORREC.TlON·

TIMING
PULSE

Fig. 1-Typical master model.

3D-MODEL

2D-MAP

Fig. 3-Block diagram of system.

tained in incremental steps of 0.01 inch and these data are
converted into the binary system and passed on to the
recording stage of the system.
THE STORAGE SYSTEM

IiP' 1OOl

Fig. 2-Metal-map coding.

gence on a single sheet. One of these color plates (called
a brown plate) has been singled out for our use because
this plate contains the contour lines which represent earth
contours and which will provide the three-dimensional
information we require for the system. If this contour
plate is used to photographically etch a copper-clad laminate sheet as is done in printed circuit techniques, a resulting map in metal will be produced which will provide terrain levels insulated from each other by the line thickness
which has been etched away. If this metal map is now
coded as shown in Fig. 2 by using electrical connections
to supply appropriate voltage levels in ratio to map elevations, the resulting prepared metal map will provide the
third dimension when scanned electrically.
AUTOMATIC SCANNING

The scanning mechanism consists of a single stylus
which is driven across the "metal map" by means of a
servodrive. The stylus makes electrical contact with the
map plate and picks up the analog signal code which represents terrain levels. The speed of the scan is proportional to the variable pulse drive in the X direction and a
shift of 0.01 inch takes place in the Y direction at the end
of each X scan. Thus, continuous map profile data are ob-

The I-inch magnetic tape storage is used in the system
primarily to provide a permanent storage medium for
model information. This method of storage is destined to
replace present methods of storing heavy molded models.
A single reel of tape, representing a portion of the earth's
surface at a set scale, may be used to reproduce threedimensional models at various horizontal scales and various vertical scales or vertical exaggeration. Also, the same
tape may be used to reproduce spherical model sections or
flat models to various map projections. Thus, this single
tape may be used to reproduce anyone of possibly 25
master terrain models.
THE OVER-ALL MASTER MODEL SYSTEM

Fig. 3 shows the diagram of the system. The prepared
metal map is placed in a scanning mechanism and the timing generator causes the X and Y servos to position a
scanner head to appropriate scan the map plate. The X and
Y servo-positioning information along with the Z-cocie
information is passed into a computer to convert the map
projection coordinates into spherical coordinates which
match the curvature of the earth's surface. The data are
then passed on to the memory device which is a magnetic
recording system. The magnetic tape is then used to drive
the model cutting mechanism through a counter storage
medium feeding the appropriate three-dimensional servodrives. The cutting tool is a high-speed routing tool which
is positioned in three dimensions by the servosystem and
thus produces a three-dimensional model.
DESIGN FACTORS OF THE SYSTEM

(See Fig. 4.) The three-dimensional models will be
constructed from Hydro-cal plaster blocks approximately
30 X 30 X 3 ~ inches thick. Hydro-cal plaster has been

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

==~

~lll
x

==~

SCAN

TABLE

Fig. 4-Typical model section.

used in the system in combination with inert agents and
lubricating agents to produce the model blocks. Because
of the dimensional stability of plaster, its good machining
properties at very high speeds, and the ease of reducing
residue to small dimensions, it provides a good material
for the system.
Profile cuts will be taken at each 0.01 inch. A cutting
speed of 30,000 to 50,000 rpm is required to maintain deflection within tolerance limits and reduce the size of residue chips to micro dimensions for ease of residue evacnation.
Maximum cutting depth is 2 inches, and accuracy is
maintained to 0.01 inch in any dimension.
Digital coding is used to obtain more accurate computation for projection corrections, interpolation of profile
coordinates, and storage recording. All the standard map
projections will be incorporated into the system, both in
the map scanning portion, and in the model input section.
Thus, flat models to various map projection coordinates
will be produced as well as curved models simulating scale
curvature of the earth.
THE CONTROL SYSTEM

(See Fig. 5.) The control system uses serial type SEAC
circuits. A I-inch wide magnetic tape is used for information storage utilizing binary pulse coding of 3 channels,
thus allowing 4 separate runs on the tape to use 12 channels. Ten bits ar~ used on the tape for each position or
word. The scanning/recording portion of the control system is programmed by means of a timing pulse generator.
This generator impresses a timing pulse channel on the
magnetic tape and pulses the X and Y counter units. The
counter units in turn operate the X and Y servos for scanning the map plate and feed the binary positioning code
for the X and Y channels on the tape. A single channel
on the tape is used for both X and Y positioning code,
since a single profile X scan produces no Y change until
the end of the scan run. Then the Y increment shift takes
over on the same channel using 4 pulses of time on the
channel to effect the O.Ol-inch physical shift of the axis
and a new X profile scan then proceeds on the same channel. While this XY scanning operation is proceeding, the
Z-code information is being recorded on a third channel
of the tape.

Fig. 5-Recording system block diagram.

TO
HYDRAUUC
DRIVES

Fig. 6-Reproduction system block diagram.

(See Fig. 6.) The tape output drive system is actuated
by the three-channel pulses of the tape unit. The XY
channel activates the X and Y counters alternately which
energizes the X and Y servodrives. The timing pulse
channel controls the register and the Z-dimension channel
activates the Z servosystem concurrently with the· XY
systems.
THE MACHINE SECTION

The mechanical portion of the equipment consists of a
heavy base table which carries the three-dimensional
drives, the milling head housing, and the residue exhaust
system. The hydraulic power supply unit provides power
to the hydraulic motors at approximately 3000 psi.
The three-dimensional servocontrol system controls
hydraulic motors which turn ball-screw actuators to position a milling head carriage in the X and Y positions.
The third Z motion is accomplished by hydraulically activating a piston which positions the milling. head drive in
vertical motion. The milling head, which rotates a tapered
fluted tool at high speed, produces a profile cut of approximately O.Ol-inch width and varying in depth as the drives
produce the profile motion.
The combination of precise tolerance "ball-screw actuator drives" and the fast acting hydraulic motor drives provides accuracies of speed and acceleration in excess of the
original requirements.

Stieber: The Master Terrain Model System
CONCLUSIONS

The research program outlined in the foregoing paragraphs has been geared to meet the over-all requirements
of the master model program. This research program has
been divided into three phases as follows.

Phase I
This phase was devoted to a detailed analysis of the
problems in the form of a design study. This included
studies relating to mapping and cartographic techniques,
data programming, data storage, machining or forming
methods, and systems study to correlate these units into
a workable design.

Phase II
Preparation of design drawings, specifications, and the

Discussion
In answering the questions on the model
terrain system, Mr. Stieber had with him
R. E. Hock, of Technitrol Engineering
Co., Philadelphia, Pa., which is reducing
this technique to practice.
J. S. Seely (Southern Railway Systern): Can overhanging cliff configurations
be handled?
Mr. Hock: We can handle anything up
to approximately 85 degrees; anything
greater than that is so vague in out- mapping that we don't worry too much about it.
E. L. Harden (Westinghouse Electric
Corp.): Is there interpolation between
contour lines or does the cutter make steps
corresponding to the contour lines?
Mr. Hock: In one method of coding,
which is to etch away the areas between
the contour lines, we coat the map with a
conducting paint, a resistive paint in which
you get contour smoothing between the
lines. We actually convert this smooth
analog voltage to digital steps between lines
so that if you are cutting an exaggerated
scale, you do get digital changes between
your contour lines. In the method in which
you are etching away the lines and leaving
the areas, there is no easy way of interpolating between contour lines.
Mr. Ebeling ('Otis Elevator Co.):
What is the maximum map scale ratio?

construction of a prototype model of the design to prove
feasibility of the over-all system were covered here.

Phase III
The final phase included the addition of complete computer control to the prototype model and reworking of
components as necessary to carry the project through the
various research phases, which were proceeding in close
conformity to the original program schedule.
This project is believed to be the first step in a series
of completely automatic devices for the correlation of
cartographic and photographic intelligence to maps and
models.
It is expected that this equipment when used for model
production will consist of scanning units located in a
centralized location, with model production units at various
activities concerned with model usage.

Mr. Hock: On the XY axis, they are
in the ratios of one half, 1, 2, and 4, in the
Z axis, one half, 1, 2, 4, and 8. It was
easiest to obtain binary values at some
later date. These will probably be converted
to decimal.
Mr. Maetra (RCA Labs., Princeton,
N.J.): What is the limit of the smallest
change delta Z that can be recorded? Is
this comparable with the precision that an
operator can obtain manually?
Mr. Hock: The smallest increment of
movement in any axis is five thousandths
of an inch and it is better than manual
accuracy.
Question: How do you take into account the cutter center offset correction?
It would appear that no correction is made,
in which case the three-axis part will be
in error.
Mr. Hock: The cutter is a tapered tool,
tapering. down to fifteen thousandths at its
tip and has a fifteen-to-one aspect ratio.
So, the offset is approximately seven and
one-half thousandths. Now, we are trying
to obtain accuracies of plus or minus one
one-thousandth of an inch so we have almost really taken up our accuracy in the
cutter offset.
However, we do ignore it, as you have
stated.
Question: When will the system be
producing three-dimensional maps?

Mr. Hock: We actually have the Z axis
at the plant. We are working on the servosystem at the present time and the X and
Y axes are under construction at a subcontractor. However, we have a prototype
of the XY axis which is a converted milling
table that gives us limited travel of approximately three by six inches and this
we hope to be operating in January, 1958.
Mr. Winslow (ABMA): In the preparation of the original map for photographic
work, do you use a color or line to indicate the contour line?
Mr. Hock: The original map is a black
line map and it is a transparent negative
with the contour lines in black and from
this we produce the etched-either the area
map or the lined map.
Question: What are the common contour intervals used?
Mr. Hock: These vary from ten feet
up. to several hundred feet, depending upon
the horizontal scale of the map. The elevations are stored as earth-centered spherical
coordinates.
Question: Do you use the output computer to convert to flat proj ection ?
Mr. Hock: Actually this conversion of
the coordinate system is the next step in
the program. Presently, we are working 011
using the existing projection in reading
and storing in that projection and cutting
in that projection.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

A Coordinated Data-Processing System and Analog
Computer to Determine Refinery-Process
Operating Guides
c.

H. TAYLOR, JR.t

GENERAL DESCRIPTION OF THE SYSTEM

HE coordinated data logging and computer equipment, which has been constructed for Esso Standard
Oi1's Belot Refinery, Havana, Cuba, measures and
records the true value of 101 process variables and 11
operating guides. The measured variables are gas flows,
liquid flows, level, pressure, oxygen percentage, and temperature. The inputs to the logger representing flows, level,
and pressure are in the nature of 3 to 15 psig signals. The
oxygen percentages and temperature signals exist as dc
mv signals. Unique in the system is the incorporation of an
analog computer which calculates 11 operating guides.
These are computed at the end of a readout cycle from the
logger. The operating guides computed are the following:

T
r

1)
2)
3)
4)
5)
6)
7)
8)
9)
10)
11)

Carbon burning rate,
Catalyst-circulation rate,
Catalyst-to-oil ratio,
Ratio of feed-to-reactor catalyst hold-up,
430 0 FVT conversion-corrected,
Percentage weight of hydrogen in coke,
Per cent weight of carbon make on total feed,
Heat duty of top pump-around system,
Heat duty of mid pump-around system,
Regenerator superficial velocity,
Reactor superficial velocity.

The logger can be adjusted to give a complete readout
of all process variables every 10 minutes, 30 minutes, or
hourly. At the end of the hourly logging cycle, information required by the computer is fed into the computation
circuits which proceed to calculate the 11 operating guides
noted above. As the computation for each guide is completed, the true value is logged on the output typewriter.
Thus, a given set of guides is based on information given
in the logging cycle immediately preceding the computation
period. Scheduled logging is supplemented by "on demand" logging and can be initiated at any time by the
operation of a push button.
Two electrically actuated automatic typewriters with
separate tape punch have been furnished to record the
outputs of the automatic logger and analog computer. The
ta~e punch produces a 5-channel punched-tape output
smtable for the actuation of IBM equipment. In order to
provide ~or proper utilization of this tape, the following
mformatlOn accompanies the data:
t Fischer and Porter Co., Hatboro, Pa.

1) A symbol to identify "on demand" readouts.
2) Tabulating card-advance and card-eject signals for
each group of 15 data points.
3) Tabulating card number for each 15 data points.
4) Time to the nearest minute for each tabulated card.
5) Unit-identification number for each tabulated card.
6) An identification character on each card in order to
differentiate preset hourly readings which are of interest to accounting and other groups for manual or
more frequent readouts demanded by the operator
but not essential to tabulating card computations.
7) Two additional identification characters on each card
in order to identify ultimate data users.
The equipment is housed in two cabinets in the follow""'
ing manner. One cabinet contains the transducing equipment for the process variables, and the programming
equipment for the data logger; the second cabinet houses
the computing. circuits. Both cabinets are arranged to
permit a continuous air purge. In addition, the computer
cabinet was provided with an air-conditioning system in
order to dissipate the electrical heat generated by the
computing elements. All equipment has been designd to
conform with specifications for a Class I, Group D, Division 2 area as defined in article 500 of the 1953 National
Electrical Code.
I t should be noted that among the process variables being logged are three gas flows which have been pressure
and temperature compensated. All flows are printed as
hourly averages and 24-hour totals. Five of the temperatUres are printed as hourly arithmetic averages, and 61 are
printed as instantaneous values. All necessary linearization of thermocouple inputs, the conversion of customers
3-15 psig pneumatic signals to digital signals, and the extraction of square-root functions have been provided in
order to check the over-all accuracy of the automatic logger. These points are printed out on the log sheet before
each readout cycle. A dead weight loaded precision pneumatic comparator has been provided to check the pneumatic-to-digital transducer which is used for the pneumatic-signal inputs.
The accuracy for the logged-process variables is as
follows:
1) Temperature -I- 0.25 per cent.
2) Pressure -I- 0.5 per cent.
3) Compensated flows -I- 1.0 per cent.
4) Integrated and averaged flows -I- 1.5 per cent from
10 to 25 per cent of full-scale flow, -I- 0.75 per cent

Taylor: Data-Processing System for Refinery Process

from 25 to 50 per cent of full-scale flow, and -+- 0.5
per cent from 50 to 100 per cent of full-scale flow.
The accuracy of computation for the operating guides
is -+- 2 per cent.
USE OF AC COMPUTER SIGNALS

Because the computer must be capable of operating at a
100 per cent duty cycle, that is to say, 24 hours per day
for 365 days per year, the signals handled by the computing circuits are in the nature of 60-cycle ac voltages. This
was done in order to produce a high degree of reliability
in the equipment. Most conventional ~omputers today utilize signals which are dc voltages. In order to obtain the
best possible degree of accuracy and minimum drift, it is
necessary that all dc signal voltages be checked for proper
calibration before the problem is solved. This becomes a
frequent maintenance procedure in the case of a continuously-operated piece of equipment. In order to eliminate
the need for a periodic check of all signal-voltage accuracies and amplifier balance, the signal voltages in the computer are obtained from a group of transformers. These
have been designed to operate on a primary voltage of 220
volts, 60 cycles. In practice, the transformers are operated
on 115 volts, 60 cycles. We have therefore supplied twice
as much iron in the laminations as actually required. This
insures that the transformers operate on the linear portion
of their magnetic-characteristic curves. This is done so
that we are certain that the coupling flux within the core
does not approach the saturation level and'that the secondary voltages contain a minimum of distortion and bear a
constant proportionality to the primary voltage. The transformers are wound with a turns-ratio accuracy of -+- 1 per
cent. The signal voltages are then padded to the desired
accuracy by means of adjustable rheostats connected in
the secondary circuits. In order to further insure repeatability between the units handling a multiplicity of signals,
all transformers are loaded equally. Furthermore, they
were constructed with laminations stamped from a uniform batch of iron in order to further insure repeatability
among units. The result is a signal-source module of five
volts, 60 cycles. When signals of greater magnitude are
required, additional transformers are used with their primaries connected in parallel, and their secondaries connected in series.
For example, if a signal of, say, 22.00 volts were required, we would supply five transformers with their primaries in parallel, and their secondaries in series. This, it
can be seen, results in a "voltage stick" which is 25 volts
long. We take the signal from the low end of the voltage
stick to a point 22 volts from the end.
This reasoning was applied since a computer operating
with signals which were obtained as described above will
require no standardization or reference to a secondary voltage standard, nor will it require any special regulation of
line voltages. It is felt that this results in a more reliable system of equipment, and restricts computer down-time to a
program of preventive routine maintenance.

TYPES OF COMPUTER INPUTS

The information fed into the computer is derived from
several sources as follows:
1) Logged during a readout cycle and stored as a shaft
position prior to the start of the computation period,
2) Derived during the computation period and stored as
a subroutine for use later in the computing cycle,
3) Fed into the computer manually.
Fig. 1 shows a representative segment of an information
time-flow diagram. As can be seen, the computer solves
the 11 equations in sequence, beginning with (1) and proceeding to (2), (3), etc. In order to solve a particular relationship, it is necessary to obtain the values for all variables involved in the equation. As the diagram indicates,
these may consist of parameters which have been logged
during the previous readout cycle, variables which· have
been manually set, or those which have been stored as a
subroutine during a previous computation. When all of the
necessary data have been assembled, the information is
fed into the computer, and when the output device has
come to balance, the desired solution is printed out on the
logger typewriter. The computer programming circuits
then proceed to assemble the data necessary to solve the
next equation.

ill

C8~~,

~ K

__ B

----0-

-:1-

--B

c -ID- -:~i--c
Dl ---rn- ~
~ -0,

~"Ol

o~-®-.t'+ ~--Da

;:; I

:~~_~_=-:~= _= ~ ~ l:I:~~~
Illm~Q

I
I

II'
, I I [g~ ..

,I:
L __ .'TO

10<:(10

1'1 Eil
[ I 1- -

T.
y

__ ~o E.G('s 6 tIC

I
, 'L - - _

TO EG/'$

'---: --TO EQ'S

[g=

LOC,C,ED

=M AI'-IUAL

$ 10

6,7~IO

VARIAeL~ (I\I'-IALCC, II--\PUT).
AI4AlOG. INPLII.

AN ALoe, ,0 OI"'IiAl CONVEl<,'510N AI'-IO PI<,IN,OU, OPE.RATIO.i'\,

[B = sueROUilNE

SI0R.AG,E OPER.AilOI'<,

Fig. l-Information time flow diagram.

When this information is available, it is fed into the
computer circuits which proceed to solve the next relationship, etc. When the last operating guide, or (11), has
been solved and typed out, the computer circuits are disengaged from the logger for a period of one hour at which
time the next computation period is initiated.
A

TIME-SHARED GENERAL-PURPOSE COMPUTER

In order to perform the necessary computation with a
minimum of equipment, it was decided to assemble a
group of computing elements consisting of algebraic SUmmation elements, coefficient modules, and electronic multiplication or division elements, and to program their input
and output circuits to solve for the value of each operat-

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

ing guide serially. This means that a given computer building block may be used repeatedly with a minimum duplication of equipment. This time sharing of computer "elements" makes it possible to assemble a so-called general
purpose computer and, in effect, makes it a specific purpose machine through programming. To determine the actual number of computing blocks required, a study of the
relationships to be calculated was made, and a block diagram for each relationship was established. The equation
requiring the greatest number of elements then determined
the diversification and number of moduJes to be included
in the computer. Since all other relationships are of a simpler nature, we use the building blocks over and over
again to solve these less complicated equations.
Fig. 2 shows a typical computer block diagram. The relationship to be solved, in this case reactor superficial velocity, is shown at the top of figure. The input-signal information is shown at the left-hand side of the figure. The
block diagram indicates how the various signals are modified and combined in order to solve the desired equation.

computer relay programming is used to accomplish this
switching.
Fig. 3 shows the circuits for a typical time-shared summation amplifier. In this case, we desire first of all to add
voltage El to E3 in one equation and then, at some later
time, to add voltage E2 to E4. It will be noted that the input signals are connected to the amplifier by means of
computer programming relays with contacts Rl and R2 2S
shown. The signals are switched by Form D (or make
before break) contacts which insure that the input to the
amplifier is not open-circuited during the switching operation. The stage gain of this particular type of amplifier is
made equal to unity by suitable adjustment of the feedback
resistor RF and the input resistances R r - 1 and R r - 2 • Furthermore, it can be seen that, to minimize any dc drift
within the amplifier itself, a chopper-stabilizer module has
been included. Should it become desirable to check the dc
output level as referred to the input, a test switch has been
provided as shown. This switch decouples the amplifier
from any existing input signals and returns the summing
junction to ground through a suitable resistance. A dc volt
meter can then be placed from the output of the amplifier
to ground, and potentiometer RA manually adjusted to
reduce the dc unbalance to zero.

caR.
MW

--------=-:-1.
r-r-"O----T--1

r-- I
r-r--<>--1I---+--+-I

e OI

~<
e01

1--1
e~

5io+T.
336AR(l4.7+PR)
MLl-l
p~

~=ALGiEa!(A\c SWMMI\.l"IONOf'ER.

\4.1

MU=!v\ULlIPlICAilOI'ol OPERA1\ON.

tlV=OIVISIOt-l CPFRAl"\ol-\.
C =FIXEO .COEFFICIE\o.I1 OPIOR.

100
~B600~-"---------'

Fig. 2-Typical computer block diagram.

e~I=("I-te.)

eo.. = (....+e4)

The references shown in the diagram refer to the analog
(5)-®=COMl'lJiE~ I"ROG,RAMM1S.R. R£\./I.'{ COlolil\CiS
signals as ac voltages with respect to ground, that exist
Fig. 3-Summation amplifier.
at the various points in the circuit. The equation is generated, so to speak, by starting with single terms which are
Fig. 4 illustrates how these same principles are applied
combined by addition, subtraction, multiplication, and division until the desired results are achieved at the output to an operational amplifier being used as a coefficient mod(shown at the right-hand side of the figure). It was neces- ule. That is to say that the input voltage is multiplied by
sary that we know the variation in magnitude for each of a fixed constant depending upon the position of the slidthe individual input signals so that we could assign appro- .ing contact on a multi turn potentiometer associated with
priate signal voltages to these inputs during the process of the desired constant. Here again, it will be noted that the
switching operation is accomplished by Form D contacts
scaling the computer parameters.
(or make before break) so that neither the amplifier-feedTYPES OF COMPUTER BUILDING BLOCKS
back loop nor the input-summing junction is left openIn order to time share the various computer elements, circuited during the switching operation.
The chopper-stabilizing amplifier is again included,
it was necessary to switch a given module between the several circuits in which it was to be used. To illustrate how as well as the test switch, should it be desirable to check
this was achieved, diagrams are included showing how the the dc level of the output with respect to the input.

Taylor: Data-Processing System for Refinery Process
Further reference to the diagram will show that the polarity of the input signal determines the alegbraic sign of
the term it represents. Thus, if we desire to add two signals together, we arrange to feed voltages of like polarity
into the summing amplifier. If we desire to take the difference between two voltages, the signals are fed 180 degrees
out of phase. This simply means that the secondary leads
from the signal transformer are reversed, resulting in the
necessary phase reversal. It should also be noted that the
operational amplifiers used invert the signals fed to their
inputs; that is to say, if a signal with polarity 1-2 is fed into
an amplifier, the output voltage will have phase 2-1. It was
therefore necessary to study the relationship being solved
in order to determine the phase relationships between the
various input signals.

e,= .107 A

MAY. !:>I'"

IHPU1~\

fLiNCTICi'-I SWI1CH

OW1PLl1Q

CHOPPER STII.\:>ILlZED
ELE.CTI.);)(.'0'BI/'M)

= COMPU1EI<

OWTPllT

FCRMl.>LlIPL'f

/>,MP\.IFIE.R.<;,

PRO,,-RA"'IIVIER Rl:.LA,{

CON1AC1~.

Fig. 5-Multiplication and division circuit.

means for manually adjusting any dc unbalance which may
occur in operation.
I--II@
e,

MEMORY SYSTEMS

<.---1

--I

0il

~---1

iE51'

@
+D.C.

@
1M

lOOK

TEST

@
p::- <.
eo.

f--I

~z
eo;!.

f--I

POli'-l;

e02.=~e'Z.
MMER. RUP-'f C:Ol'n .... C"Ts.

Fig. 4---Coefficient amplifier.

Fig. 5 illustrates the method of handling an electronic
multiplier-divider element using ac analog signals. Since
the device is a true multiplier, it was necessary to avoid
the sine-squared relationship which results when two ac
signals are fed to the inputs of the module. It will be seen
that one signal is converted to a dc voltage by means of a
full-wave rectifier and forms the input applied to terminal
No.2. This voltage is then modulated by the signal applied
to input No. 1. The output signal is proportional to the rms
product of the input signals in the case of multiplication,
and to the ratio of the rms amplitudes of the inputs in the
case of division. Contacts on the computer programmer relays determine whether the block will be used for multiplication or division. Auxiliary operational amplifiers CD and
CM were provided in order togive the proper output voltage relationships. All amplifiers within the multiplierdivider were chopper stabilized in order to minimize any
dc drift within the element itself. The circuit shown was
arranged to provide a multiplication and division element
with unity gain. In order to increase reliability and minimize drift, it should be noted that all computer building
blocks were provided with chopper stabilization and also

Information stored in the computer is derived from 3
sources:
1) Information stored as a shaft position within the
logger itself.
2) Data which exist within the logger instantaneously
during a readout cycle, but which must be maintained for use as a computer signal.
3) Information stored as a subroutine during the computation period.
Process variables which are read as integrated quantities such as compensated flows and hourly average temperatures exist as shaft positions in the logger. Fig. 6
shows the circuit for a typical compensated gas flow. In
this arrangement the position of the contact on the accumulator slide wire is a function of the flow which has
been pressure and temperature compensated. During the
logging cycle, the flow accumulator slide wire is connected
to the logger readout device. During the computation cycle,
however, an ac voltage from a computer signal transformer is impressed across the terminals of the slide wire, and
the voltage developed from the contact to the low end of
the slide wire is returned to the computer as an ac signal
for use in the computation circuits. It can thus be seen that
the accumulator slide wire is time shared between the logger circuits and the computer circuits. Moreover, there are
computer parameters which are derived from signals that
exist in the logger only instantaneously as the logger programmer scans the customer input information. For example, instantaneous temperatures, as given by thermocouples, must be stored at the time that they are read during the logging cycle and must be maintained in memory
in a form which. can be used by the computer. This was
accomplished by means of a small electromechanical servosystem. The input signal from the thermocouple is fed to
the logger readout device. A retransmitting or follower
potentiometer mounted on the shaft of the logger readout
device transmits a dc voltage to a storage' servosystem located in the computer cabinet. The shaft of this servo then

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

'3-I~*'~IGo"'~L FgOM
'fl.OW iRA~~DUC~R

3-I?W~\4Nl\L FROM
FLOW i~Al-\SDUC~R

M.V.CS\6NAL FROM
iHERMO~OUPLE.

-------4~--------------~\lrl----------------~----~
ACCUMlJLAiOR
MOiOR.

A.C.
LI~~

OP!!.R~il~'"

where the signal does not go to zero at the minimum range
of its variation. For (3), potentiometers P 3B and P 30 are
used in an arrangement similar to that shown for (2). In
this case, potentiometer P 3B is shown as a retransmitting
)or follower potentiometer mounted on servostorage mechanism SV1. The signal to be used in (4) is generated across
P 4B which is shown as a potentiometer mounted on one of
the compensated flow accumulators.

~--------------------------------~M~------------------------------~

GEARIl'lG, :

I2Q

ACLIHE

I
I

'-0 A.C. UNE.
{ACCLJMULAIOR' GCARIN6:- - - - ;
VIA LOC:::;~E.R.
RE.SE.i
\VI
I
PROGR.t>.MMER., ,MOiOR
,
I

--=--k8J'
®

t.oc:$(:,E~

'
t

'TO
READOUT
DEVICE. VIA LOGGE~

PROGRAMM;::R..

:®

-----------------IT'_----.

lOGGERRA~E

AOJUSiMEt-li

,:9 C9 (§l
TI'o/V"V'oNV\'\I\A-

ACClJN\LI L~il OR.

iO COMPUTER.
VIA COMP liT ER
?RO~RAMME_R

{

'SLIDE WIR.E

.....--------------'T'-----------------'f-------J
...------IT"\.....,.-----'f'--.--,

~
TO EQitl

(5)

T
I

~
TO E.G! "*'2

T
I

® @ ®

~
TO"Q ~5

T2.

T
I

~
TO EQ ~4

ACC;'COMPEJ.I5ATED FLt;)W ACCUMULATOR. "OTE~ilOMaiER. •.
SVI =STO"RA~E. SERVO RETIiiANSMl-'-Tlto\~ POTE~TIOMEiER.
/VI = MAt-Il.lALW saT Ii'\PUi ,"OTENTIOMEiER.
1-2. =IN!1>TAN'ANEOI.IS 'PO,-ARIT'{ t)E.SIG,NA.iION.

COM ?Lli ER,
RA~"E.

A'i>jUSTMENl

® = COtv\PUiER PROGRAMMER

=
CJ
""">FORM.,,
COMPl.I'TER 5Iq~AL

RE.LAY CONiAC'TS.

A,C.lIN~

Fig. 6-Typical compensated-flow circuit.

assumes a position proportional to the'value of the input
temperature being measured. A follower potentiometer
mounted on the shaft of the storage servo is then used as
a computer input information by impressing across its terminals an ac voltage from a computer signal transformer.
Subroutine storage within the computer itself is
achieved in a similar manner. The ac voltage representing
the quantity to be stored is fed to a storage servo whose
shaft again assumes a position proportional to the stored
quantity. Retransmitting or follower potentiometers
mounted' on the shaft of the storage servo are then used
in subsequent computations.
SIGNAL SOURCES

The signal-source transformers were treated in a manner similar to that of the computer building blocks themselves. A single transformer is time shared between the
circuits comprising several of the relationships to be handled by the computer. Fig. 7 illustrates this point. The
diagram shows a single signal-source transformer which
has been time shared between four equations. The signal
to be used for (1) is generated across potentiometer P 1B
which is shown as a manually-set variable. The voltage to
be fed for (2) is generated across potentiometer P 2B in
series with rheostat P 20. Rheostat P 20 has been included
in this case to provide zero suppression for the instance

=cOM"UTE.R f'ROG,"RAMMER RELA'I CON'TACiS.

Fig. 7-Representative signal-source diagram.

Potentiometers P 1A, P 2A, P 3A, and P 4A have been included as range adjustments set so that the voltage developed across the signal potentiometers is of the correct
magnitude, depending upon the scaling desired for a particular variable. Computer programming relays operate
contacts R1, R2, R3, and R4 which are associated with
equations (1) through (4), respectively. The relay contacts
in the primary circuit of the signal-source transformer are
used to reverse the phase of the transformer secondary
voltage. As mentioned previously, an instantaneous polarity
of 1-2 is given the designation of a positive signal in the
computer, whereas an instantaneous polarity of 2-1 designates a negative signal in the computer. Reference to the
diagram will show that the signals used in (1), (2), and
(4) are shown to have a negative sign, while the signal
generated for (3) is shown as a positive signal.
In this manner, a single signal-source transformer has
been time shared with a corresponding reduction of three
in the number of transformers to be supplied. This results
in a saving of physical space within the computer cabinet
as well as in a reduction in the cost of the necessary components.
TIE-IN BETWEEN COMPUTER AND LOGGER

Fig. 8 indicates how the various circuits previously
described were arranged to complete the coupling circuits
between the automatic data logger and the computer. It
should be noted that the addition of the computer in no
way affects the operation of the data logger----it was designed to act as a supplementary or auxiliary device.
As previously stated, the logger programmer scans the

Taylor: Data-Processing System for Refinery Process
-

---I

PR;~~"~~ER
I
_ _ _ _ _ .J

NULL

B/>"lA~CE.

R~LA'(

R~~"E

SELECTOR

POI~T

ADV~~CE

PULSE

Fig. 8-Coordination between logger and computer.

process analog-input information at least once an hour. At
the completion of the hourly logging cycle, the logger programmer sends a signal to a stepping switch located in the
computer programmer. This signal advances the stepping
switch to position no. 1 and initiates the computation cycle. Position no. 1 on the stepping switch operates a group
of equation relays associated with the first relationship to

be solved. Contacts on these relays then proceed to select
for the computer input information which has been derived from logger storage data as previously described,
and manual input data. The equation relays also serve to
connect the computer building blocks in the arrangement
necessary to solve this first relationship. The output from
the computation circuits is fed directly to the computer
readout device. This unit is an electromechanical self-balancing ac potentiometer. A retransmitting or follower potentiometer sends a dc signal back to the logger readout
device causing the two servos to track. Since the logger
readout device is a multi range self-balancing potentiometer, it is necessary that range information be supplied
to it from the logger programmer during the computation
cycle. A digicoder, coupled through gearing, is used as an
analog-to-digital converter which sets up a digital output
proportional to a shaft-position input from the logger
readout device. The digicoder output is then fed through
a divide-and-decode network to the logger typewriter. The
characters present in the output are scanned by the character programmer which actually operates the solenoids on
the typewriter mechanism. Tape punch information is generated as the typewriter is recording the value of the computer solution.
During the time that the solution to (1) is being typed
out, a feedback-point advance pulse is sent back to the
stepping switch in the computer programmer which then
advances to position no. 2. The equation relays associated
with operating guide no. 2 are than energized and the cycle'
repeats as described above for (1). The programming continues until a solution is obtained for each of the 11 relationships to be computed. At the completion of (11), a
feedback pulse is generated by the computer stepping
switch and returned to the logger programmer. The stepping switches associated with the logger point programmer
then return to the home position where they remain until
the initiation of the next logger cycle.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

System Characteristics of a Computer Controller
for Use in the Process Industries
W. E. FRADyt

'
E

AND

INTRODUCTION

EFORE the detailed design of a computer may be
begun, it is necessary to set up some fairly detailed
specifications which define the operating characteristics of the proposed system. It is the purpose of this
paper to show how these specifications were developed for
the RW-300 digital-control computer, the first computer
designed specifically for process control. The specifications for this computer were developed as a result of a
number of studies carried out on specified industrial processes over a period of almost three years.
.

THE PROCESS CONTROL PROBLEM

The functional and environmental specifications for a
digital system arise explicitly or implicitly in answer to a
number of questions which can be raised about the job the
system is to perform. The job to be done by a processcontrol computer is that of making adjustments in process
variables to attain some specified process objective in the
face of variations (in raw material characteristics, ambient conditions, etc.) over which no control can be exercised. In the course of answering questions about the process-control jobs we will show how the basic specifications
for the RW-300 were developed. A description of the computer will complete this paper.

Inputs
What is the source of the information, and how may it
best be translated into the language of the machine? The
data entered into a process-control computer is fundamentally of two different kinds: data from process instruments (measuring temperature, flow, pressure, liquid level,
chemical composition, viscosity, density, etc.), and data
supplied by the operator as requests for special operations
from the system, or as changes to be made in system operation. The instrument data is fundamentally analog in
character, normally in the form of an electric or a pneumatic signal. The signal may continuously represent the
quantity being measured by the instrument (as a thermocouple voltage continuously represents the temperature),
or it may represent the physical quantity being measured
only at intervals (as the output of a chromatograph represents a composition by a peak voltage or by the integral of
a slowly varying voltage). In addition to these analog instrument signals, there may be digital signals designating
the mode of operation of the instrument.
t The Ramo-Woodridge Corp" Los Angeles, Calif.
t The Thompson-Ramo-Wooldridge Products Co Los Angeles
'"
Calif.

PHISTER~

Although the analog input information could be transcribed manually by an operator and inserted into the computer control system in digital form, this would be an inconvenient slow operation, subject to human errors of
transcription.
The digital information inserted by the operator is most
conventiently presented in decimal form, so that the operator can prepare it easily. However, it may be desirable
to permit the operator to initiate special requests and
changes by pressing a button or operating a switch.
What kind of information must be represented, numericalor alphabetic? As can be inferred from the answer to
the first question, the information read into the computer
is fundamentally numerical in nature. The precision of the
input data is limited by instrument precision. A precision
of better than 1 per cent of full scale is unusual in common process instruments.
What is the rate of flow of information from the
source? Data from the continuous type instruments are
available at any time. From instruments like the chromatograph, data is available only periodically. Typically, a
chromatograph signal might be sampled periodically every
ten seconds over a period of five or ten minutes. Although
part of the instrument data is available continuously, the
variation in process variables is usually slow enough that
the computer control system need to sample the instruments no oftener than once every five minutes or so. The
number of instruments which must be so sampled varies
from process to process and from problem to problem. A
typical complex process may require as few as 25 inputs
or as many as 250 inputs.

Data Processing
What must be done to the information? Must it be altered; if so, how? A process-control computer must be
able to handle many different kinds of computations and
manipulations of process input data. Typical of the kinds
of data processing required are data interpretation, calculations for optimal control, data logging or printout of
process information, and checks for hazardous process
conditions or for instrument failures. By interpretation of
data we mean the translation of readings from process instruments into numbers corresponding to the physical
quantities these readings represent. This interpretation
may be as simple as the application of a scale factor. It
may include a linearizing operation like one which must
be applied to a thermocouple to translate voltage into a
temperature. Or it may require the solution of a set of
simultaneous linear algebraic equations, as are required to

Frady and Phister: Computer Controller for Process Industries
compensate for interferences between a number of chemical compounds analyzed by a mass spectrometer. The calculations required for control in general require the solution of very nonlinear algebraic equations. These equations vary widely from process to process, and there seem
to be no characteristics common to all of them. The datalogging operation requires that information be printed out
in a digital-decimal form. The checking of process instruments and process conditions for malfunctions requires in
general the comparison of observed or calculated data with
certain standard values established by the operator. These
comparisons and the decisions based on these comparisons
comprise the checking operation.
How much time is available for processing the information? The data interpretation and control calculations need
to be carried out at a frequency determined by the dynamics of the process and by the rate at which significant
changes take place if the variables are slow and the time
constants involved in a process are fairly long. A control
calculation once every five minutes to once every hour or
half hour is sufficient for most processes. The data-logging
calculation may also vary depending on the state of the
process and the desires of the operator. A complete logging operation of all process variables and related data
once every five minutes is most frequently required. The
alarm-checking operation may again be one which should
be carried out very frequently or relatively infrequently
depending on the variable in question. Some process
checks must be done once a second; others can be carried
out as infrequently as once every fifteen minutes to an
hour.

Outputs

output data. Typically, the number of process outputs is
at most about half the number of inputs.

Reliability and Maintenance
What effect would a machine error have on the information flow, and how would it affect the operation being performed? May the operation be interrupted for emergencies or for regular periods of preventive maintenance?
Conventional instruments used in the process industries
are typically very reliable. Operation over periods of several years without preventive maintenance and without
failure is not uncommon. However, every process instrument has some probability of failure, and the process engineer, in designing a control system, takes this into account and assures that the control system is fail-safe, that
is, that a failure of some instrument or even some combination of instruments will not result in a disaster. A digital-control computer, by its very complexity, is not likely
to be as trouble free as conventional and very simple
pneumatic instruments. However, as might be expected in
view of the fail-safe precautions commonly taken, the
problem involved in incorporating a digital-control system
into a process is different in degree rather than in kind
from the problem of installing the conventional instrumentation. The entire control system must still be failsafe, regardless of the reliability of the computer.
Nevertheless, the practical effectiveness of a control system of this kind is very much dependent upon its reliability, and upon the ease with which it is repaired when failures do occur. High reliability and great ease of maintenance are very important. In addition, it must remain
practical to do preventive maintenance with the least possible interference with normal computer operation.

What must be the output rate of the process information? The rate at which output adjustments a~e made on Environment
the process is again a function of the dynamics of the
What are the environmental conditions under which the
process and the rate of change of the uncontrollable varia- system must operate, and what effect should these have on
bles. In most processes an adjustment once every fifteen the system characteristics? Environmental conditions for
minutes to a half hour is ideal. In some processes more computer control systems in process industries can be expected to vary widely from installation to installation and
frequent or less frequent adjustments may be desirable.
What is the purpose of the output information? What to be very difficult. Wide temperature variations, from beform should it be in to accomplish this purpose most effec- _low freezing to somewhat above 100°F, are frequently entively? There are two principal forms of output informa- countered as are wide variations in humidity. Corrosive
tion, just as there are two forms of input information. The gases and vapors of one kind or another are often presfirst is data calculated by ~;he computer which must be used ent in the air. Large quantities of dust are not at all unto adjust process variables. The second is data in funda- usual. Vibration due to the proximity of heavy machinery
mentally digital form which must be supplied to the opera- can be anticipated. Electrical power supplied to the comtor. The data supplied to the process may itself be of two puter may be locally generated, with the result that line
forms. It may be necessary for the computer to send a voltages and frequencies change by 10 or 20 per cent
digital, on-off signal to turn an instrument on or to effect of their nominal values. And it can be expected that comsome calibration or to open or shut some valve; it will puter use and maintenance will be in the hands of inexbe necessary to make adjustments in instrument settings perienced operators who are unimpressed by delicate or
by means of analog signals which correspond to digital fragile equipment.
numbers calculated by the computer as the proper setting
SYSTEM SPECIFICATIONS
for the instruments. Added to this is the more conventional digital output system required for data-logging purThe answer~ to the above questions resulted in a set of
poses.- As was mentioned before, this requires deCimal rough system characteristics which served to guide the de-

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

signers when detailed decisions were made about the computer's operation. These system specifications may be
stated very briefly as follows.
Because of the complexity and variety of the problems
to be handled by a process-control computer, and because
of the importance of flexibility in these applications, the
basic machine should be a stored-program computer. The
computer should have available a very large memory for
storage of programs of instructions. The important requirement that the system be fail-safe suggests that the
computer outputs, used to adjust process variables, be employed with conventional process controllers which compensate for the second-by-second variabilities in process
conditions. The computer itself therefore need be of only
moderate speed.
An input-output system capable of handling some 250
analog inputs, 100 analog outputs, and a similar number
of one-bit inputs and outputs is necessary. In order that
the process input-output system be suitably matched to
computer speed and to typical process dynamics, all inputs
should be made available to the computer and all outputs
adjusted by the computer at least once a minute.
A conventional decimal input-output system is also re(luired, but since it is not necessary to read in or print out
large volumes of data, relatively low-speed devices are
satisfactory.
Because of the inherent lack of precision in instrument
input and output equipment, it might appear that the great
precision obtainable from a digital computer would be unused in process control applications. It is certainly true
that the ten-decimal-digit word length useful in some scientific computations is not required here. However, a computer precision somewhat greater than that supplied by the
instruments is very desirable to make scaling problems
easy for the programmer. An input conversion system accurate to one part in 256 or 512 (8 or 9 bits) would be
adequate, and a word length two or three times that is appropriate. Since most of the computer's operations do not
require human intervention, the binary number system is
suggested, with decimal output and input conversions
handled by the computer when necessary.
T~~ reliability and environmental specifications emphasize the importance of mechanical and electrical ruggedness and ease of maintenance.
DETAILED CHARACTERISTICS

The R W -300 computer controller is designed to fulfill
all of the requirements above as economically as possible.
Its detailed charact~ristics were worked out by planning
a hypothetical. computer, putting it into typical industrial
process-control systems, and then evolving a more detailed set of specifications while altering the computer's
characteristics to meet new demands. This whole operation
was carried out, of course, with still another objective in
mind: that of providing an ultimate system which would
be .cheap enough to permit reasonable payoff periods in
these applications.

The result, which will be described in the following sections in some more detail, is a transistorized (for reliability) general purpose, stored-program digital computer. It
has a magnetic drum memory (for large capacity at low
cost operable over wide temperature variations) and operates in a serial mode with fixed-length binary words. It
contains analog and digital input and output facilities, the
number of which can be increased or decreased without
affecting the internal logic of the basic computer.
The computer uses an 18-bit binary word for numbers,
consisting of a sign bit and 17 magnitude bits. The magnitude of each number is less than one. This word length is
approximately twice the word length of the analog-to-digital and digital-to-analog conversions and is compatible
with the accuracy of industrial instrumentation. An infrequent number of cases arises in process control requiring
the use of double-precision arithmetic operations. Because
of their infrequent occurrence it was not necessary to provide automatic double precision. These operations can be
programmed through the use of the other instructions in
the machine's repertoire. The instruction system contains
19 instructions of the one plus one address type. That is,
each instruction specifies the address of one operand and
the address of the next instruction to be executed. Instructions are stored in the magnetic drum memory as two
adjacent words. Thus, 36 bits are used to store an instruction. The first word of an instruction pair contains the
operand address and the execution time of the particular
operation. The second word contains the next instruction
address and the instruction code. The execution time field
gives the programmer the capability of specifying the
number of bits used in the multiplication, division, shifting, and digital-input instructions. These 19 instruction
codes can be divided into three categories. The first category contains the basic arithmetic operations of add, subtract, multiply, and divide. The second group contains the
conditional transfer or program-branching type of operations. These are transfer on a zero number, transfer on a
negative number, transfer on overflow, and compare magnitude. The stop instruction is also put in this category.
The third category of instructions contains the data handling operations for loading and storing either of the two
principal arithmetic registers, transferring information
between these two registers, shifting of numbers in the
registers, merging and extracting numbers (logical add
and multiply),· and the digital input-output instruction.
The time required for execution of the various instructions is as follows. This includes reading the instruction
from memory, obtaining the operand, and completing the
operation where a minimum access time is allowed for obtaining both the instruction and the operand.
Add and subtract
MUltiply
Divide
Branch
Load
Store

0.91
2.99
2.99
0.65
0.65
0.78

milliseconds
milliseconds
milliseconds
milliseconds
milliseconds
milliseconds.

Frady and Phister: Computer Controller for Process Industries
These speeds of operation and the command repertoire are
compatible with data-processing requirements and information-flow rates encountered in process-control systems.
The internal structure of the R W -300 can be divided
into the following operational units: a magnetic drum memory, a control unit, an arithmetic unit, a digital input-output unit, an analog input-output unit, a test and maintenance panel, and an operator's control panel.

Memory
The addressable memory consists of 64 tracks of 128
words each. One of these tracks is reserved for a memoryloading program which will load the memory from
punched paper tape. This track is unalterable by the programmer as a safety precaution. A second track of the 64
is used for a 16-word circulating register for fast access.
The remaining 62 tracks of 7936 words are used for general storage. It is possible to write into only eight of these
tracks at a time under computer control, thus giving 1024
words of variable storage in addition to the circulating
register. The eight tracks are selected by means of an accessible connector plug between the writing circuits and
the memory unit. The normally used 8 tracks have both a
reading and a writing head which are separated by 32
words such that a number may be read from a track, operated upon, and stored back into its original address without waiting 'an entire drum revolution. The memory capacity for program constants and variable storage was determined on the basis of programming several typical industrial processes. Additional memory capacity was added
over and above that determined in the study process since
additional use of the computer once installed in a process
will surely be made, thus requiring greater storage capacity for both program and constants.

Control Unit
The control unit of the computer consists of registers
and counters to store information concerning the instruction and the sequence of steps in the execution of the instruction. Two circulating registers on the magnetic drum
are used to store the operand address and next instruction
address. The track selection portions of these addresses
are transferred to a selection register at the proper time.
Sector address coincidence is determined by serially comparing the sector address portions of the instruction to a
sector-identification track on the drum. A third circulating
register on the drum is time shared with the arithmetic
unit but is used to store the execution time of an instruction when used in the control unit. A second flip-flop register is used to store the actual instruction code. Two flipflop counters are used to distinguish the digit times in a
word time and to sequence the steps involved in executing
an instruction.

Arithmetic Unit
The arithmetic unit of the computer consists of two
main circulating registers, the time-shared register men-

tioned above, and an adder. The principal arithmetic register contains the results of instructions and holds one of
the operands in the majority of instructions. The second
arithmetic register is used to hold the multiplier and remainder for multiplication and division. Because the contents of the second register are readily interchanged with
the principal arithmetic register, it can be used as a one
word time, fast access, temporary storage. The time-shared
register is used for multiplication and division. It is not
addressable.

Digital Input-Output Unit
The digital input-output unit contains the basic facilities
for reading in 6 bits from a paper-tape reader and for
putting out 6 bits for a paper-tape punch and/or typewriter. A large number of digital inputs and outputs other
than the paper tape and print inputs and outputs is possible without changing the basic computer. The total number of addresses available for digital input or output devices is 64 and the maximum word length for these inputs
and outputs is 18 bits. These digital input-output facilities
are more than adequate for the alarm output and operator
input instructions encountered in process control.

Analog Input-Output Unit
The analog input-output system does not require programmed control from the computer. All analog-input
quantities are converted to a digital number and stored in
specific addresses in the memory. All analog outputs are
read automatically from the memory and converted to
analog quantities for control. Input quantities are obtained
from the memory by the computer as are any other numbers stored in the memory. The number of analog inputs
and outputs required for process control varies considerably between applications. The maximum number of inputs
could be as high as 512, and the number of outputs could
be as high as 256. Changes in the number of analog in-:puts and outputs used does not change the basic computer.
The input-output system is a cyclic system in that all inputs and all outputs are read during a fixed length of time
and then the cycle is repeated. This makes the latest information available to the computer at any time where the
maximum delay in an input is a matter of a few seconds.
Ten binary digits are used in converting analog-to~digital
and digital-to-analog. The basic full-scale analog inputs
fall into two categories. These are low-level signal inputs
and high-level signal inputs. The standard low-level input
is 0 to 10 millivolts while the higher-level inputs are stand-:ardized at 0 to 10 volts. The latter may be obtained from
instruments which have current outputs sufficient to give
o to 10 volts full scale. The analog outputs are standardized at 0 to 5 milliamperes, although higher current or voltage outputs can be specified without loss in accuracy. A
single converter is time shared for all analog inputs and
all analog outputs, and switching of inputs and outputs is
done both with relays and electronic switches.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Fig. 2-Module bottom etched wiring.

Fig. I-Computer module and insert cards.

Fig. 4-Magnetic drum memory (cover removed).
Fig. 3-Computer subframe holding 14 modules.

Test and Maintenance Panel
The test and maintenance panel contains the facilities
for code checking of programs, marginal testing, and for
on-line trouble shooting of the machine. The normal automatic sequence in executing an instruction is to first obtain the instruction from the memory, then to obtain the
operand, and finally to perform the indicated operation. It
is possible to interrupt this sequence of operations from
the test and maintenance panel and to do these operations
in two steps at a manually controlled rate. The first step,
under the control of the operator, picks up the instruction
, from the memory address and through the aid of a built-in
oscilloscope the operator may inspect the address of the
operand, the address of the next instruction, the instruction to be performed, and the execution time. A second
step permits the computer to pick up the operand and to
perform the indicated operation. The operator may inspect
the results of the operation and note any changes in the
next instruction address. The oscilloscope also allows the
operator to look at the contents of the arithmetic registers
and other important logical points in the computer. Neon
lights are also provided which indicate the contents of the
control registers and certain designated flip-flops.

Controls to adjust the computer voltages and c1ockpulse amplitude for marginal checking are located on the
maintenance panel. Digital input switches on the test and
maintenance panel can be used for program checking
(break-point switches) so that portions of programs may
be checked without running through the entire program.

Operator's Control Panel
The operator's control panel contains the push buttons
and indicating lights to start the computer, to stop the
computer, to turn the power on and off, to resume at the
point of stopping in the program, and to load the memory
from punched paper tape.
The high reliability, ease and speed of maintenance, and
environmental immunity specifications are by far the most
important requirements for any process-control computer.
RW-300 reliability is achieved by using only high quality
components, maintaining rigid quality control on the use
of these components, and derating all components a considerable amount. Conservative circuit-design techniques
with wide tolerances, for component variation, load variation, and voltage variation, are used throughout. The RW300 computer is almost entirely transistorized. Both
germanium and silicon diodes and transistors are used in
the machine.

Otis: Optimized Control through Digital Equipment

Fig. 5-Test and maintenance panel (top), and operator's
control panel (lower left).

The construction of the machine is unitized and modular
in nature for rapid location and replacement or repair of
units, subunits, and components. The building block of the
basic computer is a module as shown in Figs. 1 and 2.
These modules plug into frameworks where the interconnection between the modules is minimized. Each module
contains several flush-etched wiring inserts where the
etched wiring on the module between the inserts represents the majority of wiring in the machine. Components
are also mounted on the module bottoms. Fig. 3 shows the
computer subframe which houses the arithmetic and control unit, a portion of the memory unit, and a portion of
the digital input-output unit. This modular construction
greatly reduces the maintenance time both in locating a
trouble and repairing the trouble. The use of etched-wiring
techniques increases the reliability of the machine over
those using conventional wire and solder techniques. Each
insert and each module contains test points which are accessible when the machine is in operation.
The magnetic drum, as shown in Fig. 4, is a sealed unit

Fig. 6-The RW-300 digital-control computer.

so that corrosive gases and vapors, dust, etc. will not enter.
The construction of the drum is similar to that used in
drums for airborne applications so that temperature variations and mechanical shock and vibration will not alter its
operation. Fig. 5 shows the RW-300 computer with the
test and maintenance panel exposed. Fig. 6 is an over-all
picture of the RW-300 showing its size, which is that of
a conventional office desk.
CONCLUSIONS

Specifications for the RW-300 were developed as a direct result of studies of industrial processes .. Because the
computer was designed specifically for application to process control, it will make available to this application the
flexible, sophisticated computing and control ability necessary to implement integrated control systems.

Optimized Control through Digital Equipment
E.

optimum control of a process we mean the
achievement of a series of objectives in the production of a specific product. The objectives to be
achieved are primarily of an economic nature, although
they frequently are expressed in terms of the process input and output. In other words, the problem is one of
t Daystrom Systems, Div. of Daystrom, Inc., Lo JoUa, Calif.

OTrs t

maintammg product characteristic and quality with the
minimum use of raw materials and at a minimum production cost. Furthermore, the problem includes maintaining
continued product output of specified characteristics in the
face of changing plant conditions and variations in raw
materials, as long as the cost of the product is within the
limits specified by a competitive market.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

At present, the control of a process is inadequately relieved by resorting to the use of "minor control loops"
and a human operator to dose the major loop around them.
(See Fig. 1). These minor loops consist of a transmitterrecorder-controller combination which measures a process
variable and maintains it at a desired level. The control of
this variable is effected with respect to a set point without
regard to the state of the total process or to the value of
any of the other variables. In some cases the value of one
variable and its excursions is used to control another variable in a configuration called "cascade control." However,
this type of control, although more sophisticated than the
simple minor loop, can only be used in the few cases where
a simple, known, and non-time-varying relationship exists
between two variables while the over-all control remains
effectively composed of the series of minor loops. The integration of each minor loop into the whole control system is then effected through the operator, who observes the
process state on different recorders and adjusts (adapts)
the controller to process conditions.
In other words, by adjusting a set point, after he has
observed the state of the process as presented by numerous
recorders, the operator's skill and knowledge of the process
are resorted to in closing a second 10Qop around the "minor
control loops." Through this major loop the interactions of
the different variables are now integrated into the control
system.

RAW MATERIALS
MINOR

LOOf'S

ation of the process, he rarely, if ever, knows within a
reasonable period of time whether this is the best combination, "i.e., whether the process is at maximum effictency.
It is becoming evident that the operator can not attend
the major loop efficiently because of increasing process
complexity. An attempt has therefore been made to close
the control loop through equipment rather than thrQough
the operator. This is not meant to replace the operator but
rather to place him in parallel so that the control equipment can function effectively and sufficiently fast in the
making of decisions. These decisions may be routine; they
are almost invariably numerous. The operator can therefore be relieved of making many routine decisions, and
can intelligently monitor the process and provide over-all
direction in parallel with the computing-controlling equipment. (See Fig. 2.)

RAW MATERIALS

LABORATORY DATA
~ETI~~T(H

T
H

E
COMPUTINGCONTROLLING
EQUIPMENT

I
I

-----. I
L-l-----

I
I

T=TRANsrv1ITTER

R=RECORDER
C=CONTROLLER

PRODUCT

LABORATORY DATA

Fig. 2-0ptimized process control.

o
C
E

S
S

1".
I "",-•.-'

I T=TRANSMITTER
IRoRECOFaR
I C. CONTROLLER

PRODUCT

Fig. I-Process control today.

Because of the complex nature of today's processes,
however, a human operator can not keep track of all the
necessary variables and their interrelationships, as well as
their effects upon product cost and quality. And, although
he is more intelligent than any machine that can be conceived, the human operator fails when it becomes necessary to digest large amounts of data and respond with adequate speed to an increased system complexity.
Furthermore, even though he might achieve an apparently satisfactory combination of settings for the oper-

Since the computing-controlling equipment is placed in
parallel with the operator, it must be capable of communicating with both the operator and the' process. Thus, it
must be able to accept signals directly from the transducers and/or transmitters, as well as from the Qoperator, and
its output must be recognized by, and be intelligible to, the
controllers and to the operator.
'After considerable study of various processes, a system has been designed to close the loop around minorloop controllers. This system centers in a general-purpose
digital computer and, as is the case with any other computing system (industrial or military), the peripheral
equipment becomes important. So besides the computer
we have system components such as input multiplexer,
analog-to-digital converter, computer input-output equipment, and CQontrol output.
Although many hours could be spent in discussing the
characteristics of these component systems, at this time we
shall content ourselves with indicating their main design
criteria.

Otis: Optimized Control through Digital Equipment
EQUIPMENT CHARACTERISTICS

Before discussing the particular characteristics of component systems, let us consider three of the major decisions that influenced the design of the over-all system.
First of all, it was decided to maintain minor-loop analog controllers for two very practical reasons. 1) They do
provide continuous control of the different variables. In
order for the computer to duplicate such control, it being
time shared among all the variables, it would have to be
designed to operate at a tremendously high speed-one
well beyond that considered today to be reliably attainable.
2) In case of computer failure, the process would be held
to the most recent set points, changes of which could still
be made manually.
The second major decision concerned the data processing equipment. Here, the decision to be made was whether
to use a general-purpose or a special-purpose computer.
This decision depended upon the consideration of numerous factors. The main reasons for deciding to use a general-purpose computer were its flexibility (required to
control a complex process under varying process conditions) and its capability of adapting-or better yet, selfadapting-its program to varying conditions in the process. When discussing the computer, we shall see how this
capability to adapt itself enables it to provide meaningful
control signals.
Last, and most important, the reliability of the equipment, which must be capable of operating continuously,
must be carefully considered before installation in a process plant. This requirement for reliability dictates the use
of solid-state components throughout and a minimum use
of electromechanical devices. Keeping in mind the retention of the minor loops, the usage of a general-purpose
computer, and the employment of solid-state components
properly derated in circuits designed and constructed to
withstand an adverse plant environment (temperatures to
120°F, high humidity, and a generally corrosive atmosphere), we can now look at the component systems required to close the major loop.
SYSTEM INPUT SECTION

The input section (Fig. 3) supplies the control system
with the data that determine the operating conditions of
the process, the present settings of variables to be controlled and, of course, the program indicating the control
variables and the method of control.
These inputs are derived from two sources, the process
variable measurements, and the operator.
The process variable measurements include those which
yield information on the state of the process. These can be
temperatures, flows, and levels, as well as physical and
chemical characteristics of raw materials used, and the
characteristics of the end product.
Types of analyzers (instruments which are used to
measure the physical and chemical characteristics of both
raw materials and end products) will be chosen for the

particular process. While, control variables such as temperatures, flows, levels, etc., are derived from transducers
where the signal level can be as low as 10-50 mv full scale.
Since these signals are electrical analogs, they must be
quantized to at least one part in one thousand. The input
multiplexer must therefore be able to handle such lowlevel signals and to restrict noise levels in excess of 10 !J.V.
If the multiplexer can not handle such low-level signals,
an amplifier is required for each input. However since
we have found that there is a large number of inputs
(300-1000), the cost would be prohibitive. Therefore the
use of amplifiers must be avoided.

SELECTION
REGISTER
FLOATING
DIFFERENTIAL
INP<.;T - - - - - h t------+----l

Fig.

3~System

input section.

Furthermore, in order to circumvent ground-loop problems, the input. multiplexer must be able to switch both
sides of the signal line. The most reliable component that
can switch both sides of a low-level signal line and still
not introduce unpermissible noise, is the mercury-wetted
relay. This component has displayed long life (billions of
operations), lack of contact bounce, and a very stable contact resistance. The Clare HG2A series has been found
satisfactory both from the point of view of reliability and
of drive requirements. However, like any other component
handling analog signals, there are noise problems that have
to be contended with. Noise introduced by these relays is
composed of 1) a transient portion which lasts for approximately 3-5 milliseconds and 2) a dc portion. The transient
portion of the noise is caused by flux build-up and decay,
while the dc portion is caused by thermocouple effects resulting from thermal gradients existing between the two
sides of the relay.
The transient portion of the noise, although minimized
by using appropriate circuits, still exceeds the permissible
5-10 !J.V level. However, as will be shown below, this transient is rejected through the use of an appropriate analogto-digital converter.
The dc portion of the noise is minimized by appropriate
packaging and by providing a "shorting-bar" wiring ar-

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

rangement inside the relay. In this manner we have found
that the noise level can be reduced to below 10 !LV (actually
of the order of 1 to 5 !Lv). This enables us to maintain the
required signal accuracy of 0.1 per cent or better.
These relays are driven directly from transistor circuits
utilizing the RCA 2N217 germanium· transistor in a random-access switch arrangement. This feature, permitting
us to connect any input variable at random to the system
as directed by the computer, has been incorporated for
three very important reasons: 1) It permits switching to a
standard input for recalibration purposes at any desired
time, 2) it enables the equipment to sample discrete variables at time intervals shorter than the scanning cycle
(this is necessary in order to scan quantities that must be
integrated with respect to time, such as flows), and 3)
under a self-adaptive computer program, the equipment
can sample more frequently those variables that become
critical under varying process conditions.
The analog-to-digital converter must also possess special
characteristics determined by the signals generated in a
process plant. Besides the fact that the desired signals are
very low in level, they usually contain electrical noise produced by plant power equipment, such as motors. In addition, noise pick-up is frequently found when long-signal
leads (100-500 feet) are required. Because of the noise
and ground loop problems mentioned above, the converter
itself has been designed with the following features:
1) It provides input isolation required because of differences in ground potentials throughout the plant
with respect to the control system ground. This
ground potential difference can be as high as 5001000 volts during transient conditions such as
storms. Furthermore, this differential-isolating type
of input rejects common noise.
2) It can resolve 10 mv (or higher signals), with accuracies to 0.01 per cent.
3) It integrates over the sampling period of l/5 to 1
second. This sampling rate of 1 to 5 conversions per
second might at first appear to be slow, but the speed
has been selected to conform to the type of signal
present in a process plant where there is noise pickup. If rapid conversion were required, a tremendous
burden would be placed upon the computer if it were
to average many of these readings. Using an integrating converter, however, the computer is freed to
perform correction and control computations while
the converter is digitizing a reading.
The converter is completely transistorized and known as
the DADIT (Daystrom Analog-to-Digital Integrating
Translator). It integrates over the entire period of time
that the input is connected to it (1/5 second) and is not
affected by transient noise present in the relay multiplexer
(5-msec duration). Actually, the longer the integrating
period, the better the noise rejection. Although the multiplexer and the analog-to-digital converter are matched in
speed to cope with the "noisy" type of signal that they are
expected to switch and quantize, they might not function

fast enough if readings are to be made from a number of
variables at intervals closer than 1/5 second apart. By the
same token, since a "faster" converter would not provide
adequate filtering, and hence meaningful signals, two or
more converters are used when it is necessary to solve the
above-mentioned problem.
The multiplexer and analog-to-digital converter, supple:mented by any special equipment used to tie in product
and material analyzers, form the input to the system. It is
these units which monitor the condition of the process as
well as its performance.
In addition, the system is given instructions by the operator, such as pertinent results of laboratory analyses concerning a new raw material, market information, or other
commands the operator wishes to introduce. The operator
may even want to enter new programs. For this purpose a
paper-tape reader and a typewriter have been provided.
The reason for including a typewriter rather than a keyboard is so that a record containing instructions given to
the computer will be always available. The typewriter is
used for the manual introduction of data as well as for
the control of the computer.
COMPUTER

Once the data have been converted to the digital form,
they are no longer affected by component noise and drift,
and from then on we deal with digital accuracies.
The computer, in receiving these data, must first operate
upon them, linearize and scale factor the numbers in order
to account for different transducer characteristics and, in
general, convert a set of numbers to meaningful quantities
representing physical measurements as required by the
computer program or for logging sheets. The values of
some measurements have to be compared with alarm set
points while others have to be integrated and correlated in
order to provide information concerning the amount of
output, the process efficiency, and the determination of
product quality. This part is easy to program and is designed to yield information concerning the state of the
process at the present time.
The computations that determine the control signals are
complex (because we are dealing with complex processes)
and nonlinear time-varying in their characteristics. They
are therefore difficult to define in popular mathematical
terms, and the control program is largely biased by knowledge of the process, the interactions of different control
variables, and their effects upon the output (product).
The process is usually described with a set of nonlinear
differential equations. The object is to make these equations linear for a particular process state, although they
are over-all nonlinear. This can be achieved by creating
relationships which define the coefficients of these equations in terms of the measured variables (process state),
the control set points, and the process output. The system
then can, for a set of inputs and outputs, select the appropriate set of coefficients, and by so doing, create a
piece-wise linear approximation of a nonlinear process.

Otis: Optimized Control through Digital Equipment
The computer can then solve these equations with respect
to an optimizing criterion and create the control changes
that will yield optimum performance. Actually the computer is expected to iterate through many solutions, choose
the most attractive one, and introduce it into the process.
It will then observe the actual results obtained as compared
to the expected results, and will compensate for discrepancies by readjusting sets of coefficients. In this fashion
the computer continuously adapts its mathematical model
of the process to the actual case, accounting for different
plant conditions that occur, many of which could not have
been anticipated at the time of system design.
It is this capability of self-adaptation that dictates the
use of a general-purpose computer.
The computer must be fast with respect to the process
which, in computer language, is fairly slow.
The computer designed for this purpose is a 50-kc serialbinary machine. Word length is 20 bits plus sign and
parity. This provides a computational accuracy of one part
in 106 , or well in excess of the accuracy of the measurements. It has a coincident-current memory that can be as
large as 16,384 (214) words. It is a single-address machine
with a speed of 1.3 msec for an addition, and 10.1 msec for
a multiplication, including look-up. It has the full complement of arithmetic operation commands and branch and
shift commands, as well as special ones used for controlling
the input and output functions.
The computer, through its "analog-input" command, can
select a particular input, connect it to the analog-to-digital
converter, and then proceed with another computation while
the signal is being converted. One-fifth second later (at the
5 integration/second rate) it causes anew input to be
connected to the converter, accept the integrated reading
of the previous input, and proceed with the new computations.
Again the computer, like all the circuits in the input
section, is completely solid-state, uses approximately 3500
transistors, (RCA 2N217), and 3000 diodes.
SYSTEM OUTPUT

The output of the system serves the dual purpose of
introducing the required control changes to the process and
of informing the operator of the state of the process, the
changes made, and any other information derived from the
input variables. (See Fig. 4.)
It is very difficult to generalize concerning the design
of actuators since they depend upon the characteristics of
the actual equipment. Sometimes an on-off control will
suffice, while in other instances variables are controlled in
discrete steps. A digital stepping motor, such as the Sigma
Cyclonome stepping motor (magnetically indexed rotor)
and the Digitork, announced by The Teller Company, frequently can be used for this purpose.
The accuracy with which set points are set depends
upon the process and the particular variable. It is definite,
however, that today equipment can control much narrower
ranges than an operator can.

The output for the operator is derived through tape
punches and typewriters. The computer can select one to
eight punches if the system so requires. The reason for
using punches rather than typewriters directly is based on
two considerations: 1) The particular punch used is faster
than a typewriter (Western Electric punch type BRPE
#1,60 characters per second), and 2) they are considered
to be more reliable than the relatively complicated typewriter. The typewriter can be "down" yet the system can
continue to operate since the data are accumulated on
punched-paper tape.

OUTPUT PUNCHES IPAPER TAPE I

r----~,

L--~_......JI
I

,
,

LINK

TYPEWRITERS

I
I
I

__ ~-L
I
,

I
I

.I~

~OPERATOR

!I~
I

r---'-----. I,

----I- _

lSTATISTICAL INFORMATION',
PROCESS RECORDS
IOFF-NORMAL RECORDS
I

Fig. 4---System output section.

At this point, a word concerning reliability is again in
order. The responsibility placed upon the control equipment described above is tremendous. In order to meet the
requirements, this equipment must operate without failure
for many months, if not for years. Extreme care therefore must be exercised in choosing the components to be
used and in· designing its circuits. There is room neither
for components that have not been tested over long periods
of time, nor for critical circuits. Normal maintenance must
be performed while the equipment is operating, and check
problems should be run through automatically at predetermined intervals of time in order to ascertain proper
functioning. Finally, rapid troubleshooting procedures and
faulty component location are musts.
In conclusion, rising processing costs are forcing in'dustry to adopt automation. Optimized control will make
new processes economically feasible and, therefore, practical. The challenge is here, and the future will show how
well we are meeting it.
ACKNOWLEDGMENT

The author gratefully acknowledges the helpful assistance in preparing this paper given him by Eric Weiss,
David Taylor, Charles Taylor, Bill Waddell, and Wilbur
Erickson, all of Daystrom Systems, Division of Daystrom,
Inc.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE
Discussion

Question: In your opinion, is Daystrom's Heath analog computer adaptable
for use in the system described? If so,
wouldn't this simplify the system by eliminating the analog-to-digital converter?
Mr. Otis: While you are right in remarking that an analog computer would
eliminate the use of an analog-to-digital
converter, it would not meet the requirements in terms of reliability as well as
computational accuracy, time sharing, and
flexibility that we were able to design into
the present system.
Question: Is the system described for a
specific installation?
Mr. Otis: No, it isn't, although the first
two are committed and will go in specific
installa tions. The design is based on a
study of many types of processes.
Question: Would you please repeat the
arithmetic operation?
Mr. Otis: I assume by this you mean
the speed of the machine. It is 1.3 milli-

seconds for addition and 10.1 milliseconds
for a long multiplication and division. Shift
commands fall into the 1.3 milliseconds or
what we call the short commands.
Question: Is the coincident current
memory the only memory? Is there any
drum?
Mr. Otis: No, the coincident current
memory is the only memory in this system.
However, the machine is designed to work
with a tape deck.
Question: In the presence of interaction
in a process, how is optimum adjustment
of the variables achieved? What criterion
is established for behavior of multivariable
systems?
Mr. Otis: It is a difficult question to
answer in the sense that the criterion can
only be established if you have a particular
process that you are talking about. But, as
I indicated in the paper, you usually have
many sets of simultaneous equations and in
solving them together, you take into consideration the interaction of the different
variables.

It is this interaction between the
different control variables that this equipment introduces into the control system.
We want to take into consideration this
interaction. The criterion is usually an economic one. You want to produce a product
of a specific quality at minimum cost. This
might be achieved by the use of minimum
input power, the use of minimum raw materials, maximum catalyst life, or minimum
processing time, whichever the criterion
might be. You might have three or four
criteria and go down the line until you
find one that you will give you the best
answer.
Question : Would you please repeat the
type of your memory in your computer and
mention its capacity?
Mr. Otis: It is a coincident current
memory, transistor driven, and depending
on the system we can have from 1000 to
16,000 words in this memory.
In other words, the computer has the
circuits addressing up to 16,000 or 215
words.

Real-Time Presentation of Red.uced
Wind-Tunnel Data*
M. SEAMONSt, 1\;1. BAINt,

AND

W. HOOVERt

HE effective use of wind-tunnel testing in determining aerodynamic properties of a body is very
much dependent upon the reliability and speed with
which wind-tunnel data can be reduced. The ability to
provide reduced aerodynamic coefficients in real time, or
on-line, greatly increases the operating efficiency of the
wind tunnels and thereby reduces expensive wind-tunnel
time required for each test. This paper describes a system
for presenting reduced wind-tunnel data in real time for
the two wind tunnels at the Jet Propulsion Laboratory
(JPL'J.
The requirements for data-handling equipment and datareduction procedures for wind tunnels throughout the
country are quite diverse, and depend upon the windtunnel design and the type of tests for which they are used.
The supersonic wind tunnels involved in this description
are used for force tests, pressure tests, and miscellaneous
research studies, and include a variety of force-balance
systems. Consequently, the problems associated with on-

line data reduction for these tunnels can be considered as
representative of the problems associated with tunnels
generally.
Real-time reduced-data presentation requires a system
consisting of three major parts: 1) the instrumentation
necessary to convert force, moment, pressure, and angular
measurements into a form compatible with available computing equipment; 2) an operational system including the
computer program and methods for accomplishing the data
processing; and 3) a system for presenting reduced coefficients within a specified time interval in a form allowing use of test results to control the test program. An
earlier data-reduction system providing reduced data on a
daily basis has been described in the literature. 1 Most
elements of the new on-line system have now been developed. The completion and installation of the new system
will be accomplished step by step and will not involve
wind-tunnel downtime. Prime considerations in developing the new on-line system have been economy in capital
investment, high reliability, and flexibility in handling the
variety of test types.

* This paper presents one phase of research carried out at the
Jet Propulsion Lab., California Inst. Tech., under Contract No.
DA-04-495-0rd, sponsored by the Dept. of the Army, Ordnance
Corps.
.
t Jet Propulsion Lab., California lnst. Tech., Pasadena, Calif.

1 W. R. Hoover, ]. J. Wedel, and J. R. Bruman, "Wind-tunnel
data reduction using paper-tape storage media," 1. Assn. Computing
Mach.} vol. 3, pp. 101-109; April, 1956.

INTRODUCTION

Seamons, Bain, and Hoover: Real-Time Presentation of Reduced Wind-Tunnel Dalta

WIND-TuNNEL TESTING

SYSTEM REQUIREMENTS

Wind-tunnel tests are conducted Ly immersing an accurate scale model in the wind-tunnel air stream and recording a set of readings related to physical quantities such
as forces, moments, pressures, or angles. Most tests requiring on-line presentation of reduced data are force tests
or pressure tests. A point of force-test data consists of a
set of forces and moments with respect to some reference
system, the angles of attack and roll, the pressures necessary to specify tunnel operating conditions, and the freestream Mach number. A point of pressure-test data consists of a set of readings giving the pressure at points on
the body surface.
At JPL, force and moment measurements are usually
made using two types of balance systems. The six-component external balance system resolves three independent
components of force and three independent components of
moment about a reference system fixed with respect to
the tunnel. The strain-gauge system isolates forces and moments about a reference system fixed in the model. The
readings from the two balance systems are four-digit
numbers. The external balance readings require a code
digit to indicate the set of balance constants needed in the
data-reduction process.
Pressure tests are conducted by measuring the pressure
at fixed positions on the body surface. Measurements are
made by connecting each point with a pressure transducer
which transmits a signal to the automatic pressure readout system.
The system is capable of digitally recording as many as
192 pressure readings and of automatically ratioing each
reading to a prescribed pressure.
The purpose of force-data reduction is to convert test
readings from the balance coordinate system to a coordinate system suitable for engineering purposes. This
process involves a sequence of changes of scale, rotations,
and translations to obtain dimensionless aerodynamic coefficients. A general data-reduction scheme has been
evolved utilizing a sequence of vector by matrix products
which includes the transformations for both the external
and internal balance systems. In reducing the data, two
classes of constants exist, those fixed for a complete test
and those changing as the test progresses. The air-off zero
correction is a point taken with air off and at zero angle of
attack and indicates the balance condition for zero forces
and moments. The static tare is the change in balance
readings under air-off conditions caused by movement of
the model center of gravity during model pitch angle rotations. Moment transfer distances are dependent upon the
model configuration.
Pressure-test data reduction consists of computing ratios
for all model pressures to a set of known pressures, in
addition to computing profile averages, local Mach numbers, drag, and other aerodynamic forces. In general, pressure-test data reduction is much simpler than the forcetest reduction used at JPL.

One of the primary considerations in the presentation of
wind-tunnel data is the selection of the proper form of
raw-data record. Accumulated experience with the papertape storage system has verified the advantages previously
claimed/ and the real-time data-reduction system will retain this feature. The large volume of data output from
wind-tunnel operations requires fast and reliable data accumulation equipment which does not limit the operational
speed of the wind tunnel. The system must accept information from either of two wind tunnels and convert it to a
standard form for data reduction and presentation; also,
the system must be flexible enough to handle all categories
of force tests presented by a wide variety of engineering
requirements.
Real-time presentation of reduced wind-tunnel data requires that all of the final data be presented before normal
model or tunnel conditions are changed for a succeeding
run. It is desirable that all data-handling units be integrated and utilized in such a manner that real-time results
used for monitoring the test correspond to the final coefficient tabulations and plots required for engineering reports.
System requirements necessitate tabulating and plotting
directly from the raw-data record. This first inspection
serves to determine whether raw data appear reasonable
and sufficient (curves are smooth and defined); also, it
provides sufficient information to allow testing to proceed
on the basis of raw-data presentation in the event of computer breakdown.
The operating rate of the wind tunnel is 15 seconds per
point; this is an average time required to change an independent variable in the test procedure, allow the tunnel to
reach a stable condition, digitize the balance readings, and
punch a point of raw data into tape. A run of data includes a series of related points; for each run of data,
there is an additional two minutes available, thus increasing the average time available for computing purposes on
a run basis to 25 seconds per point.
The basic computer system to be used is an ElectroData
Model 202. To increase the efficiency of the system for
use on data-reduction problems, several modifications were
required. A second photo reader and input order were
added to the computer, giving the system two independent
input stations which are internally controlled and can be
actuated either manually or by programmed command
words. A second teletype punch with independent output
commands was added to the system. Presentation of windtunnel data in both tabulated and plotted form from paper
tape requires two output tapes with different data organization for use in the plotter and tabulator, respectively.
INSTRUMENTATION

Force Tests
The six components of force sensed by the external
hydraulic balance are measured on an automatic servocontrolled beam balance. The balance position is converted

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

to digital form by means of a shaft-position digitizer. The subsequently recorded on punched paper tape. The 192
force components obtained from the internal strain-gauge pressures can be scanned in approximately 40 seconds.
balance are measured by a strain-gauge bridge with servo For a lesser number of pressure tubes the scan time is corfollow-up which also drives a shaft-position digitizer.
respondingly shorter. The output of the pressure transThe digitizers employed are the double-brush decimal ducer can be scaled to any arbitrary value, thus obviating
encoders manufactured by the Coleman Engineering Com- much of the data reduction. As an example, if it is repany. These devices require a set of readout relays before quired that all model pressures be recorded as a ratio to
the digital data can be transmitted to the tape punch. In stagnation pressure, it is a simple matter to calibrate the
order to ease the interpretation of raw data, it is necessary system by feeding stagnation pressure into the machine
that the output readings be recorded in both plus and and adjusting the encoder to read 1000. The accuracy of
minus values. The conversion of digitizer readings to cor- the multipressure measuring system is better than 0.2 per
rect plus and minus decimal numbers is a relatively com- cent of full scale. This accuracy is achieved by the use of
plex operation requiring a sequence of relay closures. a single pressure transducer, allowing detailed observation
Since the readout relays are time-shared among the digit- of any transducer zero drift or calibration shifts. In adizers, the relay cycling time would be prohibitively long dition, the method of scanning provides for the gauge
and would slow the scanning rate if it were not for a novel to be connected to a vacuum system before each reading,
method of readout devised at JPL which requires only one thus eliminating a major source of gauge error, the hysteresis effect.
relay closure time per digitizer, approximately 15 msec.
The scanner records the digitizer and keyboard readings
ona punched paper tape using a Teletype BRPE-2 60digit-per-second punch. The slow-speed scanning of datasource words is accomplished by telephone-type stepper
switches while the high-speed scanning of individual digits
is accomplished electronically using magnetron beam
switching tubes and transistor switches.
The data are recorded on punched tape and then verified on a tabulator; in addition, selected words in the scan
are plotted on an automatic tape-controlled plotter (the
raw-data plotter indicated in Fig. 1). The raw-data plotter
can plot as many as twelve components on a single 30- by
30-inch sheet of paper. After the data have been displayed
on the tabulator and plotter, the tape can be read on a
Fig. I-Block diagram of wind-tunnel data-reduction system.
high-speed tape reader which is controlled by and transmits
information to the computer. After data reduction, the
After being punched the tape is fed to a high-speed tape
final data tapes are read by tape readers located near the
computer's output punches, and these readers are con- reader controlling the Burroughs Sensimatic printer. Setrolled by and transmit information to the final data tab- lected pressures can be plotted on the Electronic Associulating and plotting facilities. The tabulating machine is a ates Variplotter for the presentation of pressure contour
Burroughs Sensimatic used as a word-at-a-time printer. curves.
The data for a word are assembled in a magnetic-core
DATA REDUCTION OF FORCE TESTS
memory which together with associated circuitry controls
The computer program is written to conform to a standthe type bars of the Sensimatic. The final data are plotted
by three Electronic Associates, Inc., Model llOOD Vari- ard reduction procedure which converts information from
plotters. These plotters use 11- by 17-inch graph paper the coordinate system of the force balance to aerodynamic
coefficients in any or all of four coordinate systems which
and employ programmable symbol print~rs.
may be specified in a test. The same program is used to
Pressure Tests
reduce data from either internal or external balance sysThe data from pressure tests are recorded on a multi- terns; the procedure accounts for characteristics which are
pressure measuring system developed at JPL.2 This system functions of the tunnel, suspension, balance, and readout
scans as many as 192 pressure sources by means of pres- system.
The basic program for reduction of force tests consists
sure selector tubes. The unknown pressures are channeled
of
eleven steps, each written as a separate routine or subto a single pressure transducer;. the transducer reading is
routine.
Essentially, each of the steps is written as a
converted to a binary-encoded four-digit decimal number in
matrix-by-vector
multiplication, and any change required
a high-speed analog-to-digital converter, and the data are
usually consists of a minor alteration of a matrix at a
particular step. In some instances it is necessary to bypass
2 M. B. Bain, "A multipressure measuring system," IRE TRANS.
ON INSTRUMENTATION, vol. 1-6, pp. 18-22; March, 1957.
or reprogram a step completely; either type of change
WIND TUNNEL
CONTROL ROOM

COMPUTER ROOM

Seamons, Ba:in, and Hoover: Real-Time Presentation of Reduced Wind-Tunnel Data
affects only an isolated portion of the entire program. This
program structure minimizes the amount of pretest programming and checkout prior to the test date.
Preparation of a test for reduction consists mainly of
setting up an efficient method for handling run parameters
such as configuration constants, roll angles, Mach numbers,
and deflection constants. As standard practice, all combinations of constants are prestored in computer memory and
programming changes are provided to select and check
proper constants as the run number changes. Infrequently,
the schedule of constant changes overtaxes the limited
memory capacity and a second photoreader is used to introduce required changes.
Fig. 2 is a flow diagram outlining the sequence of operations on a point of raw data as executed by the computer program. Assuming that the basic program, the required program alterations, and all fixed constants have
been prestored in the computer, the input is actuated for
read-in of the first data point from the accumulation system. The raw data are permuted to a fixed order, and a
check is made to determine whether the point is an airoff-zero point. (If the point is an air-off-zero it is flagged
and stored in memory for later reference.) Introduction of
an air-off-zero point indicates the necessity for changing
run parameter constants; selection of such constants is
made either from memory or from an external photoreader tape. A coded word structure is used to flag input
of points from a static-tare run; pitch and rolling-moment
data from these points are used to form a static-tare table
which is in general applicable to a series of related runs.
Upon read-in of an air-on point, the static tare for the rawdata angle of attack is determined and corrections for airoff zero and static tare are applied to the raw data. The
corrected raw data are then reduced through the bodyaxis coordinate system. A set of code words is used to
determine to which of the coordinate systems the reduction is to be carried and the form of the aerodynamic coefficients, which are punched on two separate output tapes;
one tape is for tabulation on a Burroughs word printer
and the other drives automatic plotters for final data plots.
The pointwise reduction program as written reduces sixcomponent external balance data to body-axis coefficients
in 15 seconds. Results are presented with no increase in
inaccuracy and with all anomalies accounted for. Extension of the reduction to additional coordinate systems requires approximately 2 seconds for each system desired.
These times include the operations necessary to scale and
otherwise adjust all output to forms acceptable to the listing and plotting equipment.
In the proposed system, the total data-handling time for
a typical point of tunnel data will be approximately 23
seconds. Operating times of system blocks are as follows.

1) Digitizer, scanner, and punch require 1Yz seconds per
point; during this period the independent variable for the
next point may be set in the tunnel. 2) Raw-data tabulation and plotting will require 12 seconds; if the computer
is ready to accept a point of data when readout is completed, the raw data point will first be read into the computer and then into the raw-data tabulating and plotting
system in such a manner that reduction and raw-data presentation operate concurrently. 3) Computer read-in, reduction, and punchout of final tabulating and plotting tapes
require 18 seconds per point. 4) Tabulating and plotting
of final results require 3 seconds per point.

Fig. 2-Flow diagram of reduction program operations.

In the actual data presentation process the system will
always lag tunnel output by one or two points; however,
model pitch-down time incurred during the course of each
run will allow the system to "catch up" by the time the
run is actually completed. The times listed are basic times;
varying requirements of test procedures may necessitate
compromise between the amount and rate at which data
are taken and final results presented.
CONCLUSIONS

The system described for the real-time presentation of
reduced wind-tunnel data at JPL provides the broad flexibility required by the varied test programs conducted. The
system utilizes components now in operation with the
present data-processing system. Much of the necessary
computer programming has been accomplished and the
over-all system design has been completed.
The system will provide a real-time tabulation and plot
of reduced data for most standard tests and will have the
ability to provide reduced data with a time delay of one run
for most of the possible variations on the standard reduction.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

The Mechanization of Letter Mail Sorting
1. ROTKINt
HY mechanize the storing of letter mail? After
all, billions of letters are mailed and delivered
every month for just a few cents each.
The answer lies in the rapid increase of the volume of
mail. Fig. 1 shows how rapidly this volume is increasing.
Note that the ordinate is logarithmic and the curve is almost a straight line. Fig. 2 shows that the mail volume is
increasing even more rapidly than the population. From 6
pieces of mail per person per year in 1847, it has climbed
to 350 pieces of mail per person as of 1955, and we can
expect this number to be doubled by 1980. About 150,000
people are involved already in the sorting of letter mail.
At the present rate of increase, it will soon be difficult to
find enough suitable people in the country to sort mail
manually.
Thus the Post Office Department is forced to mechanize, simply to be able to accommodate the exponentially
increasing volume of mail. However, it also can expect
other advantages, such as speed, accuracy, reliability, and
economy of operation, and all of this with no increase in
personnel. The personnel now employed in the sorting of
letter mail need not fear layoffs, because it is the publicly
expressed policy of the Department that no one will be
laid off as the result of the adoption of sorting equipment.
So much for policy matters. Now to get down to engineering.
At the National Bureau of Standards Post Office Project, we started with these basic assumptions:

&0
50
40

1) The manual system of sorting letter mail has been in
use and under study for many years. It is very unlikely that it can be improved appreciably. Therefore,
any major improvement must be the result of introducing mechanization.
2) The U. S. postal system is too large to mechanize
all at once. Besides, this would not be a prudent way
to proceed even if it were possible. Therefore, mechanization must be introduced progressively.
3) At least initially, the mechanization of a post office
should not affect the nature of its output or input.
If a post office sorts outgoing mail to 2500 destinations and incoming mail to 600 letter carrier routes
before mechanization, it should continue to do so
after. This insures that the mechanization of one
post office will not force reorganization of the whole
postal system.
4) We cannot hope to do more than help the Post Office
Department get started in this field of mechanizing
letter-mail sorting. That Department will introduce
many improvements as a result of operating expert National Bureau of Standards, Washington, D.C.

j/
10

7
6
5

j
I

I-

8:~

o. 7

o. 6

o. 5

I
'I

'it!:
2

o. I

- - - NOOATAAVAlLA8LEFDRPERIOD 1848-1885

YEAR

Fig. I-Mail volume growth per year.
400

300
200

",00
~
90
'" 80
:l..70
_ 60

:l.

~b
'"
--'
"'

10
89

'"
:\

.,'

~ 40

:l.

,.,

/
I

,,
,,

f== ~-'-

f::= ==.ZI-- 'r-J.

1--,. NO DATA AVAilABLE FOR PERIOD 1848-1865 I
I I I
YEAR

Fig. 2-Ratio of mail volume to population.

ience. Hence we must not delay the mechanization of
post offices while we strive for perfection in equipment or systems. Rather, we must freeze designs
and procedures as soon as practicable to speed the
day of actual operation.
5) The greatest improvement potential lies in those operations that take the ~ost man hours per letter. This
leads to the conclusion that manual letter~mail sorting is the operation most deserving of attention, because each letter is individually sorted three to eleven
or more times. The average must be about six, including the last sort in which the postman arranges
the letters in the order of his walk.
Having established this "axiomatic" base, we started
our study in three directions simultaneously:

Rotkin: The Mechanization of Letter Mail Sorting
1) An examination of the physical characteristics of
letter mail,
2) A study of the mail flow through present sorting
offices, and
3) A study of existing equipment for sorting mail.
Here only a brief summary of the third is given.
There are systems in use in other countries in which
the human operator reads the mail, interprets the address,
and indicates the proper destination in the output of the
machine to which each envelope should go. But a machine
helps him do the purely mechanical work by bringing letters to him and taking letters from him, each to its designated output bin. Such a 'machine makes it possible for a
human being to sort letters about twice as fast on the
average as he can do manually and to sort, not to a maximum of fewer than a hundred output destinations, but to
three or four hundred. Such machines have been in use in
Holland since about 1935, in Belgium since about 1950,
and in England since the Spring of 1956.
However, these machines would not eliminate multiple
sorting of mail for large post offices in the U. S., because
these post offices sort to more than three or four hundred
destinations. Moreover, the operators must be of even
higher caliber mentally than manual sorters, since they
sort to three times as many destinations at double speed.
This is a serious drawback, because it limits severely the
fraction of the population capable of doing this work.
Finally, these machines are not designed to take advantage
of any future standardization of the mail. It was therefore
concluded by the National Bureau of Standards personnel
that these machines would not constitute a final solution to
the mechanization of letter-mail sorting in the United
States, although they may find their place in the smaller
post offices of this country just as they have in foreign
countries.
A more promising system is the one that is being developed in Canada under the leadership of Dr. M. M.
Levy. Dr. Levy has proposed that each letter be standardized by having its address converted into a dot code by
the first human being who reads the address in the post
office. In this way, subsequent readings of that address
can be carried out by machines, and no further manual
reading of addresses is required until the postman is about
to deliver the letter to its final destination. Thus, the number of human readings for sorting is reduced from an
average of about six to one. In manual sorting today there
are about six sorts on the average, including that of the
postman ordering his route. Now although the conversion
of the address into the dot code may take longer than the
ordinary sorting of a letter manually, it does not take six
times as long; the difference represents the savings
in manpower achieved by the adoption of this system. The
exact saving is not known as of this time, although we
hope to have such information in a comparatively short
time.
There is a further potential advantage in this type of

standardization. A large fraction of the mail has its origin
in business houses or firms which specialize in direct mail
advertising. Such firms maybe induced to imprint the dot
code on the envelope themselves. In those cases, the Post
Office Department would not have to use any human
readers. We have also urged the Post Office Department
and firms in the business-machine field to adopt a standardized type font based on a 5 X 7 mosaic for machine
reading. The use of such a font in a standard size for
addressing business mail would also eliminate the need for
human readers.
Having established our axiomatic base and studied what
was available from others, we proceeded to develop our
own system using the Canadian idea. We did not adopt
the Canadian equipment, because it was not suitable for
our post offices. All history is omitted here, and we shall
describe our system as we visualize it today.
Fig. 3 is an artist's conception of one section of a
mechanized post office. In the foreground are the codeprinting stations. Mail that has been cuLed, faced, and
cancelled is brought to these stations. Human operators
read the addresses and operate keyboards to rewrite the
addresses in a standardized, abbreviated form. These
standardized addresses are printed on the back of the envelopes in a dot code not very different from Teletype
code. The printing is done with phosphorescent ink to enhance contrast during subsequent mechanical reading.

Fig. 3-Artist's conception of one section of a
mechanized post office.

Provision is made at this stage for a rough sort of the
mail by classes. It may prove operationally advantageous
not to code immediately all of the mail that reaches these
stations, and therefore provision is made to put some of
the mail aside into several categories. For example, since
in most large post offices between a third and half of the
mail is local, it may be operationally advantageous, in
order to speed up the outgoing mail, to put the local mail
aside for later detailed address coding. Similarly, it may be
advantageous to separate out mail requiring special handling, such as air mail, special delivery, and registered mail.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Also, provision must be made to eliminate from further
operations defectivemail.e.g .• mail which has insufficient
postage or an incomplete or improper address.
One of the classes will consist of the mail that has been
coded for immediate sorting. This .mail is transported
bulk-wise to the boxes shown mounted on the large machine in the rear. Each of these boxes represents an automatic code-reader input for the large machine which is the
distributor. In each code reader, the dot code on the envelope is illuminated by ultraviolet light and then read in
darkness by photocells. The resulting electrical address
signals are then fed to one of the three translators represented by vertical rectangles on the right side of the illustration. One of these operates for outgoing mail, another
for incoming mail, and the third operates when the mail
for each carrier is arranged in the normal order of his
walk. The appropriate translator accepts the address signals, and translates them (in terms of its stored scheme
of distribution) into electrical signals representing the
proper output bin on the distributor. These output signals
are sent to the distributor in synchronism with the envelope itself. The distributor then delivers the envelope to
the designated output bin.
This, in very broad outline, is the system of mail sorting
that the National Bureau of Standards now has under
development. What are the problems associated with this
system?
The chief problem,· of course, is to develop criteria,
methods, and data for judging the relative merits of various sorting schemes. To this end, we have investigated
the sorting operations of several post offices such as Washington, Baltimore, Philadelphia, Los Angeles, and San
Francisco. We may add Chicago and St. Louis to this list
later. We study the relative volumes and physical characteristics of the mail from all sources in each of these cities.
We study the mail flow pattern through each post office,
and we determine the destinations to which it must sort.
There is much too much mail in each post office for us to
conduct a complete piece count, so we have developed statistical sampling procedures. These data enable us to determine the system parameters for any mechanized scheme of
doing the work of each of these post offices. Our comparisons are thus real instead of conjectural.
In addition to the over-all systems problem, there are the
many equipment and procedures problems. Some of these
will be examined in the order in which they would arise in
following an envelope through a mechanized sorting operation. Thus we start with the code printer.
First, what kind of a code shall we use? It must require
as little operator time as possible, be easy for operators to
learn and to remember, and result in a high speed of operation with a very small error rate. An early report on code
development was made to the Post Office Department.
Second, what type of keyboard shall the operator use?
We expect to use ordinary typewriter keyboards, initially
at least. nut we are not sure that these are best; and tests
are being run to see whether the direct digital keyboard

used by the Canadians or a rather elaborate multichoice
keyboard proposed by the Dutch may not be superior.
These are human engineering problems and in order to
answer them, we have been forced to run tests with human
beings. The same test series is being used to determine
the human engineering aspects of the codes. These tests
are expensive. They take a long time to run. They must
be very carefully designed, and the results must be very
carefully interpreted so as not to be misleading.
.
Third, a more straightforward engineering problem, in
what form shall the abbreviated address appear on the envelope? Our choice has been to use phosphorescent dots in
a code similar to Teletype tape. We arrived at this choice
f or three reasons:
1) Only optical sensing equipment can be relatively distant from the surface on which the pattern is imprinted and still retain sharp focus;
2) Only in an optical pickup can sensitivity be independent of the distance between the pickup device
and the surface from which it is reading; and
3) The use of phosphorescent ink allows us almost
complete immunity from anything that may be
printed on the envelope in the way of extraneous
matter such as advertising, and it gives us a very
good signal-to-noise ratio. Phosphorescent dyes are
not used today in envelopes nor in the inks used on
envelopes. Furthermore, we have proposed that such
use be forbidden in the future. There is really no
reason for anyone to want to use them, so that this
prohibition should not work any hardship on anybody.
N ext we come to the translator. It is technically possible
to develop an electronic device which is so fast that it can
translate address signals into output signals for a whole
post office by time-sharing its services to several distributors. The alternative is to develop a device cheap enough to
be used with a single distributor, but required to make
translations at a correspondingly slower rate. Which shall
we use?
One contractor has developed laboratory versions of
both kinds of equipment. We believe the future belongs
to the more potent equipment, because it will be more compact and economical. However, for the present, we are
perfecting the more modest version; so that, in the event
of breakdown, only one translator is made idle instead of a
whole post office.
The translator under intensive development is somewhat
similar, but by no means identical, to the punched metalcard device used by the Bell Telephone System in its longdistance automatic switching. It will be small, cheap, and
fast enough. The author hopes that J. Rabinow, the inventor, will prepare a detailed history of the development
of this device. It would make a fascinating story.
Finally we come to the distributor. First, there is the
question of its general design. There are two ways of dedesigning such equipment. One way is used by the Bel-

Rotkin: The Mechanization of Letter Mail Sorting
gians and the British in their designs. We shall call this
an external control system. It makes use of a main distributor frame, which is essentially a mechanical device
for pushing envelopes, but which carries no intelligence
whatever. The instructions to this device, as to when and
how it shall divert envelopes into the proper output bins,
come from separate control equipment which must run in
strict synchronism with the main distributor. Such devices
have been built which work satisfactorily. However, in
order to insure strict synchronism between the distributor
proper and its control, the construction must be very precise and stable, and the flexibility of layout is limited. We
have chosen, therefore, to adopt a modification of another
system employed by the Dutch and others. This we call
the self-control system. In this, each envelope has associated with it, a mechanized form of the output bin address
as it moves through the machine. Each output bin also has
a mechanized address associated with it. The two work
together like key and lock. There is no auxiliary control
unit. Thus, it makes no difference how far this envelope is
carried, nor in what direction, because its instructions
travel with it.
Another question that arises in connection with the distributor is: For how many output destinations should it be
designed? From the point of view of the postmaster who
has a job to do, it is desirable to design this equipment to
handle the same destinations that he now sorts to manually. This means that the introduction of the equipment
would not force any changes in the mail transportation
system. So far as other post offices are concerned, they
would not know any difference after the equipment was
installed. However, this implies, for a place like Chicago,

Discussion
Question: Do you care to say anything
about mechanical facing of mail incidental
to mechanical mail sorting?
Mr. Rotkin: Since this is not a mail
handling specialist group, I had better explain that facing means arranging the
envelopes so they all face the same way.
It is done at present for the convenience
of the canceling machine as well as that
of the sorter. This is not part of the work
that the National Bureau of Standards is
doing. We pick up the mail after it has
been faced and canceled. It is perfectly
feasible from a technical point to do this
mechanically and the Post Office Department is working on it in conjunction with
at least two contractors.
Question: What type of distributor is
visualized for the final route sorting of
the carrier mail?
Mr. Rotkin: The same kind that would
be used for distributing outgoing and incoming mail. It is our objective to have one
distributor do the work of all three kinds
of sorting, either by time sharing or by
just dividing the distributor into sections
appropriate to each sort.

for example, sorting to about 5000 destinations. It may
not be feasible, for cost or space reasons, to build so large
a distributor.
These problems have been described very briefly and
in rather general terms. We are attacking them much more
specifically, using, mathematical models, computer simulation, statistical studies, and engineering trials. It is hoped
that each of these problems will be the subject of a separate paper at a later date by some member of the
project staff. These papers will be prepared for internal
use of the project and the Post Office Department. If they
prove to be of interest to others, we are sure the Post
Office Department will permit distribution.
The National Bureau of Standards Post Office Project
staff is not doing all these things by itself . We are getting invaluable help from the Post Office Department in gathering
and interpreting postal data. Most of the very ingenious
engineering features of the sorting equipment are due to
the Rabillow Engineering Company, our sorting equipment contractor. Prof. Harry H. Goode of the University of Michigan, who is chairman of the Technical
Program Committee of this Conference, has been our invited critic, and a very conscientious and useful one. It
is he who has encouraged us to make earlier and greater
use of mathematical models. Many others have also helped
in other ways.
This has been a very quick review of the work of the
Post Office Project at the National Bureau of Standards.
It has run very lightly over a large and complex field. vVe
hope that, in time, ways will be found to cover the ground
more thoroughly for the benefit of those who may be interested in greater detail.

I t is more likely to be by time sharing
because of the way the work comes into
the post office. It is quite feasible to do
this. The· number of sorts involved is
comparable. For example, in Washington,
we sort to about 2500 or 2600 outgoing
destinations. We sort to about sao or 600
carrier routes and each carrier makes several hundred stops. I don't know the exact
number but it is in the neighborhood of
500.
In Chicago, for example, they sort to
about 5000 or 6000 outgoing sorts; for incoming mail, they sort to about 5000 or
6000 carrier routes and each carrier also
has somewhere in the neighborhood of 500
stops.
Therefore, the same equipment can be
used for all of the sorts and the only difference required would be in the kind of
instructions the machine gets from what I
call a translator.
Question: Will it be necessary to standardize letter sizes to attain optimized
mechanization?
Mr. Rotkin: It is not necessary; it
certainly would be helpful. We are trying
at present to determine what the largest
letter will be that we will accept into this

mechanization sorting system. This is one
of the reasons for measuring physical
characteristics of the mail. Letters come
in all sizes from very, very small to quite
gigantic things called fiats. It obviously
does not pay to design a machine that will
handle anything as large as fiats because
it doesn't occur in a high enough percenetage
of the cases.
We don't know whether this will be
the proper cutoff point, but tentatively,
we are working on the assumption that
6 X 12 x:VB is big enough. I expect that
after we have done our statistical work,
we will find out this is too big, and then
we can afford to make our cutoff for a
smaller size of envelope.
Question: Is character reading a part
of the system?
Mr. Rotkin: It is not a necessary part.
If we could have good character reading,
it would be helpful. This is why we would
like to have people standardize· the style
and size and format for addresses for business purposes. It would make the problem
of character reading almost child's play
compared to what it is today with the wide
variety of sizes, type fonts, and· general
arrangements of addresses.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Preparations for Tracking Artificial Earth-Satellites
at the Vanguard Computing Center
D. A. QUARLES, JR.t

INTRODUCTION

T is considered appropriate to remark at the outset
that part of the computer programming system subsequently to be described has been used extensively in
performance of tracking calculations for artificial earthsatellites. This operational use commenced with the tracking at the Vanguard Computing Center, Washington, D.C.,
of the first Russian satellite last October 5, the day after
the satellite was launched, and has continued during the
interim.
To perform with great speed the calculations which
would be required in the tracking of any artificial earthsatellites launched during the period of the International
Geophysical Year ending December 31, 1958, an IBM
Type 704 Electronic Data-Processing Machine was installed
at the Vanguard Computing Center. The purpose of this
paper is to describe the flow of operations and calculations
performed using the programming system developed to
handle these tracking calculations on this stored-program
704 computer .
. Various stages involved in the data processing will be
dIscussed; stages commencing with receipt of raw observa~ional data and extending to provision of computed orbital
mformation specifying the predicted future motion of a
satellite,: and extending eventually to provision of a comprehensive history of its past motion. Deadlines pertain to
processing observational data and to distributing predicted
positional information.
It is emphasized that the only calculations considered in
this paper pertain to satellite tracking and orbit determination. It is the preparation for such calculations only-not
for calculations pertaining to satellite or launching-vehicle
structural desig.n, not to launching-vehicle trajectories,
and not to studIes of the earth and its atmosphere-that
the Vanguard Computing Center has been responsible.
Furthermore, the Vanguard Computing Center, which is
owned, staffed, and operated by the IBM Corporation
under a contract with the Navy, is largely concerned with
problems involved in processing positional information obtained from a satellite's continuous radio transmission.
Problems involved in acquisition and processing of optically or visually obtained observations of a satellite enabling continued use of new observational data for orbit determination if the satellite outlasts the life of the batteries
which power its radio transmitter, are primarily the concern o.f another group, also using a 704 computer, and located m Cambridge, Mass. Naturally, however, both the

IBM Corp., New York, N.Y.

Cambridge and Washington groups are interested in using
observations obtained by both radio transmission and optical or visual methods.
The responsibility for establishing the orbit computation
procedures rests with the working group on orbits which
includes: Dr. J. W. Siry and J. J. Fleming of the Naval
Research Laboratory; Dr. Paul Herget, Director of the
Cincinnati Observatory; Dr. G. M. Clemence and Dr.
R. L. Duncombe of the Naval Observatory. The detailed
mathematical formulations were developed chiefly by Dr.
Herget with assistance from Dr. Peter Musen, formerly
of the Cincinnati Observatory.
The author of this paper has had responsibility, together with those working under him in New York, N.Y.,
for planning and preparation (synonymously "programming") of the system of instructions used by the 704 computer for accomplishing the desired orbital calculations,
and for the programming of certain special calculations.
These special calculations were performed during the development of the mathematical formulation, and guided the
course of this development.
IBM staff members of the Vanguard· Computing Center, assisted by the New York system development group,
will handle the operation of the programming system on
the 704 computer. The operation of the system is highly
automatic, with many minor decisions being made by the
computer according to rules provided to it. Nevertheless,
from the processing of raw observations, to the computation of orbital characteristics and predicted positional information, the operation permits certain major alternative
techniques to be used at various intermediate stages of the
calculations. Decisions regarding some of these major alternatives can be supplied in advance to the computer,
enabling its subsequent automatic handling of the desired
choices of these alternatives. However, if these decisions
are not made in advance or if it is desired to alter any of
these decisions during the course of the calculations the
prog.ramming system is designed to make this conveni~ntly
pOSSIble by manual intervention. On hand to assist in any
such decisions and to interpret the calculated results will
be Dr. Herget and others responsible for the formulation.
GENERAL DESCRIPTION OF PROGRAMMING SYSTEM

A very general description of the structure of the programming system will now be given before discussion of
mathematical techniques employed and of their flow of
operation. From the outset of the planning, it became apparent that the system should be designed to permit convenient choice not only in the methods of computation

Quarles: Preparations for Tracking Earth-Satellites

used, but also in the order of their use. To enable great mediate results. The system is designed to enable the obflexibility in these respects, it was decided to make the sys- servational data, subsequent to its preparation in decimal
tem card-controlled. That is to say, the choice and arrange- punched-card form, to be supplied as input to the comment of control cards in an input deck would, barring puter, optionally from these punched cards directly, or
manual intervention, govern the choice and ordering of any from binary-coded-decimal tape prepared on a card-to-tape
operations subsequently performed by the computer. The peripheral device. To check for occurrence of random materminology "macro-operation" was adopted to denote a col- chine error, the internal calculations are generally perlection of subroutines linked together to perform a broad formed in duplicate with comparison of check sums, while
orbit-computation function.
transference of information between high-speed magnetic
The component parts of a macro-operation were fre- core storage and auxiliary magnetic tape or magnetic drum
quently needed in other macro-operations. Any such com- storage is checked by check sums, and special checks are
monly-needed component part was then usually written as made for input-output operations. In addition to the
a subroutine with sufficiently general specifications to en- printer, magnetic tapes, and card punch, the use of which
able its use in the various macro-operations in which this has already been indicated, the cathode-ray tube display
type of operation was required. This frequently-adopted and cathode-ray tube recorder are available as output depractice of programming by subroutines, rather than di- vices. Programming has also been done to enable optional
rectly by macro-operations, had several significant advan- use of these devices to provide plotted output in directtages. Not only would this practice eliminate much dupli- visual and/or filmed form.
cation of programming effort which would otherwise have
FLOW OF OPERATIONS AND COMPUTATIONAL METHODS
occurred, but also it would tend to keep localized, perhaps
to a single subroutine, effects of a minor change in formuThe flow of operations performed in processing the data
lation. It may be remarked that for a research endeavor will now be discussed, together with brief descriptions of
in a new field, such as the establishment of the formula- the computational methods involved. As a starting point,
tion for tracking artificial earth-satellites, some changes suppose that some of the continuous radio transmission
in the formulation should not be unexpected.
from a satellite has just been received by one of the speIt was decided to assign one auxiliary storage magnetic cially-designed receiving stations, called "Minitrack statape of the 704 computer to serve as a system tape, each tions," during a single transit of the satellite over the stablock of information on this tape comprising instructions, tion. This transmission is subdivided at the Minitrack staconstants, and control information required by just one tion into individual observations, up to about thirty in
macro-operation. Thus each control card (the contents of number and evenly spaced in time. Each observation then
which are fixed) possibly followed by one or more input- consists of phase-difference readings, one for the east-westdata cards (the contents of each of which are variable) oriented radio antenna of the station, and one for the northcorrespond to one macro-operation. In general, when a south-oriented antenna, both corresponding, after small
macro-operation control card is read at the card reader, a adjustments, to a single instant of time. A typical time intersystem subroutine retained in high-speed magnetic core - val for these observations is one second. The associated instorage causes (with a minimum of tape motion) the ap- formation, comprising the phase-difference readings, first
propriate macro-operation information to be read from the and last times for the readings, and certain information
tape into high-speed storage immediately prior to its use. identifying and specifying characteristics of the observing
In general, the complete set of instructions required for station, will for convenience be called a "message." Each
each macro-operation consists not only of those instruc- message is transmitted in triplicate to a control center at
tions transferred for the macro-operation from magnetic the Naval Research Laboratory, and thence, still in triplitape to high-speed storage, but also of utility subroutines cate form, via teletype to the Vanguard Computing Center.
(e.g., input-output subroutines) commonly needed by A device at the Vanguard Computing Center automatically
various macro-operations, and retained in the high-speed converts the teletype tape to decimal punched cards.
This message is then ready to be processed by one of
storage. The auxiliary magnetic drum storage is used to
store certain input parameters which are subject to change, the macro-operations of the programming system for the
and also output of certain macro-operations which may 704 computer. This macro-operation, which" is always the
serve as input for one or more subsequent macro-opera- first macro-operation to process any of the messages, intions. Comparatively small amounts of output are directly cludes four principal subroutines. The first subroutine is
printed and/or punched on cards. Larger amounts of out- designed to load the message into the high-speed storage
put are written in binary-coded-decimal form on magnetic of the computer, transforming information from compact
tape for subsequent printing on a magnetic tape-to-printer form on the input-data cards (or, optionally, input binaryperipheral device not connected to the 704 computer, and coded-decimal tape) into a more suitable form for its suboperated independently of it. In particular, one such tape, sequent use. The second subroutine performs an editing
called a "log tape," was assigned to preserve a detailed function, comparing the triplicate items of the message for
chronological history of calculations performed, including exact agreement. This second subroutine reduces the size
not only input and output quantities but also many inter- of the message if permissible when, for a given item, at

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

least two out of three of the corresponding items do not
agree exactly. The third subroutine performs several adjustments to the data due to certain characteristics peculiar to the Minitrack station which received the data,
due to passage of time during the recording of a single
observation, due to radio refraction, and in order to convert the phase readings to directional information. Then,
each adjusted and converted observation provides the approximate direction of the satellite from the observing
station at a certain instant of time. It is to be noted that
no direct measurement of the distance of the satellite from
the observing station is yet available. In fact, satellite distances will be derived as output of orbital calculations.
Only after such calculations, shortly to be discussed more
specifically, does such a distance serve as an input quantity.
The processing, thus far, also checks to assure that at
least three individual observations remain so that the
fourth subroutine may have reasonable assurance of being
able to perform its function. This function is to fit a leastsquares parabola to each of the two sets of direction components which were derived collectively from east-west
and north-south phase-difference readings. The principal
output of this macro-operation, then obtainable, consists
of a single "smoothed" direction of the satellite from the
observing station, expressed in a local coordinate system
and corresponding to an instant of time. This instant of
time is centrally located with respect to the time range of
the set of raw observations which, it is recalled, were
obtained from a single Minitrack station during one transit
of the satellite over this station. This principal output is
provided in both punched-card and printed form. Certain
subordinate output quantities, such as any discrepancies
between the triplicate messages and standard errors of the
least-squares parabolic smoothing operations, also are
printed. Also, as in the operation of all macro-operations, a
detailed record of input, intermediate output, and output
is preserved on the log tape. In the subsequent discussion
of the flow of calculations, it will be understood that preservation of information on a log tape, and printing of output information which is required for more rapid surveillance, are included in all macro-operations, whether or
not explicitly mentioned (see Fig. 1).
Several methods have been programmed in order to
enable computation of a preliminary orbit from such observations. (Here, and subsequently, the qualification
"parabolically smoothed" is understood when referring to
an observation.) One of these methods makes use of two
observations which, roughly speaking, are suitably widely
spaced in time, to obtain a preliminary circular orbit.
Other available methods develop a preliminary elliptic
orbit from three or four observations which are neither
too closely nor too widely spaced in time. If the observations were too closely spaced in time, inaccuracies in the
observations could seriously reduce the accuracy of the
result, while the particular methods used would be invalidated in the case of observations spaced too widely.

Fig. 1-Processing of Minitrack (or Minitrack-simulated)
input. Macro-operation 1.

Fig. 2-Preliminary orbit determination. Macrooperations II, III, IV.

These methods are based on iterative techniques due to
Gauss, and are applied after transformation of coordinates
of observations, expressed in their original local systems,
to an inertial coordinate system. In each case, a reference
position vector, and a corresponding velocity vector, are
obtained for the satellite at some instant of time. Then,
from this reference time and corresponding position and
velocity vectors for the satellite, the programming provides
for computation of certain quantities serving to characterize the orbit in question in different ways. For example,
such computed quantities are the period of the satellite's
revolution, the inclination of its orbit plane, and, in the
case of the elliptic orbits, also such quantities as the semimajor and semiminor axes of the orbit, and the perigee
(,i.e.) closest to earth) position of the satellite. The programming for each of these preliminary orbit computation
methods constitutes a separate macro-operation ( see
Fig. 2).
After having obtained a preliminary approximate orbit,
other methods have been programmed to enable its improvement and, as the orbital characteristics would constantly be subject to change, also its updating. One macrooperation consists primarily of a procedure for numerical
integration of the differential equations relating, by Newton's law, the components of forces acting upon the satellite, to the components of its acceleration. Each of the
three force components contains an expression for the
force on the satellite due to gravitational attraction of the

Quarles: Preparations for Tracking Earth-Satellites
nonspherical earth, and an additive term for force due
to an admittedly fairly rough estimate of atmosphericdrag force acting upon the satellite. As is well known, one
of the reasons for launching artificial earth-satellites is
the desire to gain further information about the structure
of the upper atmosphere. Until more is learned about the
upper atmosphere, in particular about variation of atmospheric density with altitude, the components of drag force
in these differential equations may be subject to significant
inaccuracies. The other contribution to the force components, due to gravitational attraction of the earth, may
also introduce significant error because of our inadequate
knowledge of local variations .in the earth's gravitational
field. There are, of course, further errors in the numerical
integration procedure caused by inaccuracy in initial values
of the satellite's position and velocity, due to replacement
of derivatives by finite differences, and due to growth of
error during the numerical integration computations.
Though its accuracy is limited by such errors, this numerical integration macro-operation is expected to be very
useful, not only because its output consists of predicted
positions of a satellite spaced in time by an arbitrarilychosen time interval, but useful also, in combination with
a method of differential correction, for orbital improvement and updating ( see Fig. 3).
The macro-operation for this differential correction
method obtains corrections to position and velocity vectors
corresponding to a reference time which may be periodically updated. These corrected vectors may be used, as
before, to obtain new orbital characteristics. The input
for this differential correction procedure consists of observations (preferably including, because of changing
orbital characteristics, the latest available) and predictions, at the same observational times, obtained from output of numerical integration by 6-point Lagrangian interpolation. So-called equations of condition are computed,
one set for each observation, after making an improved
adjustment to the observations for refraction. The differential corrections are then readily obtained from leastsquares solution of these equations of condition. The processes of prediction by numerical integration and differential correction may be perfonlJ.ed iteratively in attempt to
bring predictions and observations in close agreement
(see Fig. 4).
A more complicated alternative technique, which is also
planned to be used for prediction and orbital adjustment,
is based upon three further macro-operations. In brief,
one of these macro-operations uses a modification of Hansen's lunar theory to compute Fourier series representations of orbital characteristics, including perturbations
due to the oblateness of the earth, in terms of a variable
representing time. A second macro-operation treats separatelydrag perturbations in these orbital characteristics
by numerical integration. By evaluations of these Fourier
series and additive perturbations, adjusted orbital characteristics and, thence, derived predicted positions, may be

Fig. 3-Predicted positions (inertial vectors).
Macro-operations V, VI, VII.

Fig. 4----'Orbital adjustment (improvement and/or updating)
Macro-operation VIII.

obtained. The third macro-operation computes differential
corrections to the orbital characteristics by a technique
similar to that used for correcting the reference position
and velocity vectors. This technique for prediction and
orbital adjustment, using these macro-operations iteratively, is expected to be a valuable alternative until the
later stages of a satellite's "life." Near the end of the
flight of a satellite, the drag force would become large
enough to invalidate the method used for the oblateness
perturbations.
At such time, it is planned to continue predicting and
orbital adjusting by the first-described techniques for numerical integration and differential correction (see Figs.
3 and 5).
Two separate macro-operations are available for transforming predictions from the inertial-vector form into
forms more convenient for use by the general public. One
of these macro-operations provides as output, for each
prediction of a specified time span, the time, latitude, and
longitude for the subsatellite position, height, and zenithangle-acquisition information. Also provided as output is
a list of any official Minitrack or optical stations to be
alerted due to the satellite's proximity, but, in the case of an
optical station, only if the favorable observation condition
of twilight exists. The other of these macro-operations provides, as output, positional information for a specified set
of times and relative to a specified station (see Fig. 6).

PROCEEDINGS OF THE EASTERN COJl.IPUTER CONFERENCE

Fig. 5-0rbital adjustment (improvement and/or updating).
Macro-operation IX.

Fig.6-Predictions (transformed for convenient use).
Macro-operations X, XI.

than the 704 computer would not have been able to do the
same job adequately. In particular, in the very early stages
of a satellite'sflight, the Vanguard Computing Center will
be concerned with developing and distributing orbital information within a matter of minutes of the receipt of
observations which first enable preliminary orbit determination. However, until at least one revolution of a satellite has occurred, observations may be so sparse that better
than a rather inaccurate determination of the satellite's
actual motion would be prevented.
The magnitude of the programming system's development, and of associated programming for special calculations, may be measured by an estimated expenditure of
between 6 and 7 man-years of work involving the writing
of approximately 25,000 instructions. The size of the programming system is a result of a complex mathematical
formulation whose programming included many different
types of Fourier series manipulations, and a result of the
system's flexibility in enabling convenient use of alternative computing techniques. To guard against possibility of
machine breakdown at an inopportune time during a satellite's flight, another IBM center with a 704 computer will
be kept prepared with operational information by means of
a transceiver and telephone on an emergency stand-by basis.
ACKNOWLEDGMENT

CONCLUSION

Several remarks seem appropriate at this point. It was
earlier stated that certain deadlines pertain to processing
observational data and to distributing predicted positional
information. Satellites, of the types presently being considered, would complete a single revolution around the
earth in approximately an hour and a half. Allowing for
communication and distribution delays in incoming data
and outgoing predictions, the speed of a satellite's motion
requires very rapid calculations to be performed in order
to enable use of the methods described in providing, sufficiently in advance, alerts and predicted positional information to observers around the world. A machine with
significantly slower speed or significantly smaller storage

Discussion
It should be noted that these answers
were taken from the latest technical information available through March 26,
1958.
Chairman M. Rubinoff (Philco Corp.,
Philadelphia, Pa.): Is a triangulation system used or planned to be used to increase
accuracy of observation? This question is
by E. A. Keller.
Mr. Quarles: In attempting to answer
this question, I shall restrict my attention
to the standard Minitrack receiving units
being used by the Navy in performance of
radio-tracking responsibilities. The disposition of these stations was aimed at providing a very high probability of receipt

The author wishes to express his appreciation to those
whose efforts and assistance have made possible. the development of the programming system. These persons include Dr. G. E. Collins and R. T. Mertz, who have been
associated with this project from the beginning of this
development, and also Miss L. Y. Chang, Mrs. N. G.
Copeland, Israel Krongold, A. R. Mowlem, J. B. Secrist,
Jr., Dr. R. W. Southworth, and N. R. Wagner.
BIBLIOGRAPHY

[1] Herget, P. "The Computation of Orbits." Published privately, 1948.
[2] Herget, P. and Musen, P. "General Theory of Oblateness
Perturbations for Vanguard Satellites" (memorandum), March
23, 1957. Presented by Dr. Herget in abstract form at the
American Mathematical Society Symposium on Orbit Theory,
April 4, 1957.

of transmission from a United States satellite by one station during each earthcircuit of the satellite. At originally expected altitudes for such a satellite, this
disposition provided an expectation that
simultaneous observations by two or more
stations would be exceptional rather than
usual. Consequently, there was an expectation of being able to triangulate simultaneous observations from these stations
only comparatively rarely. The occurrence
of higher than originally expected altitudes
has made somewhat less exceptional the
recording, and expectation for future,
simultaneous observations. It is planned to
investigate accuracy by triangulation of any
such simultaneous observations, possibly
leading to increased accuracy of observa-

tion. However, on the basis of these considerations, triangulation does not occupy
a very significant role in the present operational use of the Minitrack system.
The method of obtaining a directional
observation inherent in the Minitrack recording involves what might be termed
"triangulation" from displaced antennas,
based upon phase difference. I assumed that
the intent of the question was to gain information about "triangulation" in the usual
sense of this word.
Chairman Rubinoff: L. Elrod of Westinghouse asks, "will it be possible to modify the program while actual tracking is in
process ?"
Mr. Quarles: It is expected that modifications to the program will probably only

Quarles: Preparations for Tracking Earth-Satellites
be made after studying the results obtained
and would not actually be made during the
course of one continuous operation on the
machine. However, it is possible to exercise
a number of programmed options by manual positioning of sense switches, and to
modify the arrangement of macro-operation control cards and content of data information to be used at a later stage, during calculation.
Chairman Rubinoff: The third question
is, "Is there any provision in the subj ect
program for using observational data from
sources other than official tracking stations
such as amateur optical or radio measurements?"
Mr. Quarles: Let me emphasize first
that it is not required that the observational
data be from the Minitrack stations. On
the first slide it was noted that the data
could be either Minitrack data or simulated Minitrack data.
To go one step further, it wouldn't
even be necessary to enter the data in simulated Minitrack form if one had sufficiently reliable data obtained by whatever
means. For example, one could bypass the
whole operation of the first macro-operation which edits and smoothes the data,
and simply enter as a smoothed observation
the direction of a satellite from an observer
at a certain time, if one had such information. So, it is quite flexible as to the types
of data that could be used and in fact several kinds of data have been used. However, the qualification "sufficiently reliable"
stated above makes virtually necessary
what might be termed professional electronic or optical equipment and professional operating personnel, and seriously
limits the usefulness of amateur measurements. Of course, in the absence of professionally obtained data, amateur measurements take on increased importance.
Chairman Rubinoff: A question from
R. Isaacs of Philco, "What reports, if any,
of the programming system are available?"
Mr. Quarles: The first actually published report of the programming system
will be this paper as presented. supplemented in publication by questions and
answers. Later, I expect there will be more
detailed reports of the programming system
available.
Chairman Rubinoff: A question by Mr.
Sumpter of the Department of Defense.
"Why use cards for input to the computer?
Why not go directly from teletype tape to
magnetic tape, then into the computer?"
Mr. Quarles: As described in my paper,
there is an option of providing input observational data to the computer either
directly from punched cards or from magnetic tape prepared on a device which operates independently of the computer. This
option of the use of magnetic tape was
provided in order to enable more rapid
computer processing in the event that a
large volume of input data were to be supplied to the computer at anyone time.
Whether or not this option is exercised,
the preliminary preparation of cards has an
advantage of enabling, together with the
printed triplicate messages also produced
from the teletype operation, convenient
partial editing, selection, or rearrangement

of the data prior to the more extensive editing and processing performed by the computer.
Also, there isn't presently, so far as I
know, a commercially available device for
converting directly from teletype tape to
magnetic tape. Furthermore, there has not
seemed sufficient need for such a device in
this computer application to render it an
important consideration. More rapid provision to the computer of this input data
does not at present seem important. Using
either of the options described, the operation typically requires a small amount of
time for supplying the input data to the
computer in comparison with times for
calculation and for development of output
data on magnetic tape. In particular, the
maximum possible time saving due to bypassing preliminary preparation of punched
cards by using a hypothetical teletype-tapeto-magnetic-tape device would not under
present, or presently expected, conditions of
operation seem to justify the sacrifices of
conveniences of cards described above.
Chairman Rubinoff: From D. ]. Nemanic of Remington Rand Univac, "Can
unknown factors such as density of the
atmosphere be estimated from the differential corrections to the predicted orbit?
Mr. Quarles : Yes, it certainly is possi.ble to estimate some of these factors on
the basis of the orbital calculations in general-not only from the differential corrections. One of the main purposes of the
whole project is to make improvements in
our knowledge about the atmosphere by
examining and studying the deviations of
the predictions from the actual observations.
In particular, the calculations have indicated that the density of the atmosphere is
greater than originally had been expected
at the altitudes attained by the artificial
earth-sateIIites. However, much study of
orbital calculations is expected to be required before reasonably reliable information about density variation at high altitudes is known.
B. Zendle (National Bureau of Standards): Has it been possible to determine
the mass of the Russian earth satellites,
thereby verifying the mass values announced by the Russians? If so, how, and
to what degree of accuracy?
Mr. Quarles: Though I have not been
concerned with this question, it is my
understanding that the presently limited
i!lformation about the density of the atmosphere, and consequently about the drag acting upon a satellite, would have prevented
any independent, accurate verification of
the masses announced by the Russians.
Later, when it is possible to determine
atmospheric drag with greater accuracy,
such independent verifications should be
possible with reasonably good accuracy
from the laws of motion of a satellite
which depend upon both the drag and the
mass.
S. M. Selig (Chemical Corps Eng.
Command, Army Chemical Center, Md.):
What are the computer facilities that the
Russians have set up equivalent to Minitrack?
Is the accuracy of their predictions for

Sputnik I due to more sophisticated equipment than the IBM-704, or to better programming and better fitting or Fourier
series and other essential computations
that you mentioned?
Mr. Quarles: I do not know the answer
to this question. However, on the basis of
any predictions by the Russians which I
have seen, I do not have reason to believe
that they are using anything but comparatively unsophisticated techniques.
M. A. Hyman (Philadelphia, Pa.): Approximately how many points were calculated by numerical integration for each
elliptical traj ectory ? Wha t can be said
about the accumulation of errors during
calculation of an average trajectory? How
long did the computer require for each
trajectory?
Mr. Quarles: The typical time interval
in the numerical integrations performed to
date is one minute, and hence the order of
one hundred steps per earth -circui t. A
thumb rule which may be applied to estimating the accumulation of errors by the
method of numerical integration which has
been programmed is that the error is
approximately the three halves power of
the number of steps, divided by eight, units
in the last place of the digital precision
employed. The programming of the numerical integration enables optional use of
single- or double-precision floating-point
calculations, providing precision equivalent
approximately to eight or sixteen significant decimal digits, respectively. Including
binary-coded-decimal tape output as well as
binary-tape intermediate output, the singleprecision computation requires approximately six minutes per day of predictions
using a one-minute time interval, whereas
the double-precision computation requires
about six times as long. Both of these computations are reduced by about four minutes per day of such predictions at the sacrifice of the binary-coded-decimal tape output.
W. H. Jenkins (ElectroData): You
mentioned 25,000 instructions. Does this
complete the program? Are these operations debugged? If so, you should have
some accurate times for the processing
from the input cards or tape until the final
result of "where to look."
What are the limits of these times?
(For example, three to five minutes, or ten
to thirty minutes.)
Mr. Quarles: The figure of 25,000 instructions was intended to cover completed
instructions and modifications which were
in progress. It is expected that some additions and modifications to the system. will
be planned and developed later due to the
research character of the project.
With the exception of some current
comparatively minor mooifications, all of
the macro-operations descrihed are debugged and all are in use. Tn view of the
great variety of ways in which the macrooperations have been combined in actual
operation, permitted by the flexibility of
the design of the programming system, it
is very difficult to give meaningful over-all
time figures without extensive qualification.
Even for individual macro-operations there
usually are several modes of operation with

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

significantly different times, as indicated by
the answer to the preceding question. (See
the answer to the next question.)
J. Otterman (University of Michigan,
Ann Arbor, Mich.): Has consideration
been given to carrying out macro-operation
No. I on a separate smaller computer?
With all Minitrack stations in operation
what is the percentage of times the computer will be idle (or reserve operation
time)? Can the system handle data simultaneously on two satellites?
Mr. Quarles: No serious consideration
has been given to carrying out macro-operation I on a separate smaller computer.
In the present manner in which the programming system is operated, 24-hour-perday utilization of the 704 would permit
simultaneous handling of the tracking calculations for from six to ten satellites. It
should be possible to increase the number
of satellites which could be "simultaneously" tracked by a significant amount

when further experience has been obtained
and/or by reducing the volume of output
information.
E. H. Weiss (Applied Physics Lab.,
Johns Hopkins University, Baltimore,
Md.): You mentioned that "other checks"
are used to ascertain the accuracy of input
and output. What are some of those checks
and how reliable are they?
Mr. Quarles: I was specifically referring to the data transference checks such
as: check sums used in connection with
transference of information between magnetic tape or magnetic drum and magnetic
core storage; "echo checking" of printed
output; check sums and/or double-punchblank-column checks for certain input
cards. These checks have been found to
be very reliable.
I-{~wever, macro-operation I, for example, also contains checks which would eliminate very unreasonable data, in addition to

various other editing checks as indicated in
the paper.
W. W. Youden (National Bureau of
Standards, Washington, D.C.): How accurate were your predictions?
Mr. Quarles: The present accuracy of
predictions obtained with this programming
system varies with the satellite in question.
It has been possible to obtain considerably
better accuracy for predictions of Vanguard I than for any of the other satellites
launched to date. In part this is felt to be
due to the availability of more better-calibrated Minitrack stations for recording
observational data, but probably primarily
due to the greater perigee distance and consequent lower distortion of the orbit due to
atmospheric drag. More specifically, predictions made for Vanguard I, and commencing shortly after its launching, have
been accurate to within a small fraction of
a minute of time.

Use of a Digital Computer for Airborne
Guidance and Navigation
s.

ZADOF~ AND]. RATTNEW

INTRODUCTION

ECENT developments in computer instrumentation
have permitted a vast increase in speed and complexity with no increase in the size of the largescale digital computers designed for scientific computation.
These developments have also made possible a new application for digital computers, namely, "real-time" computation in the field of control systems.
By way of definition, a digital computer is said to operate in "real time" when it is an integral part of a physical
control system. One of the requirements for real-time
operation of a digital computer is rapid computation consistent with changes in the input physical quantities and the
output data rates required by the system.
Historically, the analog computer has been used in
control applications. However, the analog computer is
intrinsically limited in its ultimate accuracy, whereas
digital-computer accuracy can be increased with little
change in size or basic complexity.1 Problem-handling
capacity can also be increased for the digital machine with
little or no change in its size although this may imply a
change in rate. This latter is far from true for the analog

t Sperry Gyroscope Co., Great Neck, N.Y.
Von Neumann, "The General and Logical Theory of Automata" in "The World of Mathematics," Simon and Schuster, New
York,N.Y., vol. 4, p. 2070 ff.; 1956.
1 ].

computer since its complexity is in one-to-one correspondence with that of the problem it solves.
From this, it follows that there is a point of diminishing returns by way of weight and size in the use of analog
computers over the digital type as problem complexity or
accuracy needs increase.
The development of simple, reliable logic techniques has
reduced the number of vacuum tubes in many computers.
Magnetic elements and transistors are on the verge of
totally replacing those vacuum tubes still required. The
net decrease in size and weight produced by these components is further enhanced by their lesser power requirements. It should be noted that this progress is far from
stabilized.
These component developments affect the size and
weight of analog computers also, but the increases in
speed and reliability in the digital field combined with
demands for more complex real-time computers have made
the digital machine eminently practical for this purpose.
GENERAL DISCUSSION

In real-time computation the problem to be solved is
generally described by a system of nonlinear differential
equations. The analog computer is a direct physical approximation of these equations. When using digital techniques, an equivalent set of difference equations is set up

ZadofJ and Rattner: Digital Computer for Airborne Navigation

and solved by the digital computer. It is necessary that the
solution of the difference equations be asymptotic to the
solution of the differential system and that the same stability
criteria must hold. 2
Finally, the computation must be performed in a time
commensurate with the response characteristics of the
physical system.
In addition to solving systems of difference equations,
the digital computer can be used as a function generator
and a decision device in the control application.
Evidence of the progress of digital computation in the
field of automatic control is its use in airborne systems.
The Cytac system is an example of an airborne guidance
and navigation system using a digital computer in a con~
trolloop.
Cytac is a long-range, all-weather, ground-controlled
navigation and tactical bombing system. It was developed
and tested by Sperry Gyroscope Company under a contract
with Rome Air Development Center and Wright Air Development Center.
The system is built around a hyperbolic radio-navigation aid which was also developed by Sperry3 and is now
known as Loran-C. Loran-C is essentially an extension of
the principles of the standard Loran system which is
presently in use in the Atlantic and Pacific as a long-range
aid to marine navigation. Fig. 1 shows a typical configuration. A master station, Sm, transmits radio-frequency
pulses at a uniform repetition rate. The two slave stations
transmit similar pulses synchronized to the master. A receiver in the service area measures the time differences of
arrival of the master pulses and each of the slaves to
obtain a fix at the crossing of the corresponding lines of
position. Measurements are made only on the groundwave portion of the received signal.
Loran-C achieves long range by using the low-frequency
transmission within the internationally allocated band of
90 to 110 kc. Loran-C is a two-step system and obtains
high precision by making a measurement of the phase of
the radio-frequency cycles within the received pulses,
achieving an instrumental accuracy of 20 to 30 m!Lsec. The
system is fully automatic with respect to both signal acquisition and time-difference measurement and indication.
The output of the time-difference measuring receiver
is continuous fix information in hyperbolic coordinates.
In the Cytac system the digital computer is used to combine
the inherent long-term accuracy and stability characteristics of radio-derived data with the accurate dynamic
character of air-derived data in the form" of airspeed,
compass, and altimeter indications, to provide navigation
infor~ation having the best qualities of each. This was
essentially the first application in which such techniques
were used in long-range navigation and guidance.
2 H. ]. Gray, Jr., "Numerical methods in digital real time simulation," Quart. Appl. Math., vol. 12, pp. 133-140; July, 1954.
3 W. P. Frantz, W. N. Dean, and R. L. Frank, "A precision
multi-purpose radio navigation system," 1957 IRE NATIONAL CONVENTION RECORD, pt. 8, pp. 79-85.

Fig. I-Typical configuration of Loran-C stations.

DISPLAY
DATA CONDITION

I
I

RADIO
BOMBING
AND
NAVIGATION

RECEIVER

1-----+

AurO-PILOT CONTROL SIGNAL

1 - - - -...

BOMB RELEASE SIGNAL

1-_ _'"

BOMB BAY DOOR OPEN SIGNAL

COMPUTER

L____J--:Rf.AD~'OM~OD:tEO~F~OPE~RA~TlO~N'~~.,.,--=--J1

_ _ _ _' " CAMERA SHUTTER CONTROL

SEARCH.TRACK.LOST-SIG=-J

l
I

ALTIMETER

TRUE AIRSPEED
METER

MAGNETIC
COMPASS

r--

1-"I

LOADING DEVICE
BALLISTICS DATA
80MB DOOR OPEN TIME
STATION COORDINATES
STATION CODING DELAYS
TARGET COORDINATES
MAGNETIC VARIATIONS

MERIDIAN CONVERGENCE
COMPUTER PROGRAM

Fig. 2-Block diagram of Cytac system.

Fig. 2 is a block diagram of the Cytac system. The time
differences of arrival of radio signals are measured in the
radio receiver and time-measuring apparatus and are fed
out in the form of shaft rotations. These are converted to
digital data by the computer input equipment. The computer also accepts instrument panel information including
compass heading, altitude, true airspeed readings and mode
of operation of the radio system (i.e., acquisition, track,
or lost signal). These inputs are in the form of shaft rotations, which are also converted by analog-to-digital converters at the computer input, or are in the form of relay
settings. After operating on these inputs, the computer converts the digital information back to analog or relay data
to be used by the other equipment.
The computer output consists of an analog voltage to an
autopilot to guide the aircraft toward a correct bomb release point and a signal to control the bomb-bay doors and
the actual bomb release. Additional computer functions
include:

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

66
1)
2)
3)
4)

Dead reckoning in the event of lost radio signals
Computation of time-difference rates
Computation of expected radio receiver gain settings
Control of certain servo time-constants and gains in
the radio receiver during receiver switch from acquisition mode to track mode of operation
5) Generation of timing signals for control of reconnaissance-camera equipment.

Before take off, the program is stored on the magnetic
drum memory of the computer, together with the constants
required in the equations and the target coordinates.
System operation is initiated by pressing a "start" button
on the pilot's control panel. This puts the radio equipment
into its acquisition mode of operation in which it proceeds
to locate the radio pulses in time and lock on to the received pulses. The computer notes that signals have been
acquired, and normal operation begins.
In the event of lost signals the radio system gives the
computer an indication of this condition. The computer
then ignores further radio time-difference data and dead
reckons on the basis of the last reliable radio information.
It continues to feed computed Doppler rates to the radio
system so that, if the lost signal is due to a temporary interruption of signal transmission or reception, the radio
equipment will be in position to track the pulses as they are
received again. This eliminates the need for the receiver
reverting to the acquisition mode under these conditions.
On the other hand, if the radio-data indication does not
show good signals within a reasonable time, the computer
directs the radio equipment back to its full acquisition
mode.
Fig. 3 is a functional block diagram of the computer
operation. The two-step measurements of time difference
obtained from radio pulse-envelope measurement and
radio-frequency phase measurement within the pulses are
converted from shaft rotations to binary form and introduced at block 1. The coarse pUlse-envelope time differences and fine radio-frequency phase differences are compared and a pair. of consolidated time-difference numbers
are obtained. The time-difference numbers corresponding
to the center lines between the master and slave stations
are subtracted from these numbers to provide a set of
numbers suitable to geometric computation.
A conversion from hyperbolic to rectangular coordinates
is performed in block 2.
A dead-reckoning computation of the aircraft position
in rectangular coordinates is made in block 3, based on
air-speed, altitude, and heading. Heading is derived from
a Sperry J-2 Gyrosyn® gyromagnetic compass, and suitable corrections for magnetic variations and meridian convergence are provided in block 6. Magnetic heading is then
converted to rectangular coordinates and the dead-reckoned rectangular coordinates are compared with the radioderived data in the same coordinates in block 4. A portion
of the difference is fed back through smoothing block 5
to correct the dead-reckoning computation of block 3.

By feeding back only a portion of the difference, the
equivalent of an exponential smoothing factor 4 is obtained
which reduces the effects of random variations. The correction is applied to the dead-reckoned, apparent wind
vector which is substantially invariant for short periods
of time. Aircraft steering control is derived from the deadreckoned solution. Since the long time-constant smoothing
is applied to a quantity substantially independent of aircraft heading, it does not materially affect aircraft stability.
On the basis of the corrected, smoothed position and
velocity, further computation of the steering from present
position to target is done in block 8. The time-to-go and the
bomb-release point are computed in block 9. The distance to
the stations and the velocity relative to the stations (or the
Doppler rates) are computed in block 10. Bomb-ballistics
data for a range of airspeeds and altitudes appropriate for
each mission are stored in the computer, and exact ballistics for actual airspeed and altitude are derived from
the stored data in block 7 for use in the steering and
bomb-release computation.
(MEASURED)
PHASE

DIFFERENCES

. ENVELOPE
TIME DIFFERENCES

STORED
MAGNETIC VARIATIONS

CONVERGENCE OF
MERIDIANS

AurO-PILOT
SIGNAL.
PILOTS

DIRECTIONAL
INDICATOR

AIR SPEED --4~~
AND ALTITUDE

Fig. 3-Block diagram of computer operation.

The next diagram, Fig. 4, shows the two major control
loops of the system and their interaction. One of these is
a fast control loop comprised of an autopilot, airframe
controls and surfaces, heading, airspeed and altitude measurements, and dead-reckoning computation. This control
loop is a conventional autopilot arrangement except for
the fixed delay introduced by the dead-reckoning computer.
It will be noted that this control loop does not contain any
long time-constant or narrow-band circuits.
The second loop is a slow control loop which includes
the autopilot, airframe controls and surfaces, radio timedifference field, time-measuring apparatus and transfer or
smoothing function, computer coordinate transformation,
and dead-reckoning computation and computer smoothing.
This secondary slow control loop has a narrow-band
circuit in the time-measurement smoothing and a very4 R. E. Spero, "Effectiveness of two-step smoothing in digita1
control computers," PROC. IRE, vol. 41, pp. 1465-1469; October,
1953.

Zadoff and Rattner: Digital Computer for Airborne Navigation
narrow-band circuit in the computer smoothing. Were it
not for the fast control loop, severe stability problems
would be encountered. Because of the action of the fast
loop, however, only noise and signals due to errors in the
dead-reckoning computation pass through the narrow-band
circuits. Errors in the dead reckoning may be caused by
wind changes and also by errors in heading, airspeed, and
altitude measurements. Insofar as the problem of stability
is concerned, only heading, airspeed and altitude-measurement errors, and wind changes are of importance. Since
these errors are small, the secondary slow control loop
will remove any cumulative effect of such errors, but will
have a minor effect on the airframe stability.
A tertiary control loop is provided by the computed
Doppler rates which are generated by the dead-reckoning
computation and fed back to the time-measuring apparatus.
This is provided primar.ily for the purpose of providing a
memory function in the time-measuring apparatus during
a lost-signal condition, and does not affect the basic stability
or smoothing considerations.

r---------------------------COMPUTED

:
I
I

DOPPLER

RATES

~-L

_ _~

I
I

I
I
I
I

I
I
I

I
I

(SHORT ·TIME
CONSTANT)

Fig. 4-Control loops of system.

Analysis performed at the beginning of the Cytac development program indicated the need for a computer to
perform certain calculations with stipulated accuracy and
speed. In order to meet the time schedule for the experimental instrument, an airborne digital computer was procured from a contractor.
This computer is capable of performing all required
functions. The computer was programmed for a bombing
miSSIOn and the pilot could select any of three preset
targets. The operating range, distance from all transmitters, and speed 'and altitude of the aircraft were
satisfactorily executed by the computer. In addition, other
auxiliary functions for control of radio-system components were programmed for simultaneous solution. However, doubts were raised as to whether this computer was
the best possible for the .Cytac system. It seemed likely
that a digital computer developed specifically for this sys-

tem and utilizing the latest techniques and components
would prove more satisfactory.
The digital computer used in the Cytac system
is an optimum-programmed, serial, magnetic drum-storage
computer. The drum storage contains 31 order channels,
7 number channels, and one channel for modifiable orders.
Read and write are performed by separate heads, the write
head being disconnected from the write amplifier for those
channels which contain nonerasable program and problem constants. The drum also has one channel for highspeed access.
A novel feature of the drum is the varying of the spacing between the read and write heads of the numerical
storage channels. This is advantageous in reducing access
time.
The wordlength is 16 binary bits plus sign. There are
64 words per channel providing a total memory capacity
of 2496 words. The arithmetic unit consists of an accumulating register, a shift register, an operand register, and
the add-subtract matrix. The three registers are dynamic
circulating registers.
The arithmetic orders include addition, subtraction,
multiplication, division, and square root.
The input-output equipments operate through the inputoutput unit which automatically writes onto and reads from
the drum, utilizing a separate set of heads on those numerical storage channels selected for input and output. A
manual control console which is necessary for test procedures and for loading the drum contains the usual display lights and required switches. There is also an array
of auxiliary equipment for testing, monitoring, display,
problem preparation, and output recording which are not
part of the control system.
The computer was designed with the goal of a fast,
small volume, low-weight computer which would be just
adequate to perform the required functions in about one
second. In consequence, the computer was difficult to program and code. This was justifiable only because the code
would remain unchanged and be retained on the drum
once it was debugged. For those unfamiliar with the problems of an optimum-programmed computer, it may be said
that optimum programming requires extensive juggling
of orders and intermediate storage positions to achieve
adequate results.
Those parts of the computer actually part of the airborne-control loop weigh about 300 pounds and occupy
about 6 cubic feet. It is estimated that this computer could
easily be reduced to less than 150 pounds and 3 cubic feet
by using more modern instrumentation.
While the computer was being utilized, a study was
made to find a more 'suitable computer. This study concluded that a machine using magnetic-decision elements
as a basic unit and a magnetic drum as a storage device
would be better for the Cytac system.
The set of equations for the Cytac system was selected
as that best suited to the characteristics of the digital computer. These equations and their programming were de-

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

signed to m1ll1mlZe the effects of short wordlength and
truncation inherent in the computer. Several alternate sets
of equations developed during the Cytac study were discarded because of the difficulty of programming them for
the computer. The set of equations used gave satisfactory
control of the aircraft and indicated satisfactory bomb
release during several test runs made with this computer
in a B-29 aircraft.
The computer program was also tested, prior to this time
and apart from the actual system, on the Florida Automatic Computer (FLAC) at the Patrick Air Force Base.
Furthermore, complete simulation trials were made at a
Reac installation at the Sperry Gyroscope Company laboratories using the Cytac-digital computer and the radio
time-measuring equipment which had already been thoroughly debugged and flight tested apart from the computer.
The final program used about 1000 orders, 150 constants, and 50 temporary storage addresses. Five channels
of 320 words were allocated for test-program storage. The
actual length of the computation cycle was set for about
one second whereas the time required was about 0.8 second.
The one-second period was chosen since it was consistent
with airframe-stability requirements and the smoothing
factors desired.
This project demonstrated that a digital computer could
be utilized as a very flexible part of a control system with
reliability and size to make it a practical component of an
airborne system.
A typical analog computer arrangement which might
have fulfilled all the functions provided by the digital computer would have required more than 25 servoamplifiers,

21 assorted synchros, 45 potentiometers (many of which
would require special or high-precision windings), 20
servomotors, 8 tachometers, 20 differentials and assorted
gear trains, supporting hardware and electronics, and
power supplies. Using present-day techniques, this equipment could also be expected to weigh at least 150 pounds
and fill more than three cubic feet, with the question of ultimate accuracy left unanswered. On the other hand, a digital
computer designed for the same problem with present-day
techniques would require no more space or weight and
would definitely be capable of meeting the accuracy requirements. Servicing of either type of computer would not be a
pleasure, and reliability and serviceability of each would
be on the same order of magnitude.
CONCLUSION

At the time that the Cytac program began, neither the
equations to be solved nor all the functions to be performed had yet been stipulated. Faced by a short development time, an existing general-purpose computer for the
job seemed most advisable since it provided ease of making changes in programming and addition of control functions with no extra equipment development. Optimum
programming permitted the achievement of computation
time commensurate with airframe-stability requirements
with a magnetic drum-memory computer.
On the other hand, where there is sufficient time to
develop a computer best suited for the job, a special-purpose machine may turn out to be fastest, lightest, and
smallest, with a resulting loss of flexibility in making program changes with the ease provided by a general-purpose
machine.

Some Experimentation on the Tie-In of the Human
Operator to the Control Loop of an Airborne
Navigational Digital Computer System
CORWIN A. BENNETTt

INTRODUCTION

NE of the human operator's most important tasks
in contemporary bombing and navigational systems is crosshair error correction or "tracking."
Due to navigational or intelligence errors, the system's
crosshairs may not fall on the target or other reference
point. When the operator recognizes this error he sends
.

t IBM Corp., Owego, N.Y.

correcting signals, by means of a hand control, to the
computer which then corrects the display.
Typically, bombing and navigational systems have used
analog computers to process the operator's control signals.
However, when a digital computer is utilized, the operator
is faced with the new problem of seeing the results of his
corrections periodically on the display at the solution rate
of the digital computer.
With this "sampled-data" tracking, when the operator

Bennett: Tie-In of the Human Operator to a Navigational System
moves the target across the display it seems to "jump"
from point to point. The apparent discreteness is an inverse function of the inertia in the system. Since the
operator's control signals are accepted by the computer
only at sample times, part of them are ignored. This becomes particularly noticeable at low solution rates. Since
the complexity of the digital computer is determined in
part by its solution rate, it is necessary to minimize this
rate. On the other hand, if the sampling of the operator's
control produces poorer tracking performance with lower
solution rates it should be maximized. The problem faced
by the engineering psychologist is to determine a computer solution rate at which neither of these two goalsequipment simplicity and tracking performance-is unduly
sacrificed.

"recovery time" -the time it took the operator to place
the target under the crosshair to a given tolerance for a
specified initial error. Conventional statistical analyses and
significance test were performed on the data.
Fig. 2 shows a typical curve for the relationship between
recovery time and solution rate. As the solution rate is
decreased, recovery time increases; as the solution rate
increases, performance improves, and recovery time approaches that obtained under analog-tracking conditions
asymptotically. Statistical tests were applied to determine a
specific solution rate which could be considered as yielding
performance that was equivalent to analog conditions. In
most cases this rate turned out to be on the order of 10
cps-a number to which engineers could design in order
to insure no loss of tracking performance with the digital
system.

DESCRIPTIVE EXPERIMENTATION

A series of experiments was carried out over a period
of three years to provide systems engineers with design
requirements for digital tracking. While initially the question of required solution rate was the sole obj ect of investigation, later study was devoted to related sampledtracking problems and to possible ways of circumventing
stringent equipment requirements.
Fig. 1 shows the digital control loop studied in most of
these experiments. The operator's near-continuous control
signals are sampled by analog-to-digital converters. These
numbers are processed by the digital computer which,
among other things, integrates the signals. This integration means that a rate of crosshair movement is proportional to a displacement of the control which is known as
a rate or velocity-tracking control. The computer's outputs
are converted back to analog form and displayed as periodic display changes. Feedback is then provided through
the operator.
DISTURBANCES •

Fig. I-Digital rate-control loop.

In the experimentation such a loop was simulated by
means of an analog computer and relays, a spring-loaded
joystick, and a laboratory oscilloscope. The simulation was
such that sampling in time (at the "solution rate") was
carried out, but sampling in amplitude (at the "quantization level") was not. While quantization could be critical
for tracking, the systems converter resolution was such
that with a rate control no great problem existed.
Laboratory technicians and engineers served as subjects
in each of the experiments. The actual running of a given
experiment would last just a few days, although weeks of
preparation and equipment "debugging" were generally
required. Time records of error were made to obtain performance measures. The usual performance measure was

MEAN
RECOVERY
TIME,
SECONDS(t)

tOO

SAMPLING RATE, CPS(S)
t

=.--!1.-

1950

1960

Fig. I-Population growth, City of Los Angeles.

,
I

1.3

1.2

1.1

::i

..J

i
iii

1.0

11.1

~ 0.9
%
11.1
>
0.8

1950

1960

Fig. 2-Motor vehicle growth, City of Los Angeles.

0.6

0.5

0.4

~
a::
~

0.3

II>

0.2

0.1

1940

1950

1960

Fig. 3-Increase of vehicles per person, City of Los Angeles.

vantage of the abilities of the large-scale automatic computer?" It is the purpose of this paper to discuss some of
the possibilities and difficulties involved.
AREA CONTROL OF TRAFFIC

In approaching this problem the engineer sees a control
spectrum. At one end, control is accomplished by supplying advisory or mandatory instructions to the driver, who,
'in turn, executes control orders. At the other end of the
spectrum, control might be made fully automatic with the

In the manual system, control consists of gathering information on present traffic, comparing present traffic with
stored information on past traffic behavior, and supplying
information to the drivers as to how to proceed. An example of a crude form of this type of control takes place
annually on New Year's Day in Pasadena, Calif., where
extremely large crowds gather for the Tournament of Roses
Parade and football game. For several years it has been the
practice of the Chief of Police to take to the air in a blimp
or helicopter carrying police radio equipment. On the
ground, police cars are stationed at strategic control points.
By observation from the air it can be determined which
thoroughfares are overloaded, and which, if any, can carry
additional flow. This information is used as a basis of radio
commands to the various control points to cause diversion of traffic from overloaded to underloaded thoroughfares. (The police officers give instructions to the drivers
who control the cars. Here instructions are mandatory. )
The City of Los Angeles for nearly two years has had
a helicopter which is used primarily for freeway control
during the rush hour periods. Observations are made of
tie-ups or potential tie-ups and corrective action is taken.
Where a tie-up occurs, information is sent to other drivers
via radio advising them to take a different route.
Several cities have been experimenting with closed-circuit television as a means of obtaining information on
traffic behavior. It is not inconceivable that information
in the form of maps, etc., might be transmitted to the
driver by television.
Thus, with manual control, one of the principal techniques is diversion of traffic from overloaded to lesser
loaded thoroughfares.
Another technique is the control of traffic signals on an
area-wide basis. In Denver and Baltimore, there have
been approaches made to the control of traffic signals in
the city as a whole on the basis of the traffic actually present. These approaches have, however, been based on a limited number of sampling points, a limited number of control possibilities, e.g., signal cycle lengths, and communication with the driver solely on the basis of conventional
traffic signals.
To obtain the maximum benefit from control of traffic
on an area-wide basis with the manual system, it will be
necessary to have many sampling points, several forms of
communication with the driver, a large stored background
of information on traffic behavior within the area, and a
large central computing facility. Stored information must
include anticipated origins and destinations of traffic as a

Gerlough: Control of Automobile Traffic-A Problem in Real-Time Computation

function of time of day and day of week. Unusual pat2000
terns on occasions of special events must also be known.
Characteristics of the complete street network must be
:::>
o
stored in the forms of lists of parameters. Most important
of all, there must be information in the form of equations,
'"
~ 1000
curves, or simulation procedures which will permit the
computation of the flow behavior on a given thorough~
fare under varying conditions. The central computer will
evaluate the existing situation and select the appropriate
control measures.
100
200
300
VEHICLES PER MILE
Traffic signals of the conventional type will still consti·tute an important communication channel between the Fig. 4-Form of density-volume relationship for single traffic lane.
system and the driver, but other forms of communication will play an increasingly important role. There may tain increasingly larger spacings as speeds increase. If
be wide usage of changeable signs to convey special mes- some method were devised whereby vehicles could be opsages to the driver at appropriate times. For instance, neon erated close together without danger of collision, it might
signs, similar to those used on some of the Eastern turn- be possible in an extreme case to operate as high as 264
pikes to inform drivers of snow, ice, etc., may be used to vehicles per mile with no change in spacing as speeds ininform the driver of changes in turning regulations, direc- crease. For instance, Le Tourneau 2 has indicated a techtion of flow on one-way streets, closing of streets, etc. In nique for coupling of long trains of vehicles to be operated
many locations even a series of neon signs may prove to on normal streets and roadways. It seems doubtful, howbe too inflexible, and a sign made up of individual lights ever, that the motorist would ever accept an actual coumay be needed. This sign could display a moving message pling of vehicles, but if a method were achieved by which
similar to that used to convey the news at Times Square vehicles could be operated more closely at least on roads
in N ew York, or more likely as a sign of similar type con- of a freeway type, considerable economy would result.
struction but with the message not moving. Such signs can Close operation necessarily implies, however, operation of
be remotely controlled by a computer. Radios can become all vehicles at approximately the same speed. It is possible
an increasingly important method of communication to the that drivers might accept traveling at a uniform speed if
driver, and it is conceivable that the use of radio might be the benefits achieved thereby were quite clear.
mandatory for the driver just as radio is mandatory for the
Zworykin and his associates, on the other hand, have
flyer who wishes to make use of certain airports and cer- demonstrated by means of a model a method by which
tain air navigational facilities.
vehicles follow a buried conductor and in which speeds
Traffic can be sensed by the techniques to be described can be different for different vehicles, passing being perin connection with the automatic system.
mitted. 3 The guidance principle of the buried conductor
To summarize: In control by the manual system, opera- has been demonstrated for full-scale usage by a vehicle
tion is manual only in that the actual driving of the ve- designed for use in the arctic. 4
hicle is manual. Selections of routes, etc., are performed
From the standpoint of optimum control it would be
by the central computer. Benefits will come through the desirable to maintain a continuous record of each vehicle
diversion of traffic to various routes so that the load is in the system. The computing task involved, however,
spread more uniformly, and through the use of extremely would be so large as to dwarf several SAGE systems, and
flexible signal timing.
thus it does not appear that this would be feasible. Instead,
as much of the control as possible should be carried in the
AUTOMATIC SYSTEM
individual vehicles. One might visualize, then, that the
In an automatic system the driver does not have direct ultimate achievable control system would contain some sort
control of the vehicle and there are many ways in which of guidance, by buried conductor or otherwise, including
marked improvements in traffic flow may be obtained. collision prevention and automatic provision for passing as
Fig. 4 shows the form of curve relating the number of the vehicles come too close together. There should be provehicles per mile in a traffic lane and the number of ve- vision for automatically selecting the optimum routes to
hicles which can flow per hour in that lane.
various portions of the area on the basis of the traffic presIt will be noticed that at the peak capacity there are only
about 2000 cars per hour traveling per lane at a density of
2 "Trackless cross-country
freight train has all-wheel drive,"
around 100 cars per mile. In other words, under present
Elec. Eng.} vol. 75, p. 95; January, 1956.
traffic situations, the amount of unused space in the traffic
3 "Possibilities of electronic control of automobiles explored by
stream is appreciable. The drop off from the maximum Dr. Zworykin," Elec. Eng.} vol. 72, pp. 849-850; September, 1953.
4 C. O. 'O'Rourke, "Electronic trail-finding," Control Eng.} vol. 4,
flow occurs by virtue of the fact that drivers must main- pp. 117-119; May, 1957.
It:

l:
It:

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

ent and the amount of traffic going to each zone. Each
driver on entering the system could, for instance, set a
destination indicator in his vehicle. This could be a tap
switch which would select a signal to be emitted from his
vehicle and picked up by appropriately placed scanners on
the roadway. These scanners would count the number of.
vehicles going to a given zone, and a computer would select the appropriate routing accordingly.5 As routings were
computed, optimum exits for each zone would be established for the current amount of traffic. On approaching
the designated exit for the particular zone of destination,
the vehicle's emitted signal would be sensed, and the vehicle would be automatically guided to the deceleration lane
leading to the exit. Here the automatic control would cease
and manual control would begin. The manual system has
the advantage of a much lower cost in that it makes use of
the computing and control facilities of the human operator; it does not necessitate reconstruction of the existing
highways to provide the facilities necessary for fUllyautomatic control. It can thus be accomplished at an earlier
date and accomplished in a stepwise fashion.
The principal benefit from the automatic system is, then,
increased flow (i.e., increased capacity) on a given facility. A fringe benefit will be the decrease in tension on
the part of the drivers on being freed of the driving task
within the freeway system.
SYSTEM OF THE FUTURE

If, then, one may be permitted prevision, an urban
traffic system in the year 19XX may be something like
this: Long distances will be traveled on a system of freeways where control will be conducted in the automatic
mode. Entrances and exits of these freeways will connect
with one-way streets where parking is prohibited; these
streets will serve as the carriers for intermediate distances.
On these intermediate streets control will be conducted in
the manual mode; drivers will receive instructions by
means of traffic signals, special signs, and radio. Between
these intermediate thoroughfares there will be "local"
streets on which there will be no central control. That is,
the driver will have complete control subject only to conventional traffic signals.
The automatic-control equipment will consist of units
carried by each vehicle, sensing units located at appropriate points throughout the street network, and a central
control unit containing a computer.

ties for passing and for collision prevention (while collision
prevention will be mainly in the automatic mode, provision can be made to permit its use in the manual mode
as well), and 3) radio equipment, either a standard AM receiver or a special receiver for control messages.
SENSING UNITS

Sensing units will have the ability to determine for each
passing vehicle its presence, speed, and destination. The
destination will be ascertained, as previously stated, by
sensing a driver-selected signal emitted from the vehicle.
The sensing unit will have the ability to accumulate data
for later transmission via digital data link on receipt of an
interrogation signal.
CENTRAL CONTROL UNIT

The central control unit will have a programming device which periodically interrogates the various sensing
units. Origin and destination information for the traffic
in the system will be continuously accumulated with appropriate updating.
There will be stored, probably on some random-access
large-capacity medium, inform~tion on past onglndestination movements; information to be stored could
well include such items as time of day and rate of onset
for particular flow patterns, and the optim111n handling of
these patterns. Special provisions for emergency situations such as diversion of traffic from disaster areas could
be provided for in advance. To aid in the compilation of
this stored information it would be desirable for the computer to possess learning ability. One computer can serve
both the automatic and manual portions of the system,or
there can be a separate computer for each portion with intercommunication between the two.
The computer will continually compute control parameters on the basis of the origins, destinations, volumes,
and speeds of existing traffic by means of analytic relationships or simulation routines. These parameters will
provide a basis for searching the stored body of knowledge
in order to find the appropriate listing of optimum control
procedures. These procedures will be read from storage
to the control transmitter which will cause them to be executed. As vehicles pass various exits of the system exit
data will be fed back to the computer as a check on performance.
DEVELOPMENT OF SYSTEM

VEHICULAR UNITS

Each vehicular unit will contain: 1) the destination indicator composed of a signal generator, a selector switch,
and the appropriate radiation equipment, 2) automatic
tracking and control equipment to permit following a conductor in the pavement or other guidance, including facili5 To avoid confusion and disruption of such a system by visiting
vehicles, visitors would be required to stop prior to entering the
system to pick up a map and code sheet so that they could properly
adjust their destination indicators.

Such a system cannot, of course, spring into existence
full grown. It must be built in a piece wise fashion over a
number of years. While much of the computer technology
is presently at a stage which would permit the immediate
start of design, much research and development will be
required in other phases of the problem.
One thing which needs to be decided early is the form
of guidance to be used. This information should be made
available at the earliest possible date to the designers of
new freeways and automotive equipment. It is visualized

Gerlough: Control of Automobile Traffic-A Problem in Real-Time C omputat'ion
that there might be a long transition period in which some
vehicles would be equipped with guidance facilities and
others would not. It would be necessary to set a date after
which no new vehicles would be sold without guidance facilities and a still later date beyond which no vehicle would
be allowed to use a freeway-type road unless so equipped.
Systems of. intermediate streets should be developed as
rapidly as possible and can provide immediate relief to
certain existing situations.
RESEARCH NEEDED

The area requiring the most investigation is the formulation of relationships describing traffic flow and indicating
the measures for optimization. While progress is being
made in theoretical investigations by Lighthill and his associates at the University of Manchester,6 Richards,1
Prager and Newell at Brown University,8 Edie and others
at the Port of New York Authority,9 the staff of the Chicago Area Transportation Study/o and Pipes at the Uni6 M. J. Lighthill and G. B. Whitham, "On kinematic waves, II.
A theory of traffic flow on long crowded roads," Proc. Roy. Soc. A}
London} vol. 229, pp. 317-345; May 10, 1955.
S. C. De, "Kinematic wave theory of bottlenecks of varying capacity," Proc. Cambridge Phil. Soc.} vol. 52, pt. 3, pp. 564-572; July,
1956.
7 P. I. Richards, "Shock waves on the highway," Oper. Res.} vol.
4, pp. 42-51; February, 1956.
8 W.
Prager, "On the Role of Congestion in Transportation
Problems," Div. Appl. Math., Brown Univ., Providence, R.I.;
March, 1955.
-, "Problems in traffic and transportation," Proc. Symposium
on Operations Research in Business and Industry, Midwest Res.
Inst., Kansas City, Mo.; April, 1954.
G. F. Newell, "Statistical analysis of the flow of highway traffic
through a signified intersection," Quart. Appl. Math., vol. 13, pp.
353-369; January, 1956.
-, "Mathematic models for freely flowing highway traffic,"
!. Oper. Res. Soc. Amer., vol. 3, pp. 176-186; May, 1955.
9 L.
E. Edie, "Expecting of multiple vehicle breakdowns in a
tunnel," Oper. Res.} vol. 3, pp. 513-522; November, 1955. Discussion
and author's closure, vol. 4, pp. 609-619; October, 1956.
E. S. Olcott, "The influence of vehicular speed and spacing on
tunnel capacity," !. Oper. Res. Soc. Amer.} vol. 3, pp. 147-167; May,
1955.
L. C. Edie, paper in preparation for presentation at annual
meeting of Highway Res. Board, January, 1958.
10 R. L. ,Creighton,
"Speed volume relationship on signalized
roads." C.A.T.S. Res. News} vol. 1, pp. 6-11; June 21, 1957.

Dis·cussion
J. L. Jones (Chrysler Corp.): From
your paper, I received the impression that
most of the work done has been on traffic
pattern recognition to which an already
known solution may be applied. If this is
true, has any work been done on a mathematical model to which analystical processes may be applied?

versity of California,l1 there is at present no comprehensive theory of traffic flow. To bridge this lack of theory,
development of traffic simulation techniques has been undertaken at the University of California by the writer and
others/ 2 Goode and others at the University of Michigan/ 3 Wong,14 and the staff of the Road Research
LaboratJry in England. 15
Paradoxically, while traffic is a very important and complex engineering problem, the amount of high-grade technical talent applied to this problem has been exceedingly
small in comparison to the technical skills required for the
development of a single large-scale weapons system. There
are few agencies conducting continuing research in problems related with the possible use· of computers in largescale traffic control systems. To the best of the writer's
knowledge, all efforts to date have been supported by
rather limited budgets. If there is to be any major change
in the handling of traffic, such as that visualized in this
paper, there must be early recognition of the need, and the
appropriation of adequate funds by both public agencies
and commercial interests so that the needed research and
development may be accomplished in time to permit an
evolutionary change.
11 L. A. Pipes, "A Proposed Dynamic Analogy of Traffic," Special
Study, Inst. Trans. and Traffic Eng., Univ. of Calif., Los Angeles,
Calif.; July 11, 1950.
-, "An operational analysis of traffic dynamics," !. Appl. Phys.,
vol. 24, pp. 274-281; March, 1953.
D. L. Gerlough, "Automatic computers for traffic control,"
Munic. Sig. Eng., vol. 17, pp. 40-42, 60-62; July-August, 1952.
12 D. L. Trautman, H. Davis, J. Heilfron, E. C. Ho, J. H. Mathewson, and A. Rosenbloom, "Analysis and Simulation of Vehicular
Traffic Flow," Inst. Trans. and Traffic Eng., Univ. of Calif., Los
Angeles, Calif., Res. Rep. 20; December, 1954.
J. H. Mathewson, D. L. Trautman, and D. L. Gerlough, "Study of
traffic flow by simulation," Proc. Highway Res. Board} vol. 34, pp.
522-530; 1955.
D. L. Gerlough and J. H. Mathewson, "Approaches to operational
problems in street and highway traffic," Oper. Res., vol. 4, pp. 32-41;
February, 1956.
D. L. Gerlough, "Simulation of freeway traffic by an electronic
computer," Proc. Highway Res. Board} vol. 35, pp. 543-547; 1956.
13 H. H. Goode, C. H. Pollmar, and J. B. 'Wright, "The use of a
digital computer to model a signalized intersection," Proc. Highway
Res. Board} vol. 35, pp. 548-557; 1956.
14 S. Y. Wong, "Traffic simulator with a digital computer," Proc.
W! C C, pp. 92-94; 1956.
15 Several unpublished technical memoranda.

Do you advocate that automotive manufacturers consider future inclusion of a
"traffic control radio" as standard equipment? If so, what should the salient features of such equipment be?
Mr. Gerlough: Work is being done on
mathematical models, but it is progressing
slowly. Many of the investigators are in
universities and have not had budgets to
cover this type of work. In recent months

there has been some interest shown by one
of the automobile manufacturers, and it is
hoped that this will result in an increasing
rate at which mathematical studies progress.
Yes, I would advocate such a radio
as standard equipment. The specifications
should be worked out by some national
committee which should include representatives of automobile manufacturers, highway people, a.1d the FCC.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Physical Simulation of Nuclear Reactor
Power Plant Systems*
J. J.

STONE, JR.t, B. B. GORDONt,

NE method of control of a heterogeneous boiling
reactor uses the steam pressure in the reactor
vessel to control the height of the water reflector
surrounding the core. As the steam pressure increases, the
reflector height and the reactor power level decrease. Thus,
as the steam load varies, the pressure varies and forces
the reactor power to follow the load changes.
Fig. 1 illustrates, diagrammatically, the reactor system as
controlled by the height of the water reflector. The upper
portion of the pressure shell collects the steam produced
by the boiling within the reactor-core assembly and delivers this steam to the load attached to the system.
Water coolant in the lower portion of the pressure shell
covers the reactor-core assembly. Boiling of this water
within the core produces steam, and the flow of steam upward through the core results in a circulation of water up
through the reactor, and then, after passing through ports
in the annular reflector tank, the water flows down past
the core along the inner surface of the main pressure shell.
The annular reflector tank surrounding the reactor core
is partially filled with water. This water acts as a reflector
for neutrons produced by the core, and as the level of this
water decreases, the reactivity decreases. Openings around
the top of the reflector tank admit steam to the upper surface of the water in the reflector tank. The water in this
annular tank connects, via a pipe, with an external surge
tank in which a reference gas pressure is maintained. Any
excess steam pressure in the reactor over that required to
maintain the water in the reflector system at equilibrium
will cause the following sequence:
1) Flow of water to the surge tank,
2) Decrease in reflector level,
3) Decrease in reactivity,
4) Tendency for a decrease in reactor power, and
S) Return of the steam pressure to its equilibrium
value.
Initially the system was studied by an all-electronic simulation. This simulation required making assumptions concerning the magnitude of frictional forces in the hydraulic
system. It was assumed also that inertial and frictional
terms in the equations of motion of the water in the reflector system were determined primarily by the size of
the connecting pipe. To determine the validity of these
assumptions, a physical simulation of the hydraulic portion
of the system was undertaken.

* Work

performed under AEC Contract W-740S-eng-92.

t Battelle Memorial Inst., Columbus, Ohio.

AND

R. S. BOYDt

.-----~-----._ To steam

load
Steam chest
Pressure shell

Gas
----OO--supply
Pressure
regulator

Surge tonk

Fig. I-Diagrammatic sketch of heterogeneous boiling reactor with
reflector control.

A full-scale physical mock-up was constructed as shown
in Fig. 2. This hydraulic simulator consists of a reflector
tank, a surge tank, and a connecting pipe, together with
pressure accumulator tanks coupled with compressors.
Fig. 3 is a schematic of the hydraulic simulator.
The accumulator tanks (labeled A in Fig. 3) are 16
cubic feet ASME-approved SOO-psi air pressure vessels
each mounted above, and connected to, a 3-hp, SOO-psi air
compressor. These tanks provide SOO-psi air to the reflector tank and surge tank as needed.
The surge tank (labeled S in Fig. 3) is a 36-inch diameter tank so constructed to allow for hydraulic coupling
pipes up to 6 inches in diameter, and to have various connecting ports for mechanical control valves and relief
valves. An inlet air pressure regulator is used to reduce
SOO-psi air from the accumulator tank to 300-302 psi in
the surge tank. The outlet air pressure regulator is used
to release air from the surge tank when the pressure increases above 29 S psi in the surge tank.
The reflector section consists of the simulated reflector
vessel (labeled R. in Fig. 3), two pneumatic control valves
with a controller, a capacitance-type water level indicator,
and the necessary safety relief valves. The reflector tank
has a cross-sectional area of five feet2, and the water level
can be raised two feet from the low portion without any
interference from inlet air or water connections. One-half
inch pneumatic control valves are used on the inlet and

Stone, Gordon, and Boyd: Simulation of Nuclear Reactor Power Plant

where

P = reactor power, btu/sec,
i=6

(3 =

L: (3i,

i= 1

(3i = fraction of neutrons produced each mean lifetime that are delayed in the ith group,
I = mean lifetime, 10-4 sec,
Ai = decay constant for ith delay group, sec-I,
So = term proportional to neutron source,
Ci = term proportional to concentration of ith delay
group,
k = effective multiplication factor.

Fig. 2-Full-scale hydraulic-system mock-up.

ELECT~OIllIC

SIMULATOR

Fig. 3-Schematic of hydraulic simulator. R
simulated reflector
vesse~, S = surge tank, A = pressure vessels, H = hydraulic
pressure control valves.
couphng, V

outlet lines to the reflector tank, and are both controlled by
the same pneumatic signal from a pressure controller.
Safety relief valves are provided on each tank and are
set at 505 psi on the accumulator tanks and 340 psi on the
surge and reflector tanks.
The design parameter of the physical simulation is the
hydraulic coupling (H in Fig. 3). It is necessary to install some sort of damping in this portion of the system to
provide stable operation. The purpose of this investigation
is to determine an acceptable means of damping this system.
An analog computer was used in the analysis and evaluation of this reactor. The computer was used to solve the
equations describing the reactor kinetics, reflector reactivity, and steam pressure under various load conditions.
A standard group of nuclear kinetic equations were employed in the study of this reactor. These are
dP

The Battelle analog facility has a self-contained "nuclear kinetic feedback unit" to solve these equations. The
use of this unit requires only two operational amplifiers,
and saves considerable setup time.
For the purpose of this evaluation, it was assumed that
boiling commences at the point where the water temperature reaches the saturation temperature and increases,
linearly, in intensity as the water temperature increases
beyond this point. The rate of change of water temperature was computed as the difference between power produced by the reactor and power used to convert water to
steam. From these relationships the rate of steam production was determined.
The rate of change of the weight of steam in the steam
chest is proportional to the difference between the rate of
steam production and the rate of steam used to satisfy the
power demand. The pressure in the steam chest was determined from the weight of steam and the volume of the
steam chest. This volume varies inversely as the height of
the reflector, since an increase in the volume of water in
the reflector leaves less volume to be occupied by the
steam. A voltage proportional to this computed pressure
was fed to the hydraulic mock-up as the pressure demand
signal.
Two main factors affect 'Ok in this system. These are
reflector worth, which is a function of reflector height, and
steam-void fraction, which is a function of power level and
pressure. The functions used were obtained from experimental data.
The electronic and hydraulic portions of the system
were then coupled together (as shown in Fig. 4) to complete the simulation. A pressure demand signal from the
computer was used to determine the set point of the controller. The actual pressure in the simulated reflector tank
was compared with the set-point pressure, and the error
determined the pneumatic signal to the control valves. If
the pressure in the reflector tank were lower than the demand pressure, the control valve to the accumulator would
open to increase the pressure. Conversely, if the pressure
in the reflector tank were too high, the control valve to the
atmosphere would open to exhaust the pressure. The control valves were adjusted so that at set-point pressure both
would be slightly open.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

r--·------·-------J
I
I

i
I
I
I
L _

":':E...:.!R~I=- ~M~A~R_

_ _ .J
I

I
,

I
I
I
I
I

I
I
I
I

_____

a::

I
MOCK-UP
_______
__

L-

I
I
I
IL

Fig. 6-Expanded view of the hydraulic coupling.

.5 £0.229

Q.>
O>+-

§~

.cQ.>

Fig. 4-Block diagram of physical simulation.

OJ:

'en

Q.>
"
O>+c.t::

o r-O.229

r
~V

o.~

.t::Q.>

c
Q.>

r---i""

I'---

0.
s

-4/.4

-13.8

(\ ""'-/
~

g'
0

-27.6

....-

5
.Q.>

g'
~.2

/
./

----

/
y

I::

\..../

~'.2
0

Time.sec

+/000

.Q.>

Use of five baffle plates.
Step change in power demand,50-IOO per cent.

\/
l

/ \..

l./

:u3
a..

lJ aJ

+2000

+ /000

~i ' -v
I

)
J

Q.>

.,/

~.J
a

~.J

-27.6

\
\

Q.>

g'
.c

-13.8

0:.

Q.>

a::

;\\

+0.229

Fig. 7-Responses of damped system .

-----...

----i

'-"

Time,sec
Step change
per cent

power demond,50-IOO

Fig. S-Responses of undamped system.

The pressure controller used in this simulation employed proportional-plus-reset (integral) type control.
Both the proportional band and reset rate were set at their
minimum values. In addition, because of an undesirable
time lag between the pressure in the simulated reflector
tank and the demand pressure, an anticipation circuit was

included. The output of this circuit was added to the pressure demand signal to produce the controller set-point signal. This was effectively rate control. This circuit was adjusted for optimum response of the controller.
In order to complete the loop, a signal proportional to
the height of the reflector had to be fed back to the computer. The water level was indicated by a capacitance-type
height gauge. The output of this instrument was sent to
an electronic recorder. The signal from a precision poteniometer geared to the recorder drive mechanism was used
as the height indication required in the electronic simulation.
To examine this system, the simulated reactor is brought
up to power manually. The power demand signal is adjusted to design point power. When the demand pressure

Stone, Gordon, and Boyd: Simulation of Nuclear Reactor Power Plant
to the simulated reflector reaches the operating level, the
system is put on automatic control. The system thus far
described tends to oscillate. Fig. 5 shows the responses of
power, pressure, and reflector height for the undamped
system.
To establish the required frictional forces for stable operation, the following configurations in the hydraulic
coupling were attempted:
1)
2)
3)
4)

Various concentrations of steel wool,
Four 2-inch 90-degree elbows,
A I-inch orifice in a 2-inch pipe,
A system of baffle plates.

Discussion
N. Irvine ( Convair): Since you bring
the system up to rated output manually, I
take this to imply that control is most
applicable over a limited range. What are
the difficulties of control from start?
Mr. Boyd: Our study involved control
over the operating ranges of 100 per cent
of power to 10 per cent of power. However, when a nuclear reactor is brought
up from zero power it becomes very important that this so-called start-up is very
carefully handled.
We presented an illustration showing
the nuclear kinetic equations which indicated that the rate of change of power was
proportional to power. This was solved
using a special unit involving only two
operational amplifiers. Most engineers are
familiar with the fact that two amplifiers
in a loop will tend to be unstable, and consequently it's very easy for the output of
this system to go exponential. This is true
of reactors if there is too much disturbance
or error in the initial start-up; where the
power is very low, the reactor power could
be exponential. This is considered in reactor technology as the period. It turns out
that the period is nothing more than the
amount of time it takes the reactor power

Stable operation was achieved with the use of the baffle
plates. This coupling is shown in an expanded view in
Fig. 6. It consists of a section of 3-inch-ID tubing 9 inches
long with inserts and spacers to damp the flow of water
through the tube. The inserts are made of 14-gauge brass
and have 16.0-inch diameter holes drilled in each insert.
These inserts can be placed in the coupling in various combinations of spacing up to 18 inserts.
The optimum responses occurred with the use of five
inserts. These results are shown in Fig. 7.
With this stable system, the effects of changing the void
coefficient and the incremental moderator worth of the
reactor were examined.

to increase by a factor e. At very low
power, a very short period could cause a
reactor to go super critical in a very short
time.
Consider a period of half a second or
even less, in which case an operator having
to react to the situation might not be able
to react fast enough to control the reactor.
Consequently, at reactor start-up, which is
specifically mentioned here, control is done
manually. For our simulation, which is
direct analog, the feedback unit will not
operate at extremely low power levels because of this tendency for the power exponential.
O. Updike (University of Virginia,
Charlottesville, Va.): When the reflector
liquid leaves the reactor vessel, discharging to the surge tank, it is saturated and
any lowering of temperature should cause
some steam to "flash off." Could you go
into more detail as to how this flashing was
handled in the simulation?
Mr. Stone: In direct answer to the
question, this condition was not considered
in the simulation, so no detail could be gone
into there. However, the liquid from the
reactor does move out of the reactor toward the surge tank when the pressure in
the reactor is above the equilibrium value,

or when it has just risen from the condition at which it was being maintained to
some higher pressure. This would imply
first that the saturation temperature has
gone up; then, that the water going from
the reflector tank towards the surge tank
is at a saturation temperature for the pressure. The drop in pressure, primarily in
the pipeline, would perhaps be the result of
1) the change from a higher to a lower
elevation and 2) velocity heads for the
flow velocities. The resulting pressure
changes in either case were not of a maj or
amount in the pipe itself; consequently we
did not feel that they constituted a problem.
The simulator was conducted with the
amount in the pipe itself; consequently, we
didn't have to consider flashing in the experiment itself.
H. T. DeFrancesco (Westinghouse
Electric Corp., Baltimore, Md.): What
amount of time was required in the study
and programming phases of the simulation?
Mr. Gordon: Approximately three man
months went into the study and programing phase of the simulation but by far
the larger portion of that was in the study.
The actual programming phase required
about one man week.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Application of Computers to Automobile
Control and Stability Problems
ROBERT H. KOHRt

URRENT advances in automobile stability and control studies can be credited largely to modern electronic computers. Solution of the lateral control
problem, like problems from other areas of automotive
technology, has been deterred by the extreme complexity
of the automobile. Some understanding of static or steadystate stability was obtained in the late 1930's, but the solution of dynamic response to steering input was not obtained until the advent of modern high-speed computing
equipment.

AUTOMOBILE

RESPONSE

LATERAL
RESPONSE

STEERING GEAR FEEDBACK
TO OPERATOR
AUTOMOBILE
FEEDBACK TO

RESPONSE
OPERATOR

Fig. I-Block diagram of car control system.

THE AUTOMOBILE STABILITY AND CONTROL PROBLEM

Stability and control of the automobile, sometimes called
the handling problem, is concerned with providing an automobile with the proper steering behavior. It is really two
problems: 1) providing a vehicle with directional sense so
that it will run straight of its own accord, and 2) providing sufficient steering control so that the automobile may
be easily steered along some desired path.
In its simplest sense, the study of automobile stability
and control is the study of the lateral motions induced in
a car by the steering inputs of the driver. The entire system of automobile steering response, shown in Fig. 1, consists of a driver, a steering gear, and an automobile. The
first block is the driver whose information input sets the
system in operation. This may be the desire to go straight,
to pass another car, or to turn a corner. In response to the
information input, the driver decides to do something and
as a result applies a torque to the steering wheel which
turns the steering wheel to a given position. This action,
working through the steering gear, turns the car's front
wheels to some angle and finally the car begins to change
its path down the road. The automobile's change in path
is called its lateral response. Besides the flow of effects
forward from the operator to the lateral response, there
are also several feedback loops. For example, the lateral
response is fed back to the steering gear as a torque, and
to the operator in the form of visual inputs and lateral
acceleration. There is also steering-wheel torque feedback
from the steering gear to the operator.
Although the human operator steers his automobile by
a combination of steering torques and steering displacements, it is possible to study the responses to these two
types of steering inputs separately. The stability and control characteristics associated with a fixed steering wheel
or the response produced by a steering-wheel displacement
are called the "fixed control" characteristics. Conversely,
the characteristics associated with a free steering wheel,
t General Motors Corp., Detroit, Mich.

/"
/.9 = tan -1 sideslip velocity
forward velocity

z
Fig. 2-Axis system for simplified automobile.

or the response produced by a steering torque, are called
the "free control" characteristics. The remainder of this
paper is concerned with the block labeled "automobile"
and particularly its lateral fixed-control response to frontwheel steering inputs.
DEVELOPMENT OF EQUATIONS OF MOTION

The automobile motions, of concern in studying lateral
response, are shown in Fig. 2. They are yawing, rolling,
and sideslipping, and are defined as follows:
1) Yawing is the angular velocity about the vertical
reference axis (OZ) and is denoted by the symbol r.
2) Rolling is the angular velocity about the fore-andaft horizontal reference axis (OX) and is denoted
by p.
3) Sideslipping pertains to the side velocity and is described by the sideslip angle ~.
The axis system is effectively fixed in the unsprung
mass (wheels and tires) so that Y and Z axes do not roll
with the sprung mass (body, frame, and engine), but remain parallel and perpendicular to the road surface respectively. The car is considered to have a constant forward velocity along the X axis, and for the small angles
of sideslip usually encountered, the sideslip angle ~ is defined as the side velocity along Y divided by the forward
velocity along X.

Kohr: Application of Computers to Automobile Control and Stability Problems
A first attempt was made in 1953 to describe the lateral
motions of a car by use of several differential equations. 1
Later that same year, General Motors enlisted the aid of
the Cornell Aeronautical Laboratories. Cornell's extensive
experience with stability and control problems in the ~ir
craft field, particularly their advanced instrumentatlOn
techniques, was brought to bear upon this problem. 2 The
result of this joint effort was a verified set of equations
which describe a car's lateral response to steering inputs.
The set of three simultaneous linear differential equations
which represent a car's handling motions is shown in
Fig. 3.
.
.
In each of these equations, the mass, lllerha, and acceleration terms are given on the left, while the forces and
moments that cause the acceleration are given on the
right. The forcing te~ms on .the right ea~h cO.n.sist of. the
product of some mohon vanable and a sta~lhty d.envative." For example, in the side-force equatlOn, ~ IS the
front-wheel steer angle put in by the driver, and Yo is the
stability derivative with the dimensions of force per unit
angle of front-wheel steer. The stability derivatives appear
in each equation and are composed of various car parameters, like tire lateral stiffness, weight distribution, suspension characteristics, and various other terms .. In all,
there are 20 car parameters included in the equatlOns of
motion. In deriving these equations, it was assumed that
the tires, springs, and shock absorbers all behave linearly
and that the car is operating on a flat road with no wind
blowing so that the only force input to the system is caused
by the driver's turning of the front wheels of the car.
MEASUREMENT OF VEHICLE PARAMETERS

The initial step in the experimental program consisted
of determining the actual values of the mass, chassis, and
tire characteristics of a particular automobile. The yawing
moment of inertia, 1zJ was measured by hanging the car
on four cables and swinging it as a multifilar pendulum.
By measuring the frequency of the yaw osci11~tion: it was
possible to compute the yawing moment of lllerh~. T~e
rolling moment of inertia, I %J and the product of lllertla
linking the yawing and rolling motions, 1xz were both
determined by oscillating the car on knife edges about a
horizontal axis. 1% was found from the rolling oscillation
frequency and l%z from the yawing moment produced
by the rolling oscillation. The total weight ~nd the lo~gi
tudinal center of gravity locations were obtamed by a SImple weighing process.
Chassis characteristics, such as roll-spring rates, rearaxle roll steer, and front-wheel camber due to body roll,
were determined by standard General Motors Proving
Ground tests. The damping characteristics of the shock
absorbers were determined with a stroking machine, and
the damping produced by the shock absorbers in the suspension system became a simple geometrical calculation.
J

1 R
Schilling, "Directional control of automobiles," 1. Indus.
Math. Soc,) vol. 4, pp. 64-77; 1953.
2 W. F. Milliken, Jr., "Dynamic Stability and Control Research,"
Cornell Aeronautical Lab., Report No. CAL-39; 1951.

Side force equation

MV(g

+ r) + M8 hp =

Y{3~

+ Yrr + Yoo + Ycf>CP

Yawing moment equation
/z; - /xzP

= N{3~ + Nrr

+ Noo + Ncf>cp

Rolling moment equation

IxP

+ M8hV(~ + r)

- Ixz;

LpP

+ Lcf>cp

Fig. 3-Lateral equations of automobile motion.
Variable of Motion

Sensing Instrument

-----1-----------------_________

1) Left front-wheel position,
OL
2) Right front-wheel position, OR
3) Steering-wheel position,

Angular potentiometer
Angular potentiometer
Angular potentiometer

osw

4) Lateral acceleration,
5) Roll attitude, cp

'Y]y

6) Pitch attitude, ()
7) Angular yaw velocity, r
8) Angular roll velocity, p
9) Forward velocity, V

Statham lateral accelerometer
M innea polis- Honeywell a ttitude gyro
Minneapolis-Honeywell attitude gyro
Doelcam rate gyro
Doelcam rate gyro
Fifth-wheel-generator set

Fig. 4--Measured variables and associated transducers.

Tire side force and moment characteristics were supplied by the tire manufacturer who obtained the data by
running the tire on a moving drum. These tire characteristics were determined for wide variation in tire pressure
and in the load carried by the tire.
In addition to providing numerical data to insert in the
equations of motion, these tests demonstrated that the
assumption of linearity for the various car parameters was
valid for lateral motions of a reasonable magnitude.
VERIFICATION OF THE EQUATIONS

The equations of motion were verified by response tests
made with an instrumented 1953 Buick. The instrumentation that was required to measure the car's lateral response
is shown in Fig. 4. Both left and right front-wheel positions were measured and averaged to yield the effective
front-wheel angle. Since it was not convenient to measure
the sideslip angle, the total lateral acceleration along the Y
axis was measured with a lateral accelerometer. The pitch
attitude is not a lateral degree of freedom, but its measurement was made to determine whether any coupling occurred between the vertical and lateral motions. The outputs of the various motion-sensing instruments were recorded on an oscillograph.
In the early stages of the experimental work, it was
assumed that information would be recorded in the frequency range of 0 to 10 cycles per second. After the first
shakedown runs, it was discovered that both engine vibrations and wheel vibrations at the natural frequency of the
wheei on the tire, were being picked up by the various
transducers. These unwanted vibrations were removed
with low-pass filters utilizing a tuned galvanometer so that
only frequencies from 0 to 3 cps were recorded.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

TEST RUN '12
APRIL 8, 1954
TRANSIENT RESPONSE OF 8UICK TO PULSE INPUT
OF FRONT WHEEL ANGLE

Fig. 6-Rolling sphere harmonic analyzer.

F(jw) =

Fig. 5-Typical test response.

Response tests were conducted by stabilizing the car's
forward speed, then performing a steering maneuver and
recording the resulting car responses. The record of a
typical test is shown in Fig. 5. This record shows the transient responses induced by a pulse steering input. All the
motion variables are shown here. It is of interest to note
that there was no appreciable change in the car's pitch attitude during this test. Numerous tests were made using
both step and pulse steering inputs. The "step" and the
"pulse" steering inputs used in this test work are only approximations to the mathematically pure step and pulse
inputs used by servo engineers. The use of these two inputs with different harmonic content made certain that all
of the frequencies of interest were introduced as inputs
at some time during the experimental program. In addition to varying the inputs, response data were obtained in
which the stability derivatives were varied from normal.
With these response data in hand, a comparison was made
between the actual measured responses and the responses
predicted by the differential equations.
COMPARISON OF THEORY AND EXPERIMENT

This comparison was made on the basis of frequency
response, that is, the steady-state response of the automobile to a sinusoidal input of front-wheel steer angle.
The theoretic frequency response was obtained by applying the Laplace transformation to the system of equations,
and then replacing the Laplace operators by jw where (v is
the frequency of front-wheel oscillation and j is V - 1.
The determination of yawing, rolling, and sideslipping
frequency responses was then a matter of algebraic computation which was quickly accomplished by use of a digital computer.
The experimental frequency responses were determined from transient responses like the one shown in
Fig. 5. This procedure is based on the use of the Fourier
integral which, under certain conditions, enables a time
function of a system f(t) to be transformed into a complex frequency function F(jw). The Fourier integral may
be expressed as

f. rXJJ(t)e~iwtdt.

Since this integral must be evaluated along the interval
from zero to infinity, it is necessary to know the behavior
of f(t) for an infinite time. This is most easily arranged
by applying a disturbance to the system such that f(t)
reaches a steady value in some finite time, T. Under these
conditions, the frequency function F(jw) may be broken
into its real and imaginary parts as:

F(jw) = R

+ jJ

where

R = IT J(t) cos wt dt - JT sin wT
o

= - fT J(t) sinwt dt - JT cos wT.
o

These integrals may be evaluated in a number of ways. 3
One convenient method utilizes the rolling-sphere harmonic analyzer shown in Fig. 6.
This method simply involves following the curve to be
analyzed with a cross hair eyepiece that is attached to the
analyzer. As the eyepiece is moved along the curve from
the starting point (initial conditions zero) to the point
where steady state is reached, it actuates a number of
rolling ball integrators. Each integrator is equipped with a
recording dial, and after the eyepiece has completed the
traverse of the curve, the individual dials produce readings which are proportional to the real and imaginary components of the Fourier integral. The machine used in this
work produces five harmonics for one traverse of the transient curve.
In order to obtain experimental response data for comparison, it was necessary to analyze the input to the system (steering angle), and the various system responses.
Once this was done, a comparison was made between theoretic and experimental responses.
Fig. 7 shows the excellent agreement between the predicted yawing velocity response and that actually obtained
on the road. Good agreement was also obtained for the
rolling and sideslipping motion. It should be pointed out
that the use of the frequency-response technique has two
3 J. M. Eggleston and C. V\T. Mathews, "Applications of Several
Methods for Determining Transfer Functions and Frequency Response of Aircraft from Flight Data," NACA Report 1204; 1954.

Kohr: Application of Computers to Automobile Control and Stability Problems

v civg." 46.3 ft. per

sec.

l'J_

l't

!..L

0""-

P-.i

~• •
•

-MshV

-Msh

(Yaw Accel.i

It
Yr-MV ,<3 (Slideslip Velocity)

-MshV

P (Roll

Accel.)

S
(steer Angle)

Fig. 8-Block diagram of automobile lateral-motion simulation.
0,

-8
I

r----.;

<5
-40
z
«

---

-..i.

~
0

w
~ -80

:t:
Q..

~...

.6.8 1.0

2.0

4.0

FREQUENCY - cycles per sec.

Fig. 7-Theoretic and experimental frequency-response,
yawing velocity.

important advantages: 1) it provides for easy removal of
dynamic effects produced by the filters in the recording
channels, and 2) it provides a more general solution than
does a transient response. It may be noted here that the
first experimental frequency responses that were obtained
did not match the equations exceptionally well, and,
through use of the frequency-response plot, it was possible to determine some additional terms which were added
to the equations of motion.
COMPUTER SIMULATION OF VEHICLE LATERAL RESPONSES

When the equations of motion had been verified, they
were then used to study the effects of the various car
parameters on its lateral response. Both digital and analog
computers have been used in this work. The analog computer has been used only to determine the transient response, while the digital computer has been used to determine the transient response, the frequency response,
and the roots of the characteristic equation.
Analog Computer Studies
The mechanization of the differential equations on the
analog computer follows the standard procedure of summing the various quantities which determine the various
accelerations in the system, and then integrating acceleration to velocity and velocity to displacement.
The block diagram in Fig. 8 shows the general procedure that was used. Any of the motion variables, for example, the yaw acceleration r, is found by summing, horizontally, the products of the term in each box and the corresponding vertical input. The yaw acceleration is thus
found as

Fig. 9-Real-time automobile-handling simulator.

(1)

and is exactly the yawing moment equation that is given
in Fig. 3. The rate of change of sideslip and the rolling
acceleration are calculated in a similar manner. Each row
then represents a summation, with the quantity in each
block being the gain factor applied to the various motion
variables.
The complete simulation of the equations on the analog
computer requires only fourteen amplifiers. Although no
nonlinear equipment is used at the present, future work
will require the addition of some function generators and
multipliers.
REAL-TIME SIMULATOR

In addition to analog computer studies which are often
run in "slow time," a real-time simulator has been of material value in demonstrating the lateral motions caused
by steering inputs. This simulator, shown in Fig. 9, is
composed of a small, special-purpose analog computer
which can be "steered" by turning a steering wheel attached to the computer. This motion causes "steering
angle" voltages to be introduced into the analog computer
circuit. The computer then solves the lateral equations of
motion of the automobile. The voltages proportional to

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE
4

....
2
Q)
a...
en

)

- -

V Yaw

(f)

Putting t = v' and v = t' in the second integral, it becomes

oofv,
0

t'-sv' v' dt' dv',

0::W

...J
W

which differs from the first only by the interchange of
c and s.
Performing the integrations we find

a 2

- - - CONSTANT TIME INTERVALS

...J

- - - - - POISSON-DISfFiIBUTED TIMING

SCALE: 1 JOB-COMPLETION TIME
~

~~:

r--'t..~.r-:-o

1
1

~--'-[~}---o
+20V.

-2'1.

•. +
+ 20'1. ADDED

~---{~l---O
r---~u~J.---o

r---[:1----o

INPUTS

L---[",~1---o

Fig. 1-Standard circuit for computer building block.

and others interconnected to form flip-flops. Altogether,
75 per cent of the computer is composed of this one basic
circuit.
Since so many identical basic circuits are used, it became obvious that added circuit reliability could pay large
dividends in computer reliability, and thus a major reliability effort was applied. Having established the circuit configuration and made the usual laboratory experiments, a
program for the Univac Scientific was prepared using the
circuit-design equations to compute optimum-circuit parameters.2 After establishing criteria for expected parameter
variation with life, the values of all parameters were calculated to give the maximum circuit stability. The results
are impressive. Even after the beta gain of both transistors
has dropped to two thirds of the purchase value, the diode
reverse currents have increased from 50 to 400 ;J.ta, and all
of the resistors have drifted 10 per cent in the worst direction, the circuit will still suffer a 10 per cent voltage variation without failure. Thus, it has been demonstrated that
careful, detailed engineering will produce reliable circuits
that remain reliable as the components age.
COMPONENT RELIABILITY

The selection of components likely to contribute most to
reliability of the computer also required a comprehenSIve program. The decision to use transistors rather than
vacuum tubes or magnetic-core switches resulted from an
extensive investigation during which several small selfchecking computers were built and operated. Substantial
quantities of all types of components were subjected to
heat, humidity, shock, vibration, low temperatures, and
other destructive environments that might contribute information on comparative component reliability.
Particula: care to detect a tendency toward catastrophic
types of faIlure was necessary in this investigation. Components that deteriorate gradually with time would be detected and removed before failure occurred, while catastrophic failures would mean circuit failure every time. The
t~e

2 J. ~ln:an, P. Phipps and D. Wilson, "Design of a basic computer bmldmg block," Proc. Western Joint Camp. Cant.; 1957.

Fig. 2-Automatic transistor tester.

final decisions on component choice had to be based on
reliability and not electrical characteristics. Every engineer had to put reliability ahead of all other design 1 equirements, and circuits had to be redesigned to use less
efficient components where these proved to be more reliable.
Having established those components that were to be
used, it became necessary to establish controls to assure
that only these components would be used in manufacture
of the equipment. Specifications were written covering
every critical component with quality-level requirements
exceeding the most rigid military specifications. Large
samples drawn from every lot of components had to be
subjected to rigid acceptance tests at the .manufacturer's
plant and again at Remington Rand Univac. To insure
compliance with the specifications, Univac quality control
representatives are stationed at each manufacturer's p 1ant
during the production and testing of the components.
The final assurance that only reliable components are
used is the complete test of every component prior to introduction into the computer. In most cases this test is performed on specially designed automatic machines such as
the transistor tester shown in Fig. 2. This unit, with a
turntable arrangement, moves the transistor through a
number of test stations. The test circuitry and parameters
to be measured at each station are programmed on the
plugboard at the upper right. Counters at the upper left
record the rejects on each separate test while the com'Jonents are being sorted into "accept" and "reject" categories.
Extreme precaution had to be taken in the design of the
test machines so that transients or equipment failures
would not cause damage to the components being tested.
This is extremely important in the case of surface-barrier
transistors where even small transients may cause complete failure.
. Continuous improvement of final screening tests is an
Important area where much can be done toward eliminating

134

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Fig. 3-Standard circ?it packagi!lg with vacuum
tube for SIze companson.

the "weak sisters" from component lots. Improvement of
these tests is continuing as results from a large-scale component-testing program are received. 3 This program is being conducted by Inland Testing Laboratories in Chicago,
Ill., and Battelle Memorial Institute in Columbus, Ohio,
and will give life characteristics and data for screening
tests on transistors, diodes, and resistors used in the computer. Approximately 60,000 components are on test in
this program.
RELIABLE PACKAGING

Many of the components tested during the development
program showed tendencies toward deterioration under
certain environmental conditions. High humidity proved to
be the worst offender, and complete protection of the components from humid atmosphere showed prospects of improving reliability. The results of this phase of the design
effort are perhaps the most unique of all the work done to
achieve greater reliability.
A specially designed connector, mounting two etchedcircuit boards, is the standard package for all electronic
circuitry in the computer, and all circuits are mounted in
this fashion. The package design, shown in Fig. 3, provides
positive hermetic sealing. Connector pins are brought out
through glass to metal seals in the base, and a pressure
valve is provided for pressurizing the chassis with dry
gas. The seal around the base is made by induction soldering and may be unsoldered for repair of the circuitry. The
manufacturing process includes a complete bake out of the
chassis under high vacuum (see Fig. 4) to reduce the
relative humidity below 1 per cent after sealing. The relative humidity may be .checked by means of a humiditysensing element inside each chassis.
The highly reliable contact arrangement is shown in
Fig. 5. The arrangement is reminiscent of the knife switch
used in power circuits with a flattened male pin and tuning
fork shaped female contacts. Two completely independent
pairs of contacts in the female connector give redundancy
for added reliability. Of the 100,000 contacts in connectors
aD. R. Bair and P. Gottfried, "Reliability results from largescale testing," Elec. Equip.; January, 1957.

Fig. 4-Chassis are baked under high vacuum prior to sealing.

Fig. 5-Male- and female-contact assemblies used in
special reliable connectors.

used to date, not one single case of poor contact has occurred.
The etched-wiring boards mounted on the connector are
fiberglass-epoxy laminate with rolled copper foil on one
side. Extreme care is observed in selecting and processing
this material so that the finished boards are completely free
of scratches, pinholes, or other defects such as warpage or
contamination. Assembly of the entire chassis is a "cleanroom" operation with temperature and humidity control,
white smocks and gloves for all operators, and strict process control. Policy prohibits touching any component with
the bare hands. Any component that is dropped, even in a
container, is rejected. Rework is carefully controlled and
strictly limited. Complete records are kept of each operation dn each unit including the time, date, and operator
number so that assembly reference can be made during
the routine failure analysis which follows every failure.
Dip soldering is used to attach the component leads to
the etched wiring and also to connect the etched wiring to
the connector pins. These two separate operations are performed on a selective soldering machine .which permits
masking the entire circuit and exposing only the areas
where solder is desired. This procedure reduces the heat
transfer to components and permits attaching the components to the board in one operation and the board to the
connector in a following operation. Results of selective
soldering are shown in Fig. 6. Following the dip soldering
and prior to an electrical test, the completed board is given
155°F
a temperature shock from room temperature to
and then to -50°F. The temperature-shock treatment is
further insurance against marginal components or connections that might later show intermittent failure.

Raymond: A Transistor-Circuit Chassis for High Reliability

135

Fig. 6-Standard package without cover-showing
results of selective soldering.

Final assembly of the sealed chassis includes pressurization with a mixture of nitrogen and helium thus allowing
a standard helium.c.leak detector to be used for checking the
final seal.
With the final seal complete and final electrical checkout
satisfactorily performed, the chassis is installed in the
computer panel as shown in Fig. 7. Insertion is performed
with a special tool to prevent damage to the connector
pins. The chassis is fastened in place with hold-down
screws and is ready for operation.
CONCLUSION

The purpose of this paper is -to present a broad
picture of a reliability-design effort that achieved true re-

Discussion

Since the discussion at the Conference
dealt mostly with types and causes of
failures, an updated summary from December, 1957, through February, 1958, follows.
From completion of the final checkout
May 17, 1957, until March 1, 1958, the
computer has operated 1613 hours. Failures
considered in determining computer reliability were limited to those which, had they
occurred during guidance missions, would
have caused the mission to fail. They have
been seven failures in this category. They
are summarized below.

Fig. 7-Completed packages installed in computer rack.

liability. The salient points may all be summarized in the
one word-attitude. The desire for reliability must be
present in the mind of each person from the chief engineer
to the girl on the assembly line. The finest quality control
cannot make a poor design reliable and the finest design
will not be reliable if the person who builds it is careless.
Careful attention to minor details in the selection of components, the design of the circuitry, the packaging of components, and the manufacturing process can payoff in a
big way where reliability is the most important requirement.

1) Intermittent chassis-the defect has
not yet been located.
2) Intermittent chassis-defective solder
connection inside pulse transformer.
3) Defective chassis-two shorted diodes,
apparently damaged by externally
applied voltage.
4) Defective
chassis-collector-emitter
short in transistor.
5) Defective
chassis-collector-emitter
short in transistor.
6) Possible intermittent chassis-has not
been established as a definite failure,
but is suspected.
7) Two rectifier stacks in the power supply-resulted from improper design in
the switching circuitry.

Transients from the power supply are
suspected as the cause in 4) and 5). No
transistor failures have occurred since
October when' this defect was eliminated.
Two failures previously reported have
been removed from this list after detailed
study indicated that in one case there was
no defect, and in the other case, that the
defect would not have caused a computer
failure.
There have been twelve chassis removed
for reasons which would not have caused
computer failure. Of these, four chassis
were removed because of low gas pressure,
six because of high humidity indication,
and two because of defective indicator
transistors. No other chassis have been removed for any reason.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

136

A Method of Coupling a Small Computer to InputOutput Devices without Extensive Buffers
JAMES H. RANDALl}

GENERAL DESCRIPTION OF COMPUTER

HE computer referred to in this paper is contained
in a package about desk size. It is intended to sell
at a relatively low cost as compared to other general
purpose type computers and accordingly would be considered in the "small" class. 'The basic components used
are transistors and diodes. Although the design was aimed
specifically at business applications, the computer is nevertheless a general purpose machine with internally stored
program. Typical of the applications intended are the problems of small businesses such as payroll, stock inventory,
production control, interest calculations, etc. The computer
can be considered as an extension of the accounting machine system rather than an integrated data processor.

MECHANICAL ACCOUNTING MACHINE

In business applications, one must use all shapes and
sizes of business forms. Therefore, a standard mechanical
accounting machine with a large carriage suited to these
forms was utilized for printing data from the computer
and as a source of keyboard entry into the computer. The
use of the accounting machine provides the additional
advantage that format control of printing on the forms is
taken care of on the accounting machine itself and does not
have to be stored in internal memory.
This accounting machine is essentially a parallel digit
device. The maximum word size has been chosen as ten
digits (numeric only). As is conventional in accounting
machines the keyboard has ten columns of nine keys each,
one column for each digit position in the word. Depressing
a key in any particular column determines the value, zero
through nine, of the digit printed in that digit position.
(Depressing no key causes a zero to be printed.) After all
the desired keys have been depressed, the machine cycle
is initiated and all the digits of the word are printed simultaneously.
Associated with each column of the keyboard is the conventional rack as shown in Fig. 1. During the machine
cycle all ten of the racks are simultaneously driven in a
setting direction parallel to their long axes, successively
passing through positions representing digital values, zero
through nine. Each rack may be stopped at anyone of the
ten positions, depending upon which key is depressed in
that corresponding column of the keyboard. The racks are
connected to the printing mechanism so that the value of
each digit printed is determined exactly by the position
t The National Cash Register Co., Dayton, Ohio.

COMMON

5
~~~!:f;~67
34

1
0
PRINTED CIRCUIT SWITCH

TO PRINTER

Fig. I-Illustration of a rack with accompanying
solenoid and position detector switch.

of its corresponding rack at the time of pnntmg. After
all of the racks have assumed their proper positions they
are held stationary for a period of time sufficient to complete the printing operation and then restored to their
original positions.
For keyboard entry of data, the computer must detect
the positions of all racks after they have been stopped by
the keys, and transfer this parallel information into the
memory. In order to print words from the memory, the
computer must stop each rack at the proper digital position.
This is actually done by a process that pulls in a solenoid
for each rack, stopping it at the proper position as it is
moving in the setting direction.
MAGNETIC LEDGER CARD

In a large percentage of businesses, the data continually
being used in calculations are stored in printed form on
ledger cards. In general, the problem is to take data from
a ledger card, combine it with new data to get the desired
result, and bring the information on the ledger card up to
date. If the ledger card system is to be retained, it is advantageous if the data on the card is also in machinereadable language. To accomplish this, a strip of magnetic
material was added along one edge of the ledger card.
On this strip are recorded magnetically certain controls
and all of the current information printed on the card.
The computer must both record data and read data from
this magnetic strip.
To make the magnetic ledger card really practical for
the applications intended, it is necessary to store 50 to
100 digits of information on the magnetic strip. The
choice of scanning speed and recording density resulted
in a read and record rate of about 400 pulses per second.

Randall : Coupling a Small Computer to Input-Output Devices
There is a data track and a clock track on the strip and the
data are recorded serial digit and serial code (four bits
per digit). To get the maximum amount of data on the
card, a variable word length system is used which requires
that an end-of-word symbol follow each word.
Recording on the card is under control of the computer
and must be done at the rate of 400 pulses per second,
which is also the reading rate.
READING PUNCHED CARDS

In certain applications it is necessary to do some types
of distribution for writing reports, etc. The computer was
accordingly designed to read punched cards which are
sorted on avaIlable commercial equipment.
A modified IBM 026 card punch was used for reading
the cards. This punch has a duplicate feature which reads
one card and simultaneously punches the information
read into the following card. By temporarily disabling
the punches and operating in the duplicate mode, the cards
can be read one after the other at the normal reading station. This data appears as parallel-code serial-digit at a
rate of about eighteen digits per second. The computer
must take the data at this rate and store it in internal
memory.
INTERNAL CONSTRUCTION

137

In the computer are two single-digit registers of four
bits storage capacity each. They are used for certain arithmetic and control operations and also are used as buffers,
as will be shown.
READING DATA FROM ACCOUNTING MACHINE

Entering data into the computer from the accounting
machine is comparatively simple. As previously discussed,
the value of each digit of the word entered into the keyboard is represented by the differentially set position of a
corresponding rack. As shown in Fig. 1, a switch was
added to the machine which has ten parallel conductors on
one surface, extending in a direction perpendicular to the
racks and spaced the same distance apart as the digital
positions of the racks. A wiper on each rack then makes
contact with one of these ten conductors, dependent upon.
the position at which the rack is stopped. The wiper also
makes continuous contact with a single common conductor
for each rack, extending in the direction of movement of
the rack. Consequently, if voltage is applied to anyone of
these common conductors, the same voltage appears, via
the wiper, on the conductor corresponding to the position
of the rack.
When the computer receives a signal that the racks are
in position, it initiates a word cycle that scans the switch
and copies the information into memory at the same time.
That is, the digit selector selects the digit position in
memory in which to write and at the same time selects the
corresponding rack to be examined for its digit position.
Effectively, the computer is taking the parallel digit oneof-ten coded data from the switch, properly encoding and
serializing it and copying it into memory.
Since the computer is operating at the 25-kc rate during
scanning of the switch and the racks remain stationary for
a sufficient length of time to allow the scanning to be
completed, no buffering is necessary.

Both a core and drum memory were considered for this
computer. Since the capacity of the memory is relatively
small (100 words), it seemed more economical to use a
drum. However, with a drum, communication with each
of the aforementioned input-output devices would require
extensive buffering and add substantially to the cost of the
computer. Therefore, a core memory was considered. The
core memory has three major advantages. First, any address can be selected in a very short time. Second, the
read-write operation can be stopped and started almost
instantaneously and at any point in a word. Third, it is
PRINTING WITH THE ACCOUNTING MACHINE
possible to read out a word starting from either the high
Printing a word from internal memory presents a more
or low-order end. By exploiting these characteristics it was
possible to synchronize, rather than buffer, the memory to complex problem. Each rack must be stopped at the proper
the communications devices.
digital position by a solenoid, shown in Fig. 1.
The basic clock frequency of the computer is 25 kc,
A means was provided for generating a pulse each tirrie
making a bit time and the read-write cycle for a core 40 the racks are moved from one digital position to the next.
/.tsec. The word length is ten digits of four bits each, or These pulses are fed into a counter which operates to
40 bits. Reading out of the memory is always done in a produce in coded form a number representing each of the
completely serial fashion. There are two basic modes for digital positions of the racks as they move from positions
this process, "word cycles" and "digit cycles." Initiating zero through nine.
a word cycle causes an entire word of the memory to be
When a signal is received to start the print operation'
read (or written into) without interruption. With two and before the racks move, a word cycle occurs which sucextra bit times for control, this cycle takes a total of 1680 cessively loads the digits of the word to be printed into
/.tsec. A digit cycle reads only one decimal digit, or four one of the digit registers. Each digit is compared to the
bits, from the memory. The length of this cycle is 200 /.tsec number in the counter, and if any digit in the word is'
including an extra bit time for control. These digit cycles a zero, a solenoid is energized which prevents movement
may follow one after the other in sequence to read out an of the rack corresponding to that particular digit. For
entire word. However, any amount of time may elapse be- example, if the five high-order digits of the word were
tween the completion of one digit cycle and the beginning each zero, the five corresponding racks that cause the
printing of those digits would be prevented from moving.
of the next.

138

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

'fhe remaining racks then begin to move in the setting
direction. When they reach the "one" position, the counter
would then be storing the number representing that position. At this time the same word cycle is repeated, but now
in all positions of the word where the stored digits are
equal to "one," solenoids will be energized to stop all the
corresponding racks at the "one" position. This entire
sequence is repeated each time the racks reach a new digital
position.
The racks move at a constant velocity with 8 msec time
elapsing between digit positions. However, the word cycle
_~t each position requires only 1.68 msec, which, with the
aid of high-speed solenoids, is adequate time to catch the
racks. Thus, a timing system and a series of word cycles
operating at the internal 25-kc rate eliminate the need of
a word-length buffer.
RECORDING ON LEDGER CARDS

The synchronizing technique is applied in the following
manner to recording on the ledger card. A card clock is
generated internally at the 400 pulse-per-second rate. This
clock records the clock track on the magnetic strip and also
synchronizes the computer. The card is scanned in opposite directions for read and record; accordingly, the data
are recorded on a card from high-order to low-order digit
so that it may be read from low-order to high-order digit.
Upon receiving a signal to record, the computer initiates
a digit cycle which loads the high-order digit of the first
word to be recorded into one of the single-digit registers.
In the following bit time the content of the register is examined to determine if the digit is a significant one-that
is, a number other than zero. I f not, another digit cycle
follows immediately, loading the next lower order digit
into the same register and the check of the content is repeated. This process continues until the first significant
digit is detected. At this time the word scanning operation
halts and the next card clock pulse that occurs starts the
recording of an end-of-word symbol. When this is completed, the bits making up the digit stored in the register
are transferred to a single bit storage one at a time in
sequence, each time a· card clock pulse occurs. The output
6f this single bit storage as the bits are stored one after
another determines precisely what is recorded on the data
track of the magnetic strip. When the last bit of the digit
has been transferred to the single bit storage, another digit
cycle occurs which loads the next lower order digit into
the digit register. The following card clock. pulse then
begins the serial recording of that digit. This sequence of
events continues until all of the digits have been recorded.
I f another word is to be recorded, the new address is selected in one bit time and the scanning process begins
again.
If the word is found to be all zeros only an end-of-word
symbol is recorded. In this case the ten-digit cycles required to examine the word can easily occur while the next

end-of-word symbol is being recorded. It is not necessary,
therefore, to have any unused space between words on the
card.
By synchronizing the memory to a slow. card clock only
five bits of buffer storage are needed for recording any
number of digits.
READING LEDGER CARDS

Reading ledger cards is essentially the reverse of recording, but with the 400 pulse-per-second clock output from
the card being in control. Upon receiving a signal to start
reading a card, the computer first initiates a word cycle
that clears out (writes all zeros into) the address in which
the first word is to be stored. As the data come from the
cards serially, each bit is loaded into the digit register having the four bit capacity. A check is made to see if this
digit is an end-of-word symbol. If it is not, a digit cycle
is initiated and the contents of the digit register are loaded
into the low-order digit position of the word. In similar
fashion the succeeding digits are first loaded into the register and then into the next higher order digit position in
memory. When an end-of-word symbol is detected, no
digit cycle occurs and the digit selector of the memory is
reset to the low order digit position. The next address in
the memory to be loaded is then cleared. This takes place
before the next bit of data is received from the card. The
process then is repeated for each word to be entered from
the card.
By synchronizing the memory to clock pulses received
from the card, only one digit of buffer storage is needed
in the reading of magnetic ledger cards.
READING PUNCHED CARDS

The punched cards are read in a serial digit fashion
from high to low-order digit. A clock pulse is received
from the reader which signals the computer that the card
is ina position for reading a digit. At this time a word
cycle clears out the address to be loaded, and the output of
the card reader is loaded by a digit cycle into the low order
digit position in memory. When the next clock pulse is
received, the previous digit is shifted into the next higher
order position of the word in memory and the new digit
is loaded into the low order position. All of this can easily
take place between card reader clock pulses which occur
about every 55 msec. When the last digit of a word is
loaded, the word is stored in the proper position. Again,
synchronizing the memory to the card reader clock eliminates the need for any buffering at all.
CONCLUSION

By synchronizing the core memory in a small computer
to input-output devices, the buffering required is greatly
reduced. Additional savings are realized because small
existing registers which are already a necessary part of
the computer can be used as the buffers.

Robinson: Computer-Limited Sampled-Data Simulation
Discussion
M. W. Marcovitz (Burroughs Corp.,
Paoli, Pa.): Is the program for this machine stored in the core memory? If not,
where?
Mr. Randall: Yes, the program is
stored in the core memory.

T. A. Dowds (Burroughs Corp., Paoli,
Pa.): Is this machine available? How are
nonsignificant zeros stored when reading
from punched cards and from ledger cards?
Is this a single-address program?

139

R. A. Wallace (Burroughs Corp.):
What is the method of feeding the ledgers?
Is part of the information used to store
line information? What provisions are
there for checking?

Mr. Randall: No, this computer is not
yet commercially available. The method of
storing nonsignificant zeros is to first clear
(write all zeros into) the memory cell to
be loaded. The data are then loaded into
the cell starting at the low-order end of
the word until a signal is received that
there are no more significant digits for that
word. When reading punched cards this
signal is an "end-of-field" signal from the
card punch. When reading magnetic ledger
cards the signal is an "end-of-word" symbol
read from the card. The instruction format
is not single address, but rather "three-plusone" address.

Mr. Randall: The ledger cards are
driven into the carriage by a mechanism
added to the mechanical accounting machine. Line information is stored on the
card so that it is stopped on the proper line
for posting. There is practically no internal
checking. However, checking can be accomplished by normal programming methods.

The Synthesis of Computer-Limited Sampled-Data
Simulation and Filtering Systems *
ARTHUR S. ROBINSONt

'
T

HIS PAPER concerns the synthesis of systems in
which a single digital computer is to be used in conjunction with an array of output "holds" or filters
either to simulate the dynamic transfer characteristics of
a number 6f linear continuous systems or to filter random
messages from a number of continuous inputs, each of
which consists of a mixture of random message plus
random noise.
A block diagram of such a system is shown in Fig. 1.
The system inputs are the continuous time functions

rl(t), r2(t), ... rN-I(t), rN(t).
The system outputs are the continuous time functions
CI(t), C2(t), ... CN-I(t), CN(t).
Each system output is obtained from an individual continuous output filter. This filter is in turn actuated by
sampled signals periodically derived by the computer as
it moves sequentially from channel to channel.
In systems of this type, in which a digital computer has
available and is to supply continuous data, the computer
operates on sampled data only because a series of sequential operations are to be performed, each requiring a
finite amount of time. A typical simulator channel requires
time for switching between channels and for analog-todigital and digital-to-analog conversions, defined as Sj, and

*Details pertinent to both this paper and footnote I are contained in footnote 2.
t Eclipse-Pioneer, Div. Bendix Aviation Corp., Teterboro, N.J.

r, ttl --+------<>
r. ttl _______

$-1

$-2

I
I
I

DIGITAL COMPUTER
L __________________
J

Fig. I-Simulation system functional block diagram.

time for internal computer data transfers, multiplications,
additions and subtractions, defined as k j • The time required for switching and conversion is generally constant,
whereas the time required for internal data processing
and computation is governed by the complexity of the computer program. That is, in general, k j will be a function
of the number of terms in the computer program for that
channel, so that T, the system sampling period, will be a
function of the total number of program terms the computer is required to process for all channels.
The term Computer Limited has been coined to designate sampled data systems of this type, in which the system sampling period T is a function of limitations imposed
by the computer implementation. Computer Limited sampled. data systems can be contrasted to Data Limited systems, in which the system sampling rate is limited by some
fixed external constraint in the data measuring equipment,
such as the speed of rotation of a radar antenna. It is important to understand this basic difference between sys-

140

PROCEEDIN'GS OF THE EASTERN COMPUTER CONFERENCE

terns, since synthesis techniques that have been derived
heretofore for the optimization of Data Limited sampled
data systems are not applicable to the Computer Limited
problem.
It is interesting to note that Computer Limited sampled
data systems are not restricted to the simulation systems
described in this paper. In general, any system that utilizes
a digital computer in real time is a Computer Limited sampled data system, unless the available input data rate is
slower than the speed with which computer solutions can
be obtained. Many analog computing systems that utilize
multiplexed elements are also subject to Computer Limited
constraints.
An application of Computer Limited theory to control
system synthesis has already been presented. 1 It is suggested that the basic techniques to be presented in this and
in supplementary papers2 can be used as tools in the general synthesis of Computer Limited sampled data systems.
Returning to the stimulation and filtering application.
When the system illustrated in Fig. 1 is to simulate the dynamictransfer characteristics of a number of linear continuous systems actuated by deterministic input signals, the
ideal system outputs are as defined by the block diagram of
Fig. 2, where

r,ltl

hi It I

c, (t)

r-----'

- - - - - ...... - - - - -

----~L _____ -'1----- ------~------

--~--~-----

----~1- _____ ..11 - - - - -

------~-----

- - - - - -. . - - - - -

r-----'

r------,

- - - - ,1- _____ ..11 - - - - -

r-----'
L. _____ -'1-----

----4

-----+-------

-----+------- - - - - -.. - - - - - -

Fig. 2-ldeal simulator input-output relationships
-deterministic inputs.

~--I-------C,(t)

-~--------- ----f.'-....
,----:[-----1----')"
L _____ -'

-------

_______________ { -....L ___ ~-----~_---

______ _

'l"

1-_____ -'

----------- ---.f-'I-----f-----l---------'1/
1- _____ -'
f - - - - I - - - - - - CN_I(t)

h 1(t), h 2 (t), ... hN- 1(t), hN(t)
are the impulsive responses corresponding to the ideal in~
put-output relationships. When the system inputs are random functions of time of known statistical characteristics,
or when the function of the system is to filter random
messages from the continuous inputs, each of which consists of a mixture of random message plus random noise,
the ideal system outputs are as defined in Fig. 3. As
indicated, the ideal filtering response for each channel
would result if a signal equal to the noise input could be
effectively subtracted from the total input, so that only
the random messages
rm1 (t), rm2 (t), ... rmN_Jt), rmN(t)
actuated the ideal filters. Since the block diagrams of the
filtering and simulation problems are almost identical, it is
convenient to visualize the filtering problem as the simulation of ideal filters, and so to combine the discussion of
system synthesis techniques under the single heading of
"simulation."
The ideal system responses shown in Figs. 2 and 3 define
reference outputs against which the performance of the
actual system shown in Fig. 1 must be compared to evaluate the effectiveness of the simulation. When system inputs
are deterministic, the synthesis techniques to be described
permit the system designer to determine the linear digital
computer program that will minimize the integrated error
squared between ideal and actual continuous system outA. S. Robinson, "The Synthesis of Computer Limited Sampled
CC!ntrol Systems," presented at AlEE Computer Con£., Atlantlc CIty, N.]., October, 1957.
2A. S. Robinson, "The Optimum Synthesis of Computer Limited
Sampled Data Systems," D. Eng. Sc. Dissertation, Columbia University, New York, N.Y.; May, 1957.
1

Dat~

Fig. 3-ldeal simulator input-output relationships-random inputs.

puts. When system inputs are random, the synthesis techniques lead to minimization of the mean squared error.
When only the form of the channel output filters are specified, the synthesis procedure permits the filter parameters
to also be optimized.
Returning to Fig. 1, a mathematical model for a given
channel can be obtained by tracing the channel from its
input to its output. At the input the continuous signal r j (t)
is periodically connected to the digital computer input,
where the sampled analog voltage is converted to a digital
number. Each such digital number is effectively an instantaneous sample of the value of the continuous function.
I~ the period between sampling instants is denoted byT,
the conversion of the continuous deterministic input signal
rj (t) to the series of digital numbers

r j ( 0 ), r j (T ), r j ( 2T ) . . .
can be visualized as a process of impulse modulation, with
the area of each impulse equal to the value of the corresponding digital number.
Each time the computer receives new input data it proceeds through a new series of computations and delivers a
new solution at its output after the time delay required to
perform these computations. The computed problem solution is then used to actuate the output filter, characterized
by its impUlsive response, g j (t), so that as the computer
moves on to other simulator channels, the output filter
provides a continuous output response for the jth channel.

Robinson: Comp'uter-Limited Sampled-Data Simulation
The mathematical model for a single channel with a
deterministic input is shown in Fig. 4 in terms of Laplace
and Z transforms. It is convenient to describe the digital
computer by an instantaneous response, characterized
by C*(Z), followed by a time delay, described by Z-k/T.
The transfer function of the continuous output filter is
defined by G(s). The meaning of each of these symbols
is shown more precisely in Table 1. r*(t) describes the
train of impulses at the computer input. Z-1 is, of
course, simply the delay operator e- sT • The Z transform
of the computer input is R*(Z). G(s) is the Laplace
transform of the output filter impulsive response.
Dr. Salzer 3 has shown that C* (Z), the linear program of
a digital computer operating in real time, is physically
realizable only when it can be described by a ratio of
polynomials in Z-t, as indicated in Table 1. Referring
to Table II, this constraint states that the impulse sequence corresponding to the effective computer instantaneous output, characterized by P/(Z), can only be
formed as the sum of appropriately weighted past and
present values of the computer input, and past values
of the computer output. That is, R*(Z)ao corresponds
in the time domain to the input impulse sequence multiplied by ao. R*(Z)Z- 1al corresponds to the input
impulse sequence delayed by T and multiplied by a1.
R*(Z)Z-2a2 corresponds to a delay of 2T and multiplication by a2, etc. P i*(Z)Z-lb l corresponds to the computed output sequence delayed by T and multiplied by
bl . P i*(Z)Z-2b2 corresponds to a delay of 2 T and multiplication by b2 , etc. Only past and present terms can be
utilized in a real time digital computer because future
terms (positive powers of Z) will not, of course, be available. Table II also indicates the existence of the constraint that exists in an actual system implementation
between the number of computer program terms, the
channel computing time kj, and the system sampling
period T. This relationship between computer program
complexity and required computing time must always
form part of the basic problem statement. The relationship need not be linear. Any nondecreasing function
can be accepted, defined. in analytical,graphical, or
tabular form.
Fig. 5 summarizes the relationships between true and
desired system outputs for a single simulation channel
with a deterministic input. The object of the synthesis
procedure is to determine the linear digital computer
program C*(Z) that will result in minimization of the
integral of the continuous error squared. That is, the
quantity

~
R{Sl

141

r*{tl
R*{i!l

Fig. 4-Simulation system single

channel~block

diagram.

TABLE I
BLOCK DIAGRAM DEFINITIONS
00

r*(t) =

r(t)o(t - nT)

(1.1)

Z-l = e-ST

(1.2)

R*(Z) =

r(lT)Z-

(1.3)

I~O

C*(Z)

ap.?J-P

= ~~____ 6 ~(z) = Pi (Z)
1m

bqz-q

b*(z)

(1.4)

R*(Z)

(1.5)

TABLE II
COMPUTER PROGRAM RELATIONSHIPS

(2.1)

Pi*(Z) = R*(Z)[ao

+ alz-1 + Il2Z- + ... apmz-Pm]
+ b z- + ... bqmz-qm]
2

- P/(Z) [bIZ- 1

(2.2)

Z-i,

(2.3)
Tj

= ki

+ Si

(2.4)

(2.5)

T= L T j
i=O

~
I
R(s1

'*(tl
R*(SI

r------------------------l
I
!

~-------+I----~

~ffl

~----+l--------~

I
I
IL _________________________ .JI

Fig. 5-Simulationc:cannel error-determi~istic input.

is to be minimized. Note particufarly that it is the continuous error, squared and integrated, that is being used
3 J. M. Salzer, "Treatment of Digital Control Systems and Numerical Processes in the Frequency Domain," D. Eng. Sc. Dissertation, Dept. of Elec. Eng., M.LT., Cambridge, Mass.; 1951.

to evaluate system performance, not the error at sampling instants only. Thus inter-sample ripple is automatically accounted for. When only the form of G(s)
is specified, it is also required that the optimum parameters for G(s) be determined.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

142

TABLE III

TABLE IV

OPTIMIZATION RELATIONSHIPS FOR COMPUTER LIMITED
SIMULATION SYSTEMS-DETERMINISTIC INPUTS

SUMMARY OF BASIC SYNTHESIS PROCEDURE-SnIULATION
SYSTEMS-DETERMINISTIC INPUTS

(a)

Step
Number

Evaluate

From
(a)
~UkUk(S) =G(S)e-skG( -S)e- sk

!PUkUk*(Z)

=G(S)G( -S)

QE2*(Z) = - 2"(2R*(Z)R*

a*(z)a*(

Ir(S) =G( -S)e-skR(S)Hd(S)

R*(Z)

R(S)

The values of the (Pm,+1)
time functions corresponding
to .fal'E*(Z) (3.1) evaluated
at t=O.

(b)

1 ) If*(Z)a* ( : ) z+Q

+ 2~R Z ---;5'~

(3.2)

.f E2*(Z) = R*(Z)R* (~) C*(Z)C* (~) "(2~UkUk*(Z)
- 2,,(R*

If*(Z)

(~) JIukU/(Z) ----(;-)2
b*(z)b*

~ )Z+q

(±) C* (±) IF*(Z) + ~CdC/(Z)

The values of the gms time
functions corresponding to
PqE2*(Z) (3.2) evaluated at
t=O .

C*(Z)

(3.3)

The value at t =0 of the time
function corresponding to
IE2*(Z) (3.3), with T determined from the stated
relationship between sampling period and computer
program complexity.

(b)

where'
C*(Z) = ~~
.
b*(z)

~kl1k(Z)

Z[G(S)G( -S)]

Simultaneous solution of the
. (Pms+gm,+l) equations defined by steps 4 and 5.

(3.4)
(3.5)

IF*(Z) = Z[G(-S)e- 8kR(S)Hd(S)]

(3.6)

~cdciZ) = Z[R(S)R( -S)HiS)Hd( -S)]
"( .= impulse duration

(3.7)

Detailed derivations of the synthesis procedures to be
described are available. 2 It is possible, without proceeding ,through these derivations, to acquire an under,;.
standing of the capabilities and limitations of the technique, and a physical insight into the fundamental
processes that are automatically brought into play by
the synthesis procedures.
The three key synthesis equations for systems with
deterministic inputs are listed in Table III. In essence,
these equations summarize in transform form all of the
time domain convolutions, integrations and differentiations required to establish the conditions for optimum
operation of a given channel, given the total number of
terms in the computer program numerator (Pm + 1) and
in the program denominator (qm). These equations contain both negative and positive powers of Z, so that
their corresponding time functions have values for both
positive and negative time. The synthesis procedure requires that the value at t = 0 of the time functions corresponding to (3.1) (p=O, 1·· 'Pm), (3.2) (q=l," 'qm),
and (3.3) be evaluated. The results of each of the

(Pm 1) evaluations of (3.1) and of the qm evaluations of
(3.2) are then set equal to zero. This effectively con~
strains the derivatives of the integrated error squared
with respect to each of the computer program parameters to be zero. The resultant (Pm +qm 1) simultaneous
equations are then solved for the (Pm +qm l)'unknowns
(ao, al • . • a pm , bl , b qm ). This establishes the optimum
values for the computer program parameters. When
these values are substituted in the result of the evaluation of (3.3), the resultant number, (still a function of k
and T), will correspond to the system integrated error
squared. Finally, k, which is a function of single channel
computing time, and T, which is a function of total
system computing time, can be evaluated, and the inte~
grated error squared will then correspond to a known
number. Table IV, summarizes the steps involved
in the determination of the optimum program parameters and integrated error squared for a computer program of stated complexity (Pm, qm). In the event that
only the output filter form is specified, the filter
parameters can also be optimized by differentiating the
integrated error squared with respect to each parameter,
setting each equation equal to zero, and solving for the
optimum parameters.
Note that the factor 'Y that appears in the basic
synthesis equations of Table III corresponds to the
duration of the computer output pulse that actuates
the channel output filter. Normally, the duration of this

Robinson: Computer-Limited Sampled-Data Simulation
pulse is important, since the output filter is specified in
terms of its impulsive response, and both the amplitude
and duration of the computer output affect the filter
output. When the output filter consists of a hold circuit
that responds only to the amplitude of the computed
output, and is unaffected by the time it takes to actuate
the hold, a factor 1/1' must appear in the hold transform to compensate for this effect. That is, the transform of a first-order hold is simply

TABLE V
OPTIMIZATION RELATIONSHIPS FOR DATA
LIMITED SIMULATION SYSTEMS
DETERMINISTIC INPUTS

C*(Z) = _____ ~_'_*~l _____ _

(pm = 2, qm = 0),

(pm = 1, qm = 1), (pm = 0, qm = 2).

The selection of a particular procedure can be made to suit
the convenience of the system designer, since the results
are independent of the procedure.
4 G. Franklin, "The Optimum Synthesis of Sampled-Data Systems," D. Eng. Sc. Dissertation, Columbia University, N ew York,
N.Y.; May, 1955.

(5.1)

'Y1 R *(Z)R* (~) ~ + {~gkO/(Z)}+
*

R* (

If*(Z)

W (Z) = - - - - - - - - - - - - - - - -

(5.2)

1R*(Z)R* (~ ) ~ -! ~y kflk*(Z) r

(1 - e- sT )
'YS
Data Limited theory provides an upper limit to the number of potential computer programs that must be considered before the optimum Computer Limited program
can be established. Dr. Franklin4 has shown that the optimum Data Limited C* (Z) is given by the equations listed
in Table V. The derivation of the optimum Computer
Limited program requires consideration of both this Data
Limited solution and all less complex programs, and the
evaluation in each case of the corresponding integrated
error squared, subj ect to the computer complexity-sampling
period constraint. Based on these evaluations, the optimum
Computer Limited program will be evident as the program
resulting in least integrated error squared.
For example, if a given Data Limited solution results
in the program form shown in Table VI, (6.1) Computer
Limited theory would require, in effect, that each of the
potential programs shown in (6.2) through (6.4) be considered. Programs of greater length would not have to be
considered, since they would always lead to greater integrated error squared.· This is the case because Data Limited
theory imposes no penalty for computer program complexity, so that the Data Limited program cannot be improved by increasing the number of program terms. However, when the sampling period-computer complexity constraint is taken into account, a longer pFogram would result in a longer sampling period and therefore in greater
integrated error squared. In general, the advantage of the
Computer Limited approach lies in its ability to indicate
shorter programs, and therefore shorter sampling periods,
than those nominally dictated by Data Limited theory.
In implementing the procedure described above the system designer can, if he wishes, use only the five term program shown in (6.5), solve the required five simultaneous
equations to obtain a general solution to the problem, and
simply reduce the unused parameters to zero when considering each potential program in turn. This approach
is to be compared with the solution of three sets of three
simultaneous equations

143

=W1*(Z)+W/(Z)
·Pole(
p()~s
Inside
Ou tsicle
"-TI~UcitC;;cle~
illkO/'(Z)

{t(Jkll/(Z)j-I!lPYkU/(Z)j+
T

(5.3)

Poles
Poles
Outside
Inside
--Th;;-U~tcir~e-~

TABLE VI
EXAMPLE OF POTENTIAL COMPUTER LIMITED PROGRAMS

ao + alz-1
Data Limited Program C*(Z) = - - - 1- - (6.1)
1 + b1z- + bS;-2
Alternate Programs to be Considered
Program
Period
C*(Z)a_l = ao
Ta
(6.2)

o
C*(Z)b_l = __a_- '
1 + bIZ-I'

C*(Z)b_2 = ao

-t

alz- 1

C*(Z)C_l = ---~--- .
1 + b1z-1 b2z- 2 '
C*(Z)C_3 = ao
alz- 1 a2Z- 2

+
+

(6.3)

(6.4)

General Program Containing All Applicable Terms

+ ~-2
+ b z-

C*(Z) = ao
alz1 -t b1z- 1

(6.5)

When N similar problems are being simulated, the
system sampling period T is N times the channel sampling
period, so that a channel sampling period Tij will result
in a system sampling period T = NT,j. The synthesis of
the entire system can then be accomplished by optimizing
the program for a single channel. When the N problems
to be simulated are not similar it is necessary to consider
each combination of potential single channel programs
that could produce a given system sampling period, and
each potential system sampling period, evaluating in each
case the corresponding integrated error squared for each
channel. The sum of all channel integrated error squared
terms can then be observed and the least value selected as
the optimum system operating point. Further details on
such a procedure are available. 2
When the system inputs are mixtures of random message plus random noise, a slightly modified mathematical

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

144

TABLE VIII

~=,.'" ~ ~",,""<'"

SUMMARY OF BASIC SYNTHESIS PROCEDURE-SIMULATION
SYSTEMS-RANDOM INPUTS

Step
Number

From

Evaluate

E(t)

r ________

f_______________________________,

(a)

i t :
!~

~~:--~--~Cd~tt-)~

I
I

I
I
IL_______________________________________ JI

rr(S)

F*(Z)

F(S) =G( -S)e·kHd(S) rrm(S)

aE2

TABLE VII

--=0; P= 1, . . . Pm.

aa p

(b)

The values of the qms time func-

aE 2

(a)

- - = 0;

tions corresponding to E2 Q*(Z)
(7.2) evaluated at t=O.'

C*(Z)

Simultaneous solution of the
(Pm.+qm.+1) equations defined
by steps 4 and 5.

+-T

The value at t=O of the time
function corresponding to E2*(Z)
(7.3) with T determined from the
stated relationship between sampling period and computer program complexity.

(~)

( 1)2
z

q = 1, ... ,qm

(7.2)

b* -

q = 1, . . . q11lS

ab q

The values of the (Pm.+1) time
functions
corresponding
to
]f/P*(Z) (7.1) evaluated at t=O.

OPTIMIZATION RELATIONSHIPS FOR COMPUTERLIMITED SIMULATION SYSTEMS
RANDOM 1NPUTS

--~- F*(Z)z+q

(S) =G(S)e-·kG( -S)e·k
=G(S)G( -S)

rr*(Z)

Fig. 6-Simulation channel error~random inputs.

2'}'

k k

TABLE IX
OPTIMIZATION RELATIONSHIPS FOR DATA-LIMITED
SIMULATION SYSTEMS-RANDOM INPUTS

C*(Z)
,

Wl*(Z)

'}'

{rr*(Z)}+

= - ----------

F*(Z)
W*(Z) = - - - - - - - - - - {4PqkOk*(Z) }- {rr *(Z) } -

(9.1 )

(9.2)
(9.3)

= Wl*(Z)+W2*(Z)

(7.3)

Poles
Inside

Poles
Outside

T~tCirck
(b)

rr*(Z)

~kO/(Z) = Z[G(S)G( -S)]
rr*(Z) = Z[rr(S)]
F*(Z) =;:: Z[G( -S)eskHd(S)rrm(S)]

(7.4)

cdciZ) = Z[rmrm(S)Hd(S)Hd( -S)]

(7.7)

(7.5)

{rr*(Z)r{rr*(Z)}+
i
i
Poles
Poles
Outside
Inside

--Th; ucitC~cle~

(7.6)

model must be used to define the error between true
and ideal system outputs for a given channel. As illustrated
in Fig. 6, the main difference lies in the fact that the ideal
output Cd (t) is assumed to be derived from an ideal transfer characteristic h a ( t) that is actuated by the random
message alone. Since the stationary random signals are assumed present over all time; the 'continuous mean squared
system error, rather than the integrated error squared, is

to be minimized. The techniques required to synthesize a
system of' this type are essentially the same as those required to synthesize a system with deterministic inputs,
although a different set of optimization equations now
apply. The optimization equations for the synthesis of
simulation systems with random inputs are presented in
Table VII, the corresponding step by step synthesis procedure is tabulated in Table VIII, and Prof. Franklin's
Data Limited solution to this problem is shown in Table IX.
Note that the channel inputs are now defined by the

Robinson: Computer-Limited Sampled-Data Simulation
TABLE X
DETERMINING THE VALUE AT t=O OF THE TIME FUNCTION
CORRESPONDING TO o*(Z) b~ factored

(11. 2)

(11. 3)

The third approach to the problem is illustrated in Fig. 7
and tabulated in Table XII. Referring to the figure, it can
be noted that the original two-sided Z transform is the
product of a numerator and of two terms in the denominator, one with poles inside and the other with poles outside the unit circle. Both denominator expressions can be
expanded to a finite number of terms rational in positive
and negative powers of Z respectively. These expansions
can be cross-multiplied to obtain a new two-sided Z transform. This transform can then be multiplied with the numerator, considering only terms that contribute to the value
at t = 0 of the corresponding time function. The example
in Fig. 7 shows only three numerator terms and therefore
only three terms of the new two-sided transform need be
considered. Table XII further demonstrates the process.
As a simple example of the theory that has been outlined, consider the problem of simulating a pure prediction

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

146

rc * (Z) .

C*(Z) = -:;- (; +b)(i+~-aTe-bT)
2a e-2b(a+k)(1 - e-aTe-bT )

_
e2

= 1 - - - - - - - - - - -bT-

T (a + b)2(1 + e-aTe- )
2a e-b(k+a)(l - e-aNTle-bNTI)

TABLE XII
DETERMINING THE (ApPROXIMATE) VALUE AT t=O OF THE
TIME FUNCTION CORRESPONDING TO cI>rc*(Z)-METHOD 3

C*(Z)
_
e2

1
-*-(f3l Z)
1
-f32*(Z)
cI>

(Z)

1 + OAZ-l + 0.13Z- 2 + 0.04z-a

+ 0.121Z-4 + . ..
H

f3l*(Z)f32*(Z)
+ 0.33445Z+ 2 + 0.72958Z+1
1.28180 + 0.49065Z- 1
+ 0.15739Z-2 + 0.04726Z- a + 0.0121Z- 4]
[-1.2Z+1
4.24 - O.8Z~q
(12.3)
CPrc(O) ~ 4.2624
(12.4)

in each of N simulation channels when the system inputs
are random messages with power density spectra
2b
0.03/TI • The one
term program is therefore superior under the majority
of potential operating conditions. The reason for the
success of the Computer Limited program is illustrated
in Fig. 8, where the mean squared error corresponding
to (13.2) and (14.2), a = b = 1/ NTI , N = 20, are plotted
as functions of T. Note that if T were not related to
computer program complexity the two term program
would always be superior to the one term program.
This is essentially the assumption of Data Limited
theory. The one term Computer Limited program superiority arises because the two term program actually
imposes the requirement for a system sampling period
of 2N,TI , while the one term program requires a system
sampling period of only NTI •

Robinson: Computer-Limited Sampled-Data Simulation

147
XV

TABLE

EO
I

DATA LIMITED SOLUTION TO PROBLEM EXAMPLE
DETERMINISTIC INPUT

(15.1)
(15.2)
a.=b=...!..

NT,

(15.3)
(15.4)

Fig. 8-Mean squared errors-one and two term programs
-random inputs.

Further physical insight into the reason for Computer
Limited program superiority can be gained by considering
the same problem when the system input is an exponential
ret) = e- bt and Co( = 0, N = 1. The Data Limited and
Computer Limited solutions to this problem are tabulated
in Tables XV and XVI respectively.
When b = 0, the Computer Limited solution is always
superior to the Data Limited solution. The reason for this
superiority can be seen from the time responses from the
two systems plotted in Fig. 9. Note that after a delay of
2Tl the two term program immediately achieves the best
response of which it is capable, and that while the one
term program response is delayed by only T 1 , it requires
a transient period to reach its optimum condition. The
one term program is superior because its shorter length
results in a higher permissible data rate, with a corresponding reduction in achievable integrated error squared. In
this particular example there is no penalty associated with
the poorer transient response of the one term program,
since the time response to a step input extends to infinity,
and it is therefore the steady-state response that is being
optimized. For this reason the one term program is superior for any finite a.
When b is not zero, so that the system output is exponentially damped, it is possible for the poorer transient
response of the one term program to overshadow its higher
data rate advantage, in which case the two term program
will be superior.
In general, the shorter program has the advantages of
a shorter delay before a change in the input is reflected in
the output and a higher data rate, and the disadvantage of
a poorer transient build-up to the optimum response condition.
The detailed procedures that have been described in this
report pertain to the synthesis of linear systems employing
linear digital computer programs, in which the minimizationof mean squared error or integrated error squared is
an effective optimization criterion. Itis important to realize
that many other classes of Computer Limited sampled data
systems exist, and that a great deal of further work is
required to extend the basic approach to the Computer
Limited problem presented in this report to these problem
areas. For example, the problems of simulating nonlinear

TABLE XVI
COMPUTER LIMITED SOLUTION TO PROBLEM EXAMPLE
DETERMINISTIC INPUT

C*(Z)

«:2

2ae-bk (1 - e-bTe- a7')

= ---------bT a7
'Y(a

+ b)(1 + e-

(16.1)

e- ')

= ~ _ __~~~~~=-~:?~aTL__

2b
(a
b)2(1 - e- 2bT )(1
2ae-bk (1 - e-bNTle-aNTl)

(16.2)

+ e-bTe-aT )

C*(Z) = - - - - - - - - - - 'Y(a
b) (1
e-bN7'le-aNTI)

IE2 =

+
+
2.- _ ___2~2bk(1 =-~~~=NT'l __
2b
(a + b)2(1 - e-'1hNTI) (1 + e-bNTle-aNTl)

2.0

1.8

1.4

SYSTEM
OUTPUT

I
I

1.0

t-rt

I
I

1.2

(16.4)

TWO TERM PROGRAM

rII \I

1.6

(16.3)

ONE TERM PROGRAM

0.8

......

0.6

""' ...

0.4

0.0

'--6

0.2

2T,

3T,

4T,

5T,

6T,

Fig. 9-Data Limited and Computer Limited response for
single time constant-output filter-deterministic input.

systems, of utilizing nonlinear computer programs and of
employing optimization criteria other than mean squared
error minimization present considerable challehge. Many
of the actual problems to be simulated by multiplexed
real time digital computers are actually nonlinear or can,
in specific instances, benefit from the utilization of nonlinear programs. Unfortunately, the sharpness of presently
available mathematical tools seems to limit the generality
of synthesis techniques applicable to these nonlinear problems.
The solutions presented in this report benefit from the
analytic flexibility inherent in the analysis of linear systems. It is submitted that, pending the development of
comprehensive nonlinear synthesis techniques, linear
theory can provide basic insights into the fundamental
problems associated with the synthesis of Computer Limited sampled data systems, and that this fundamental understanding can in turn serve as a guide in the synthesis
of systems considerably more complex than those covered
by the basic theory.

148

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE
Discussion

A. G. Favret (American Machine and
Foundry Co., Alexandria, Va.): Explain
the difference between a one and a two
term program to perform the same operation.
Dr. Robinson: Let us say that you are
attempting to use a digital computer to
simulate a certain dynamic response, such
as the response of an aircraft. It is possible
to attempt the simulation for a given channel using anyone of a host of different
programs. A one term program would generate an output equal to' the present value
of the input, multiplied by a constant. Two
different two term programs are possible.
One would generate an output equal to the
present value of the input multiplied by 'a
constant, plus the value of the input at the
prior sampling period, multiplied by a different constant. The other would generate
an output equal to the present value of the
input multiplied by a constant, plus the

val ue of the output at the prior sampling
period multiplied by a different constant.
In general, a (Pm
qm
1) term program
operates on (Pm
1) present and past
values of the input and qm past values of
the output.
A one term program does not require
storage of a computed variable and takes
a computing time T 1 • A two term program
requires the storage of one computed variable (corresponding either to the past input or the past output) and takes a greater
computing time, for example, 2T 1 • A three
term program requires the storage of two
computed variables and takes a still greater
computing time, and so forth.
S. H. Cameron (Armour Research):
In what sense is a program "optimum" in
a Data Limited system?
Dr. Robinson: In exactly the same
sense as in: a Computer Limited system.
That is, an optimum Data Limited system is
derived by minimizing the mean squared
error between actual and ideal system out-

+ +
+

puts. The difference between systems lies
in the factor that limits the system sampling
period.
In a Data Limited system, the system
sampling period is limited by the data
source, for example, the limited speed of
rotation of a radar antenna. If the derived
data rate is slower than the effective computer speed, long computer programs carry
no penalty.
In a Computer Limited system, the system sampling period is limited by the computer speed. A Computer Limited system is
characterized by the fact that the longer
you make your program, the longer your
system sampling period is going to be.

Any system in which a digital computer
operates on continuous data is a Computer
Limited sampled data system. A system in
which a digital computer operates on data
that is already sampled could be either
Data Limited or· Computer Limited, depending on the data rate and the computer
spe~d.

SAGE-A Data-Processing System
for Air Defense *
R. R. EVERETTt, C. A. ZRAKETt,

THE REQUIREMENT

FOR

SAGE

URING the past decade, the continental United
States has faced the continually increasing threat
of enemy air attack. High-speed, high-altitude
intercontinental bombers can deliver thermonuclear weapons to any part of our country. Even though ICBM capabilities are rapidly approaching operational status, it is
firmly exp'ected that the manned bomber threat will continue and grow well into the 1960 time period. Until very
recently, we have relied on an air-defense processing system whose traffic-handling techniques were almost identical with those used during World War II. Fortunately,
there has been substantial improvement in our inventory of
automated air-defense components. These include: improved radar systems, automatic fire-control devices,
automatic communication links for ground-to-ground or
ground-to-air communication, navigational systems, and
both missiles and manned aircraft whose performance
equals the threat of the newest manned bombers. But,
successful air defense requires both good components
and intelligent utilization of these components. A longrange supersonic interceptor is of little value unless enemy

* The research in this paper was supported jointly by the U. S.
Army, Navy, and Air Force under contract with the Massachusetts
Institute of Technology.
t Lincoln Lab., M.LT., Lexington, Mass.
System Development Corp., Santa Monica, Calif. Formerly at
Lincoln Lab., M.LT., Lexington, Mass.

AND

H. D. BENINGTON*

targets can be detected and tracked at long ranges. More
important, intelligent commitment of many such interceptors requires up-to-date knowledge of the complete
enemy threat and of the success of weapons already committed.
In early 1950, the military concluded that the manual
air-defense system in use at that time could not adequately
coordinate use of our improved hardware against the
growing enemy threat. The capacity of the system was
too low; the speed with which enemy aircraft could be
detected, tracked, and intercepted was too slow; and the
area over which an air battle could be closely coordinated
was too small. The problem was one of inadequate, nationwide data-handling capability: facilities for communication, filtering, storage, control, and display were inadequate. A system was required which would 1) maintain
a complete, up-to-date picture of the air and ground situations over wide areas of the country, 2) control modern
weapons rapidly and accurately, and 3) present filtered
pictures of the air and weapons situations to the Air Force
personnel who conduct the air battle.
The Semiautomatic Ground Environment systemSAGE-was developed to satisfy these requirements.
SAGE uses very large digital computing systems to process nation wide air-defense data. SAGE is a real-time
control system, a real-time communication system, and a
real-time management information system. The basic ideas

Everett, Zraket, and Benington: SAGE-A Data-Processing System for Air Defense

Fig. I-A SAGE direction center building contains power generation and computing equipment, operational areas for directing
sector operation, and office and maintenance facilities. Data are
transmitted to this center both automatically and by voice phone.
The center communicates with adjacent SAGE centers and
transmits guidance data to weapons under its control.

of this system resulted from the efforts of Drs. George E.
Valley and Jay W. Forrester of M.LT.
A large number of organizations have contributed to the
development of SAGE since its conception in the Air
Force and at M.LT.'s Lincoln Laboratory. The International Business Machine Corporation (IBM) designs,
manufactures, and installs the AN/FSQ-7 combat direction central and the AN /FSQ-8 combat control central
including the necessary special tools and test equipment.
The Western Electric Cumpany, Inc., provides management services and the design arid construction of the direction center and combat center buildings. These services
are performed with the· assistance of the subcontractor,
the Bell Telephone Laboratories. The Burroughs Corporation manufacturers, installs, and provides logistic support
for AN/FST -2 coordinate-data transmitting sets. The
System Development Corporation (until recently a division of the RAND Corporation) assists Lincoln Laboratory in the preparation of the master computer program
and the adaptation of this program to production combat
and direction centers. At the present time, SAGE is in production; a prototype unit has been successfully operated for
some time.
SECTORS AND DIRECTION CENTERS
With SAGE, air defense is conducted from about thirty
direction centers located throughout the United States
(Fig. 1). A center is responsible for air surveillance and
weapons employment over an area called a sector. Each
center contains a digital computing system-the AN /
FSQ-7-containing almost 60,000 vacuum tubes. Over one
hundred Air Force officers and airmen within the center
control air defense of the sector. Most of these men sit at
consoles directly connected to the computer where they receive filtered displays of the computer's storage of system
status data; they direct the computer through manual keyboards at each console. The ~oston Sector is typical; its
direction center is located at Stewart Air Force Base in
New York. Its area of responsibility extends from Maine

149

on the north to Connecticut on the south; from New York
on the west to a point hundreds of miles off the sea coast
on the east.
The computer in the direction center can store over one
million bits of information representing weapons and surveillance status of the sector at one time (Fig. 2). These
bits represent thousands of different types of information.
For example, the computer generates and stores positions
and velocities of all aircraft, or it stores wind velocity at
various locations and altitudes. Within the computer, a
program of 75,000 instructions controls all automatic operations; input data are processed, aircraft are tracked,
weapons are guided, outputs are generated. Each second,
the computer can generate over 100,000 bits of digital jnformation for display to Air Force operator consoles. Each
operator receives cathode-ray tube displays which are
tailored to his needs, and he may request additional information or send instructions to the computer by means of
keyboard inputs on his console. Each second, the computer
can generate thousands of bits of information for automatic digital transmission via telephone or teletype to
weapons and missiles, to adjacent centers or higher headquarters, and to other installations within the sector.
SAGE DATA-PROCESSING

OVER
ONE HUNDREO
AIR FORCE

OPERATORS

SAGE DIRECTION CENTER

Fig. 2-The direction center continuously receives input data from
hundreds of locations within and without the sector. Some
of these data are transmitted digitally over telephone lines and
read directly into the computer; some are transmitted by teletype or voice phone and transcribed onto punch cards before
input to the computer. During one second, over 10,000 bits of
data representing hundreds of different types of information can
be received at the direction center.

How fast is this system? Obviously, response times from
input-to-output vary with the task performed. Fastest response is required by automatic control functions (such
as weapons guidance) and for man-machine communication (such as displays of requested information) . For
many of these functions, only several seconds are required
from stimulus to response. For others, several minutes may
elapse before the effects of new data are reflected throughout the system. We shall consider now, in somewhat more
detail, the three major areas which comprise SAGE data
processing. First, the sector or environment which contains the data sources or sinks coordinated by the direction
center. Next, the man-machine component: how the operators within the direction center are informed of the air
situation and how they affect its progress. Finally, we
shall describe the computing system which performs the
automatic component of the direction center function.
THE SAGE SECTOR
The direction center communicates with over one hun·
dred adjacent installations (Fig. 3). Air surveillance data
are received from several types of radars: long-range

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

150

HIGHER

•\

""---

~~j:;;~" h'rroQ'~. ~[~~S
:~'0_~T~"I
~P
T
\ L
(~A
~ ~\~ JJJ01
0

1- 1;
-

GAP fiLLER RADAR

~tl-,
SENSE UNITS
READ
DETECTED BY
EXECUTION OF
OR
"BRANCH AND SENSE" WRITE

READ
ONLY

INSTRUCTION

Fig. 2-Intercommunication facilities.

operation. The duration of the transient would depend
upon the complexity of the air situation at the time of
switchover. This transient period is minimized by maintaining certain key data upon the drum fields of both the
active and stand-by computers. Thus, an up-to-date summary of the air situation is available to the direction-center
program immediately after switchover.
The summary air-defense data stored in the stand-by
computer are periodically assembled by the direction-center program operating in the active computer, and transmitted to storage in the stand-by computer via the intercommunication-drum system. The amount of data transferred is limited by the program operating time available
in the active computer, and the drum storage available in
the stand-by computer. Operating time is a critical factor
because the direction-center program is part of a real:time control system, and any increase in operating time
degrades over-all system response. The drum storage
available in the stand-by computer is limited by the storage
requirements of the stand-by programs.
The nature of the summary air-defense data can best be
discussed in terms of the types of data tables used and
generated by the direction-center program. There are four
broad categories of data tables: input, output, display, and
central.

162

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Input tables contain data awaiting processing by the
direction-center program. These data are generated at external sources (e.g., radar data), and by the Air Force personnel within the direction center. Output tables and display
tables contain data awaiting sequential transmission to locations outside the direction center (e.g., air bases), or
display to direction-center operators. No portions of the
input, output, or display tables are transmitted to the
standby computer for storage as summary air-defense
data. The net result is loss of direction-center operation
during the switchover period. If this period is short, the
effect is not serious, because the input data are accumulated again after switchover, and the display and output
tables are regenerated by the direction-center program itself. The central tables, on the other hand, are the heart of
the air-defense program. In a broad sense they represent a
mathematical model of the air situation on which the operation of the direction center is based. It is the central
tables, or more specifically, the key portions of the central
tables that are transferred to the stand-by computer as
summary data.
The only other duplex function of the active computer
is that of monitoring the intercommunication lines to determine if a scheduled switchover is to take place. The
program operating time required to accomplish this monitoring function is negligible. Th~ switchover process itself
will be discussed later.
DUPLEX FUNCTIONS OF THE STANDBY COMPUTER

The stand-by computer must operate maintenance programs, and at the same time be readily available for operation of the direction-center program. Certain duplex functions, then, must be performed by the stand-by computer:
1) Monitor active-computer alarms,
2) Maintain the direction-center program on the standby computer drum fields,
3) Transfer and store summary air-defense data assembled by the active computer,
4) Monitor operator-inserted switch requests controlling standby operations,
5) Prepare digital displays indicative of the status of
standby computer operation.
Only the first of these functions (alarm monitoring) is an
equipment function; the others are programming functions.
Memory parity, drum parity, or arithmetic overflow
alarms that occur in the active computer cause an automatic branch of program control to test memory in the
stand-by computer. The sequence of instructions in test
memory initiate preparation of the standby computer for
switchover. Preparation for switchover includes erasure
of all maintenance programs and tables from core and
drum storage, restoration of the direction-center program
upon the stand-by drum fields (if necessary), and a final

transfer of the summary data from the active machine.
Having completed preparations for switchover, the
stand-by computer simply waits for switchover to occur,
or for a manual intervention to restore normal stand-by
status.
Maintenance of the direction-center program on the
standby computer's drum fields permits rapid recovery of
direction-center operation after switch over. The fact that
the direction-center program is properly stored is verified
by reading each program drum field into core memory,
computing the sum of the binary numbers stored on the
drum field, and comparing the result with the correct sum
(also stored on the drum field). If the computed sum is
incorrect, the offending drum field is reloaded from magnetic tape. The process of checking and loading the program drum fields occurs automatically whenever a maintenance program has destroyed the contents of a program
field, or whenever preparation for switchover is initiated
by an active computer alarm. It can also be requested by a
manual switch action.
The duplex functions of transferring summary air-defense data, monitoring operator switch actions, and the
preparation of digital displays are executed periodically.
The frequency with which these functions are performed
depends upon the mode of operation of the stand-by computer. One of three modes may be selected. Each provides
a different frequency of execution of the periodic duplex
functions, in the range of once every few seconds to once
every few minutes. This requirement for interleaving simplex and duplex operations imposes stringent requirements
for manual and automatic control of the maintenance programs. Control of the sequence of operation of maintenance programs, and selection of the mode of stand-by
operation is accomplished by manually-inserted operatorswitch actions. The running time of each maintenance program is known to the stand-by control program, and either
manual or automatic selection of long running maintenance
programs is automatically prevented if the selected mode
of stand-by operation requires frequent execution of the
periodic duplex functions. In order to relieve this runningtime restriction on the selection of maintenance programs,
the programs are designed as a collection of program units
to permit operation of long running maintenance programs
by operating them one program unit at a time.
SWITCHOVER

Switchover requires transfer of direction-center inputs
and outputs from the active to the stand-by computer, and
activation of the direction-center program in the stand-by
computer. Preparation of the stand-by machine for switchover is initiated automatically by an active computer alarm,
or manually by an operator switch action. After the
stand-by computer has completed its preparations for
switchover, and after the duplex switch has been operated,
control of the stand-by computer is transferred from the

Vance., Dooley, and Diss: Operation of the SAGE Duplex Computers

163

5) In the case of scheduled switchover, the air-defense
stand-by control program to the startover program. The
information transferred from the active computer is
startover program performs the function of activating the
stored on the proper table drum fields.
direction-center program in the standby computer, and
thereby completes the transition of that machine from
Completion of the switchover process requires that the
standby to active operation.
startover program process the air-defense data stored in
Two modes of switchover have been provided. The
the stand-by computer to make it usable by the directionemergency switchover mode is used when switchover
center program. The startover program then transfers the
occurs after the active computer has become inoperative.
control portion of the direction-center program into the
The scheduled switchover mode is used when both macore memory of the stand-by computer, and transfers comchines are in operating condition at the time of switchover.
puter control to the direction-center program.
The major difference between these two modes is in the
Sorting and extrapolation of air-defense data are peramount of air-defense data that is made available to the
formed by the startover program. In the case of emerstand-by computer. Normally, only summary air-defense
gency switchover, only summary data are available to the
data are available upon the drum fields of the stand-by
stand-by computer. These data, which were gathered
computer. As was mentioned before, these data are transfrom several central tables, occupying different drum fields
ferred to the stand-by computer periodically, and the
in the active computer, are packed together upon one drum
amount of data that can be transferred is limited by the
field in the stand-by computer. The startover program sorts
computing time and storage-space restrictions imposed
these data and distributes them among the appropriate
upon a periodic operation. In short, program operating
stand-by table drum fields. In the case of scheduled switchtime is not available to transfer a voluminous amount of
over, the air-defense data are already stored upon the
data during each cycle of the direction-center program. If,
proper standby drum fields, and the sorting process is unhowever, switchover is scheduled when both computers
necessary. The extrapolation process performed by the
are operating, more complete data can be transferred as a
startover program adjusts the air-defense data to com"one-shot" process during switchover. This wholesale
pensate for the program operating time lost during switchtransfer of data is accomplished by interrupting operaover. The process is primarily a matter of extrapolating
tion of the direction-center program just prior to switchthe position of aircraft tracks along their last known velocover, and transferring the contents of the central tables
ity vectors.
and display tables from the drum fields of the active comFUTURE DEVELOPMENTS
puter to the corresponding drum fields of the stand-by
The duplex problem is that of determining how best to
computer. Proper timing of successive drum transfers is
utilize
two computers to enhance direction-center reliabilachieved by signals transmitted between computers via the
ity.
A
secondary
consideration is that of determining how
intercommunication lines.
to
make
efficient
use of the stand-by computer without
The following conditions describe the status of the
primary
requirement that it be readily
jeopardizing
the
stand-by computer at the time that control is transferred to
available
to
perform
air
defense.
When more information
the startover program.
has been gathered regarding maintenance requirements,
1) The direction-center program is properly stored and when maintenance techniques have been perfected, it
upon the drum fields of the stand-by computer.
may be possible to utilize the stand-by computer for a
2) The summary air-defense data are stored upon one limited amount of data processing, or for simulation of
stand-by drum field.
battle conditions during training exercises. Such applica3) All traces of stand-by program operation have been tions must, of course, be designed within the ground rules
erased from core and drum storage.
established by the primary requirements of adequate
4) In the case of emergency switchover, all program stand-by computer maintenance, and availability for rapid
tables are cleared.
swi tchover.

164

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

A Digital System for Position Determination
DAN C. ROSSt

NE of the most impo.rtant functions in the air traffic
control (A TC) system is that of aircraft position
reporting. A large fraction of the equipment and
effort required in CAA operations is involved in the initiation and handling of position reports. A pilot must divert
attention from the actual control of the aircraft to talk
with ground controllers for the purpose of reporting position, or for the purpose of procedural communications
associated with position reports. Many of the adjustments
and readings of navigation instruments are performed not
because the pilot wishes such frequent information for his
own uses but because of the necessity of reporting position. A large amount of electromagnetic spectrum is consumed at present in the position reporting function, and
the need for spectrum will increase still further as air
traffic control is expanded, unless some new reporting technique is developed.
Future systems of air traffic control will utilize automatic data processing machinery to handle many of the
routine functions and will permit the human controller to
handle with safety more traffic than he can today. The
effectiveness of the combination of human controllers and
automatic computers will be greatly improved with the
advent of an automatic means of providing frequent and
accurate position reports on all aircraft in the system. Unfortunately, there is no existing system of position reporting which meets the requirements of air traffic control.
A great deal of effort has gone into the development of
several methods of providing aircraft position information to the traffic control system. The three techniques
which have received the most attention are radar, beacon
transponder, and data link. The principal advantages of
radar is that no equipment is required in the aircraft, but
this advantage is offset by the lack of identification, the
lack of correlated altitude data, and excessive noise and
interference of various sorts. The conversion of the raw
signal from the radar into a form suitable for air traffic
control requires the continuous solution of a difficult correlation problem involving either a great deal of computing capacity or the full attention of many human operators. Both the beacon and the data link overcome the
fundamental problems of radar at the expense of adding
equipment to the aircraft and of introducing a number of
difficult technical problems which are yet to be solved. At
least in the case of the data link, it appears that the various
technical difficulties will eventually be surmounted, but the
expense in terms of airborne equipment costs and total
usage of the electromagnetic spectrum may be quite high.
t IBM Corp., Kingston, N.Y.

The purpose of this paper is to present the basic principles of operation of a position reporting technique which
satisfies the present and future requirements of air traffic
control and overcomes the known technical difficulties in
the radar, beacon, and data link systems, and yet promises
to accomplish these goals with a minimum of expense. The
technique to be discussed is known as Automatic Position
T elemetering (APT). The APT system has progressed to
date through preliminary design and testing phases which
have concentrated on the radio communication portions of
the system. The results of this work indicate the feasibility
of the proposed system and point the direction for the design of a complete prototype.
The basic requirements for an automatic position reporting system are obtained from a consideration of the
present manual techniques, the characteristics of semiautomatic data processing systems, and the operational and
technical shortcomings of radar, beacon, and data link.
First, the position reporting system must be capable of
integration with automatic data processing machines and
must eliminate or minimize the manual operations required
of pilots and controllers. Second, the total electromagnetic
spectrum assigned to the system must be held to a minimum; system planners ought to regard bandwith as one
of our most precious national assets. Third, the airborne
element of the system must be kept as small and inexpensive as possible because of compounding effects on the
weight, reliability, maintainability, and cost of the total airframe system. Fourth, the system design must be based on
fundamental logical and physical principles selected to minimize the technical difficulties which have been experienced
in recent beacon and data link development programs.
A major design objective is the provision of service to
all aircraft in the system on a single radio channel. Accomplishment of this objective completely eliminates tuning operations as far as either the controller or pilot is
concerned. The complexity of both the airborne and
ground-based portions of the system is greatly reduced if
single-channel operation can be realized. A related secondary objective is the minimization of the bandwidth required for the single channel.
If a single channel is to suffice, it is mandatory that
some sort of time-division technique be utilized. This
leads naturally to completely synchronous operation for
allocating use of the channel and to discrete time-slot addresses uniquely assigned to each aircraft. Use of the
time-division addressing principle eliminates the need for
narrow-beam antennas and complex antenna-guidance
equipment.

165

Ross: A Digital S.ystem for Position Determination
Another design objective of major importance is that
the amount of information transmitted from the aircraft
to the ground environment must be minimized. If possible,
the transmission ought to be limited to a single impulse for
the sake of simplicity. Any information which could just
as well be determined at the receiver, even if this amounts
to a redetermination of data previously known at the
transmitter, ought to be so determined rather than wasting
channel capacity in its transmission. If the ideal of position reporting by means of a single impulse can be realized, then many of the technical problems involved in
pulse transmission systems are greatly simplified, in particular, the problem of interference from multipath echo
phenomena.
Noise phenomena such as atmospherics, receiver noise,
and ignition noise will plague any sort of radio communication system. In order to minimize noise effects, the
transmitter must produce high pulse power and the receiver must be designed for low-noise performance and
located in a relatively quiet environment. The carrier frequency ought to be selected in the 10DO-mc region to obtain
an optimum balance between atmospheric noise at lower
frequencies and receiver noise at higher frequencies. For
the position reporting application, the line-of-sight limitation of UHF communication is actually an advantage
because several aircraft can be given the same address provided only that the aircraft are separated by several hundred miles.
All of the requirements and design objectives introduced in the toregoing discussion can be realized in the
proposed APT system. In addition, the system is capable
of handling traffic densities considerably greater than the
densities predicted for 1975. There would be no difficulty'
in providing APT service to several thousand aircraft
simultaneously within the area covered by an air route
traffic control center.
The principle of operation of the APT system along
with the interrelationship between the major equipments
involved is shown in Fig. 1. In each major terminal area,
four ground-based receiving equipments are arbitrarily
located, provided only that the area of maximum traffic
density is central with respect to the four receiver sites.
Maximum separation of any two receivers in the group
would be about 20 miles. Each aircraft is provided with
a transmitter which emits an intense pulse of UHF energy
every few seconds. Interference between aircraft in the
system is prevented by means of time-division multiplexing techniques.
As shown in the plan view of Fig. 1, the pulse of UHF
energy leaves the airborne transmitter at instant tp and
travels outward at the speed of light. At some later instant
of time tA the wave front reaches receiver A. The detected pulse at A is used to sample the contents of a freerunning clock, recording tA in digital form. Immediately
thereafter, the binary representation of tA is serially transmitted on a digital data line to a centrally located Coordinate

A
COORDINATE
CONVERSION
COMPUTER

\
\
\

\
\

PLAN
VIEW

PA
PB

\
'-----f:

= tA

= ts

t p

= tD

"'
x, ~,e-\
I

P
ABCD

AIRBORNE PULSE TRANSMITTER
FIXED RECEIVING SITES

UNKNOWN = (X,

11,;0-, Up

TO
ATC
COMPUTER

UHF RADIO

GIG DIGITAL DATA SERVICE

I I I[[
P

B A C D

Fig. I-System diagram.

TO DIGITAL
DATA SERVICE

Fig. 2-TOA equipment.

Conversion Computer (CCC). In similar fashion, the times
of arrival of the signal at receivers B, C, and D are obtained
and transmitted to the CCC. The equipment at each of the
four receiving sites will be referred to as Time of Arrival
(TOA) equipments.
Each of the four unknowns can be determined in theory
from the four TOA measurements. One way of stating
the functional relationships between these quantities is
shown in Fig. 1. However, there is no interest in tp and
precise values of aircraft altitude cannot be obtained from
the TOA data. Thus a practical design for the CCC would
provide for the calculation of x and y only, and the altitude z must be found by another method. The output data
from the CCC is transmitted in serial form over a digital
data line to locally or remotely located A TC data processing
equipment.
A block diagram of the major sections of each TOA
equipment is shown in Fig. 2. The oscillator, shaper, and
counter constitute a high-speed digital clock capable of resolving intervals of a fraction of a microsecond. The precision of this counter is directly related to the precision of
the final position measurement (1 lLsec = 0.186 mile). The
received pulse is used to sample the high-speed counter
and transfer its contents to a shift register. The received
pulse also initiates a series of shift pulses to transmit the

166

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

TOA information to the CCC. The number of stages, N,
in the high-speed counter must be sufficiently large to insure
that no more than one end carry occurs during the passage
of a wave front across the net of four TOA stations.
To make the UHF transmission highly reliable in spite of
various noise phenomena, it may be necessary to send
doublets or triplets rather than single pulses. The purpose
of the "pulse-shape filter" shown in Fig. 2 is to produce
an output pulse when and only when the input signal
is within appropriate tolerance limits of the coded waveform transmitted by the airborne unit.
The airborne pulse transmitter for the APT system is
shown in Fig. 3. The transmitter proper consists of a highpower modulator which excites a cavity-tuned power oscillator. The approach which appears most practical employs
a hydrogen thratron discharging a delay line to modulate a
"lighthouse" triode. Pulse power levels of several kilowatts are desired. Satisfactory results were obtained during the tests with a pulse power of 1.5 kw and there appears to be no difficulty in designing economical modulators and oscillators that will produce up to 10 kw.
Positive identification is provided and intrasystem interference prevented by means of a synchronized timeslot counter in each aircraft. The counter in any given aircraft is set to the aircraft address each time that the
"framing pulse" is received from the ground. The framing pulse is sent to all aircraft simultaneously at the rate
of about once a minute. The address counter in each aircraft
counts down toward zero under the control of a mediumprecision oscillator in the airborne equipment. When the
address counter reaches zero and produces an end-carry
pulse, the modulator is triggered. Since each aircraft
would be given a different address, the UHF impUlses all
occur at different times.
The operation of the proposed address timing technique
can be clarified by considering a typical example. In Fig.
3 and Fig. 4, the following numerical parameters are assumed: framing-pulse period = 1 minute, number of addresses = 212 = 4096. It is further assumed that the aircraft address is established by means of four octal
switches. The aircraft chosen for the purposes of this
example has address 13 (octal). The framing pulse transfers the number 13 (octal) from the address switches to
the counter, and once each 15- msec the counter contents
are reduced by one. As the address counter changes from
0000 to 7777, the end-carry pulse is produced and the UHF
transmission occurs.
The framing pulse is transmitted from ground to air by
multiplexing it on the VHF and UHF voice channels to
economize on both equipment and bandwidth. Since the addressing technique proposed does not require precise timing, the framing pulse can be handled on an audio channel. Assuming that the voice signal could be cut off at
about 4 kc without serious loss of fidelity, it seems practical to use a sub carrier of about 6 to 8 kc for the framing

r------<:
~

@(0)\i)

____

ANTENNA

L- _ _ _

~----'

@@@@

ADDRESS SWITCHES

Fig. 3-Airborne equipment.

1-011-1---------

PULSE
FRAMIN:Jl.
------------------~

4096
CPM
OSC.

,,
ADORESS
COUNTER

K -0013
0 0 o 0
0 0 o 0

e.g.

0 o 0 o
0 o 0 o
I I I I I I o 0 o 0 000 o
5 4 3 2 I 076 5 4 3 2 I o

o
o

0
0

o
o

0
0

L
J1lL

I MIN - - - - - - -......

7
7
7
7

7
7
7
6

000 0
000 0
I

~ ~

4 3 2 I

UHF
TRANSMISSION
'r----------'

'-----~

Fig. 4-Timing.

pulse. In order to reject extraneous noise pulses, it is desirable to employ two or three short bursts of the sub carrier for each framing signal. The purpose of the pulseshape filter shown in Fig. 3 is to decode the waveform
used for the framing signal in order to distinguish it from
various interfering signals. The framing-pulse transmitters
are synchronized on a national basis via transmissions over
1£ channels from a central timing standard. Frequencycontrol servomechanisms are provided to keep the framing
pulses together during outages or noisy periods on the LF
channel.
It should be noted that the APT system provides both
position and identification data to the ground environment without actually transmitting either one of these
quantities. Both the position and the identification are determined at the final receiver on the ground the position
by a precision time-difference technique and the identification by medium-precision time division. To add automatic
reporting of altitude, the altitude signal is mUltiplexed on
the air-to-ground transmission. Altitude data can be readily
included in the APT system by providing a second UHF
pulse such that the spacing between the two pulses represents the altitude code. One way to produce the second
pulse is to use the end-carry pulse from the address counter

167

Ross: A Digital System for Position Determinat.ion
to drive a delay multivibrator which produces two pulses
to trigger the modulator. The delay between the two pulses
would be controlled by an electromechanical connection
from a sealed aneroid altimeter in the aircraft. In the TOA
equipment, the first received pulse would be used to start
an altitude counter and the second pulse would transfer
the contents of this counter to an extension of the shift
register shown in Fig. 2. The second pulse would also be
used to reset the altitude counter. Thus, the altitude information turns out to be the only data actually transmitted from the aircraft to the ground in the usual sense
of the term "transmission," while the position and identification are both ground-derived.
The high-speed counter used in the TOA equipment is
not synchronized with any of the other counters in the
TOA net. The effect of synchronization is accomplished
much less expensively by the provision of a set of equipment identical to that used in the aircraft but located on
the ground near the center of the net. TOA signals arriving at the CCC during the time slot assigned to the groundbased pulse transmitter are then compared with the values
which would have been obtained if the four counters had
actually been synchronized. The differences so obtained
are then held in temporary storage in the CCC and are
used to correct the TOA values received during each of
the other time slots in the frame. The additional calculations required in the CCC amount to a few subtractions
and represent no important increase in the amount of electronic equipment required ..
The principal computing problem which the CCC must
solve is the conversion of the hyperbolic coordinates represented by the TOA values to a more convenient set of
coordinates for use by the controllers and data processing
machines involved in A TC operations. The output coordinates from the CCC may be chosen to be rectangular or
geodetic with very little difference as far as the cost of the
CCC is concerned. Several possible designs of the CCC
have been considered; one based on table look up and interpolation, a second method based on an iterative technique,
and a third and most promising method based on direct
calculation. The time provided for each computation cycle
is sufficiently long that the CCC requirement can be met
with a simple design based on a serial arithmetic unit controlled by a fixed program. The special geometry associated with each TOA net can be accommodated by wellknown storage techniques. Since no human intervention is
required in the normal operation of the CCC, the inputoutput requirements are easily satisfied.
The numerical values of the APT parameters may be
chosen from a fairly wide range of values; in many cases
the range of practical choice extends over several. orders
of magnitude. The numerical values of the parameters
used in Table I are presented in the interest of clarity
and do not .necessarily represent recommended values.
Some of the more flexible parameters are: the number of

TABLE I
A

POSSIBLE SET OF PARAMETERS

Airborne Pulse Transmitter
Reporting rate:
Number of addresses:
Number of time slots:

Time-slot duration
Carrier frequency
Pulse power
Pulse duration
Timing for altitude pulse:

Enrou te phase
Terminal phase
Enroute phase
Terminal phase
Enroute phase
Terminal phase
Total

Minimum delay
Maximum delay

8
2048
256
1 X2048 = 2048
8X 256=2048

60/4096=

4096
14.65
1400
10
0.2
1000
2600

msg/min
msg/min
a/c
a/e
slots
slots
slots
msee
me
kw
p.See
p'sec
p'sec

TOA Equipment
Receiver bandwidth
Antenna height
Clock-oscillator frequency
Number of stages in clock
Clock cycle duration
Maximum span of TOA net
Message content:

Message rate
Shift rate
Altitude-oscillator frequency
Number of altitudes

Time of arrival
Altitude
Parity check
Spares
Total

5
me
100
feet
5
me
9
bits
29XO.2= 102.4 p'see
102.4XO.186= 23.1 miles
9
bits
6
bits
1
bit
3
bits
19
4096
19X4096/60=1297
0.04
1600XO.04= 64

bits
msg/min
bit/see
me
levels

APT System
Minimum altitude coverage:
Maximum range
Precision:

Terminal phase
Enrou te phase
Inside TOA net
At maximum range
azimuthal
radial
At SO-mile range
azimuthal
radial

o feet
4000
feet
90
miles
0.05 mile
0.5
1.5

mile
mile

0.3
0.5

mile
mile

time slots, the time-slot duration, and the reporting rate.
These parameters are limited only by the requirement that
the interval between reports from the same aircraft must
equal the product of the number and duration of the
time slots. This requirement is modified if a higher reporting rate is desired in the terminal phase of aircraft
flight than in the enroute ph'ase. For example, the 4096
time slots discussed earlier could be divided into two
groups of 2048 slots each. One of these groups could be
used to handle 2048 aircraft in the enroute phase of flight,
and the remainder could be divided into 256 groups of
8 slots each to accommodate an additional 256 aircraft in
the terminal phase with a reporting rate 8 times that used
in the enroute phase. By properly choosing the numerical
values involved, the pilot's attention required in the setting
of the APT address can be limited to a single setting of
the address switches just before take-off, plus the operation
of a two-position switch at the beginning of the en route
phase of flight and once again at the beginning of the
terminal phase.
The precision of the APT system in the near zone depends only on the TOA clock resolution. At maximum
range, the precision depends on the ratio of the TOA station spacing to the clock resolution. For a value of clock
resolution of 0.2 [J.sec and a TOA station spacing of 20

168

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

miles, the precision will vary from about 0.05 mile inside the
net to about 1.5 miles at a range of 90 miles. At long range,
the azimuthal precision is considerably better than the
radial precision.
The automation of the position reporting function will
eliminate a large fraction of the present communications
load. Greater precision of position data will reduce the
frequency of conflicts with a corresponding reduction in
the number of transmissions required to each aircraft. A
standard clearance signal requiring no human intervention
can be employed except in the small number of cases requiring special transmissions. It therefore appears doubtful that an automatic ground-to-air data service can be
justified for more than a small fraction of the aircraft in
the future A TC system. These reasons explain the emphasis
of this paper on the air-to-ground reporting service and
the neglect of the reverse direction of transmission.
The proposed APT system meets the basic requirements
of position reporting for air traffic control. In the interest
of economy, the special needs of the more advanced air-

Discussion
Mr. Bhippel (U. S. Signal Corps):
Wouldn't three ground stations be sufficient
for position determination?
Mr. Ross: Assuming that accuracy of
the order given in the paper is required,
then one must acknowledge that the final
results depend on variation in three dimensions and, therefore, three independent
time differences are required. The number
of independent time differences is one less
than the number of TOA receivers.
D. C. Friedman (National Bureau of
Standards, Washington, D.C.): What provision would be made for control when
there is a plane transmitter outage, possibly
unknown to the pilot? Or a plane with no
transmi tter ?
Mr. Ross: In any future air traffic control system, two-way radio would continue
to be employed; thus any outage of the
APT transmitter could be overridden by
reverting to voice transmission of estimated
times of arrival over various fixes as determined by airborne navigation equipment.
Aircraft not outfittc'd with APT would, of
course, be required to file position reports
by voice radio at all times while flying
under instrument conditions.
Mr. Friedman: How many computers
would be required for current airways?
How many ground stations? What would
be the cost for the ground and aircraft
installations?

craft ought to be met by providing additional equipment to
supplement that used for the basic functions. All aircraft
in the system need not carry high-precision navigation
equipment plus automatic two-way digital communication
merely because a minority of the aircraft requires these
devices. One of the advantages of APT is its compatibility
with all sorts of navigation techniques including contact
flying, VOR and other air-derived navigation systems,
deadreckoners corrected manually or automatically from
ground-derived data, and advanced inertial systems.
There are many design problems in the APT system
which remain to be attacked, so it is too early to predict
success. However, the investigations and tests to date indicate that the system is feasible and has several strong
points. Some of the more important advantages are: simplicity of airborne equipment, flexibility of parameter
choice, minimization of pilot attention, and perhaps most
important of all is the independence of position-measuring
accuracy with respect to malfunctions or misadjustments
in the airborne unit.

Mr. Ross: One coordinate conversion
computer and four TOA receivers are required for each major airport. Throughout
most of the United States, the enroute area
would be adequately covered by the installations at the terminals. In the areas where
the terminal installations do not provide
sufficient enroute coverage, one has the
alternative of installing supplementary
APT nets or requiring the use of voice
radio for position reports while flying
through these areas. It is too early to state
cost estimates for either ground or the
aircraft installations.
T. Kampe
(Librascope,
Glendale,
Calif.): How many ground stations are
envisioned across the country?
Mr. Ross: The number of APT installations depends entirely on the number of
terminal areas requiring automation of
the position-reporting function. The determination of the traffic level needed to
justify such service would have to be made
by the Civil Aeronautics Administration
in the case of civil airports and the cognizant military service in the case of military
airports. Further technical development and
product engineering work should be carried
out before these questions of system economics can be answered intelligently.
Mr. Kampe: How are aircraft with
malfunctioning sets to be detected, and how
handled?
Mr. Ross: It is important to note that
malfunction of the APT transmitter does
not produce erroneous position measure-

ments. However, there are two important
malfunctions to consider. First, a complete absence of the UHF pulse would be
detected at the A TC data processing center in the form of a missing position report in some particular time slot. If this
situation continued, the pilot of the offending aircraft would be notified by voice
radio to revert to voice position reporting
over certain fixes. An outage of the framing pulse would result in a very slow drift
in the aircraft identification number-a
difficulty which is easily resolved either by
automatic or manual "identification tracking" at the air traffic control center.
Mr. Kampe: How are transmissions
between different sets of ground stations,
for a specific aircraft, to be integrated into
a coordinated picture of aircraft?
Mr. Ross: The output of the coordinate
conversion computer associated with each
APT ground installation would be transmitted automatically over ground-to-ground
communication facilities to the terminal air
traffic control facility and also to the air
route traffic control center covering en route
operations in the area. Thus, each air traffic
control facility receives position data on all
aircraft within its area of responsibility.
Digital data transmission techniques presently available are entirely satisfactory for
the APT application. The necessary data
processing and display equipment at the air
traffic control centers would be designed to
include data from APT along with data
from other sources.

PROCEEDINGS OF THE EASTERN COMPUTER CO.IVFERENCE

169

Real-Time Data Processing for CAA
Air-Traffic Control
G. E. FENIl\10REt

OUR years ag~, the .Eastern ~ oint Computer Conference was held 1ll thIS same CIty. At the first session
of that Conference, Vernon Weihe, representing the
Air Transport Association of America, stated that "the
need for automatic computation and automatic data handling (in air-traffic control) is immediate and urgent." He
challenged the computer industry and the aviation industry
to meet this need with sound system design incorporating
human engineering and the rapidly advancing technical
developments of the day. The paper took note of the fact
that a start had already been made with the installation of
a magnetic drum-message storage and processing system
at the CAA Technical Development Center in Indianapolis, Ind. This present paper is somewhat in the nature of a
status report, describing how the Civil Aeronautics Administration is beginning to use electronic computers for
air-traffic control operations. In order to understand this
application, it will be necessary to consider briefly the manual operations which are to be replaced.
Air-traffic control is exercised in two types of areas. The
first type, called the terminal area, is that airspace in the
proximity of an airport where aircraft are under the jurisdiction of an approach controller or tower controller, located at the airport itself. The second type, which is called
the enroute area, is that airspace designated as Federal
Airways, which are the well-traveled highways of the sky.
In the enroute area, control is exercised from an air-route
traffic-control center, of which there are 27 within the
continental limits of the United States. A typical center
has jurisdiction over an area approximately 300 miles
across. Plans are well along to expand enroute .control
area to include all airspace above a certain designated altitude, such as 24,000 feet. This paper will concern itself
mainly with the operations of the en route area.
Fig. 1 shows an air-route traffic-control center. The individual controllers are responsible for a portion of the
area called a sector. Each sector has a tabular display in the
form of a board in which are inserted flight-progress
strips. A close-up view of one of these boards is shown in
Fig. 2. Within the geographical area of the sector, there
are several key traffic-control points generally located at
the intersection of airways which are called fixes. Aircraft
are required to report to the ground by radio whenever
they pass over one of these fixes, in order that the controller may ascertain their position and maintain proper
separation from other aircraft both in altitude and in time.

t Chief, .{\ir Traffic Control Equipment Branch, CAA Tech Dev.
Center, Indianapolis, Ind.

Fig. I-Indianapolis air-route traffic-control center.

Fig. 2-Flight-progress board.

Aircraft pilots who intend to make a flight under instrument conditions or under the supervision of CAA airtraffic control must file a flight plan with the control agency.
This flight plan will include the aircraft identification
aircraft type, speed, take-off point, altitude, route, and
destination. From the information contained in a flight
plan, individual flight-progress strips are prepared and
posted at each fix over which the aircraft will pass. Fig. 3
shows a typical flight-progress strip prepared by a traffic
controller for American Airlines Flight No. 34. The strip
shows the following data: the aircraft is a DC-7 with a
speed of 370 knots, its route of flight is from Chicago
Midway (MDW) over airways designated as V 6, V 168,
and R 15, to Idlewild (IDL). The flight was first estimated
over Youngstown, Ohio, (Y) at 0958 which was later re-

170

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Fig. 5-IBM 650 air-traffic control operation.
Fig. 4-Automatically prepared flight-progress strip.

vised to 1004 and actually reported at 1001. It was originally estimated to pass over Brookville, Pa., (BKL) at
1010 and this estimate was revised at 1016 on the basis of
the Youngstown revision. The altitude was originally
19,000 feet and this was later changed to 15,000 feet.
During a busy hour in the New York center, as many as
1200 to 1500 flight strips will be prepared. The system has
a deadline to meet in that the strips must be prepared and
distributed to the proper controller approximately 30 minutes in advance of the actual arrival of the aircraft over
the fix, in order that he may compare the time and altitude
with that of other flights in the area and make sure that
no conflicts exist. On peak days and at peak hours of
traffic, the time involved in getting flight plans to the center, processing strips, and in getting clearances back to the
pilot sometimes causes the delays in take-off that all of us
have experienced. A great deal of manual effort is involved in gathering the data, preparing and distributing
the flight-progress strips. It is in this area that the first
application of computers is being made.
In 1955, under an Air Navigation Development Board
project, system experimentation at the Technical Development Center began to determine the data-processing requirements for automatic printing of flight-progress strips.
Although the magnetic drum-storage equipment lacked
computing capabilities, a group of specially trained operators were used to simulate the computing functions. These
operators would receive a flight plan and process it by
breaking down into the various fixes over which the aircraft would report. They would then prepare messages in
the form of fix postings and send these messages over a
teletypewriter circuit to the magnetic drum equipment.
The magnetic drum equipment would subsequently read
out each message to one of several printers when it came
time for the flight strip to be displayed in front of the
air-traffic controller.
Fig. 4 shows a flight-progress strip printed during the
evaluation of this system.
During the past year, an IBM 650 computer has been
installed in the Indianapolis Air Route Traffic Control
Center in· order to carry out an operational test of printed
flight-progress strips. See Fig. 5. Traffic controllers, re-

ceiving flight plans by interphone, have been preparing a
punched card for each flight. These cards are fed into the
computer, which determines the route of flight, calculates
estimates, and prints all the strips required automatically.
A complete report of this operation has been prepared and
will soon be published by the CAA as a Technical Development Report. In addition, this subject was also presented in a paper given by G. B. Harwell of the Technical
Development Center during an IBM 650 Scientific Symposium in Endicott, N.Y., the first week of October, 1957.
Plans are now being made to expand the capability of this
installation by adding RAMAC and on-line input-output
f acili ties.
Plans are also being made to install the Model I
Univac File Computer in the traffic-control centers of
New York and Washington, D.C., by next summer for
automatic preparation of flight-progress strips. A prototype of this system is now being assembled at the Technical Development Center in Indianapolis for testing and
evaluation. Fig. 6 is a block diagram of the prototype system. In the Washington and New York installations, space
limitations will require that the computer be located in a
distant room or even on another floor of the building which
houses the center-operation area. Provisions must be made,
however, for input and output at several locations in the
center itself. In the initial phases of the operation, the
principal input to the system will be in the form of flight
plans. Flight-plan data will be fed to the computer in two
ways. Those which are received by interphone or off-line
teleprinter will be encoded by operators in the center
itself and transmitted to the computer. Those received
from another computer-equipped center or remote station
in the local area properly equipped for on-line communication will be fed directly into the computer.
One feature of the system will be the adaptation of a
scanner and speed-conversion device which Remington
Rand is producing for their airline reservations systems.
By means of this equipment, a number of communication
channels, operating at teletypewriter speeds, can work into
a single high-speed input to the computer.
At the manual input position, flight plans which are
received by interphone will be encoded by operators and
transmitted to the computer. The computer will analyze

Fenimore: Real-Time Data Processing for CAA Air-Traffic Control
CU

CAA

TELETYPEWRITER
ClRCUTS

TELETYPEWRITER
ClRCIITS

AIR ROUTE TRAFFIC
CONTROL CENTER
AREA

171

FLtGHT
PROGRESS
STRIPS

FLIGHT PROGRESS BOAROS

Fig. 6-Block diagram of prototype data-processing system.
Fig. 7-Flight data-entry equipment.

the route of flight, determine over which fixes the aircraft
will pass, estimate the times of arrival over each fix asse~ble data in the proper form for printing flight-pro~ress
stnps, and transmit the data to one or more of the flightprogress strip printers in the center area. At the same time
it will keep a record of the flight plan itself and all fix
data within its memory. Flight plans which enter the system from distant points over the communications lines will
go directly into the computer for processing.
. If a flight is to continue into an area served by another
air-traffIc-control center, flight-plan data must be forwarded to that center in order that strips may be prepared.
At th~ appropriate time, the computer wi11locate the flightplan mformation in its memory and transmit automaticall?, over the communications circuit to the proper destinatIon. A configuration of selective distributor and speedconversion units will permit one high-speed output from
the computer to serve a number of low-speed communications channels. During the evaluation of the prototype
system, the procedures of transmitting flight plans to Air
Defense and messages to terminal areas regarding landing aircraft will be investigated.
A supervisory console will be provided in the computer
area for maintenance operations, and an additional printer
wil~ ?e located. in the center area near the manual input
posltlOn for flIght plans which the computer may reject
as erroneous.
A problem area in the preparation of data for automatic
handling by computers is the rigid format required to
insure that the computer treats each part of the data received in its proper category. This is particularly true in
the case of a relatively long and involved series of data
such as flight plans which will have an average length of
110 alpha-numeric characters. Two major types of errors
may occur in the preparation of such messages. Operator
errors may range from the addition or deletion of one or
more characters, to failure to follow the proper ground

rules. Communication errors may also occur, and with the
five-unit code of standard teletypewriter systems, these
may pass through the system undetected.
In order to permit rapid and accurate composition of
flight-plan messages for transmission to the computer over
communications circuits, the Technical Development
Center has contracted with Aeronutronic Systems, Inc.,
of Glendale, Calif., for the development of FLIDEN
(Flight Data Entry) equipment. Fig. 7 is their artist's
conception of the device whereby an operator may com:'"
pose a message which is displayed on a cathode-ray tube
and stored electronically on a magnetic drum during composition. If the operator makes an error, he may backspace
rapidly or a character at a time to the position where the
error occurred. Having corrected the error, he may shift
rapidly back to where the message composition was inter~upted and continue. A form on the face of the display
deSIgnates categories of information which should be entered. All fixed characters and functional characters are
entered automatically for ease of composition and accuracy of format. The completed message is checked for accuracy by the equipment before transmission.
For detection of errors during transmission, an error
check character is automatically inserted at the end of each
line of data. This is accomplished by making a longitudinal
count of marks at each code level and adding a check bit
to make the total count odd. The check bits are combined
to produce "nonsense" or check characters which are transmitted with the message. This basic technique is described
by Vincent. 1 At the receiving end, a similar count will be
made and the resulting check character at the end of each
line of text will be compared with the check character transmitted by the FLIDEN equipment. Messages which contain
1 G.
O. Vincent, "Self-checking codes for data transmission"
Automatic C ontral, vol. 5, pp. 46-49; December, 1956.
'

172

PROCEEDINGS OF TJ-IE EASTERN COMPUTER CONFERENCE

errors detected by this means will be directed to the computer supervisor's console printer for manual handling.
The specifications under which the FLIDEN equipment
is being developed require that transmission of flight-plan
messages be made over standard teletypewriter code. However, the logical design of its internal circuitry has been
planned that transmission at higher speeds using six or
seven-level code will be possible with minor modifications
in the final equipment.
With the advent of this type of equipment in the field,
on-line filing of flight plans directly into the computer
system for flights originating in the center-control area
will be possible. This will ultimately reduce the quantity of
flight plans which are received by interphone in the center
area and manually prepared for insertion into the computer.

Reducing the clerical workload of controllers in the
CAA air-traffic control system is only the first step. Anyone who is familiar with the background of thinking on
air-traffic-control systems and the great volumes of studies,
papers, programs, and system configurations produced
during the past ten years may well ask, "What about pictorial displays, radar inputs, data links, automatic position
reports, etc.?" This introduction of computers into what
is presently a completely manual operation should speed
the day when these very desirable features will be included.
With the formation of the Airways Modernization Board
and the rapid progress that is being made in their dataprocessing and display program, it is anticipated that airtraffic control operations will soon be provided with powerful tools commensurate with the jet-age traffic-control
problem.

Design Techniques for Multiple Interconnected
On-Line Data Processors
F.

J. GAFFNEyt

INTRODUCTION

HE application of data-processing techniques to
large-scale nation wide industries, such as the transportation industry, has resulted in the need for remoting of subscribers and for multiple data processors
regionally located. These equipments form elements of a
complete integrated data-processing system or network of
systems to solve such problems as reservations control on
a complete system basis for railroads and airlines. Early
applications of digital data-processing equipment to solve
the reservation problem were concerned with limited area
problems. 1 The improvement and extension of· these systems to integrate the entire reservation-space problem has
been a logical evolution.
SYSTEM CONSIDERATIONS

In a reservations system, large numbers of inquiries as
to availability of space as well as actual transactions involving sales, cancellations, and waitlistillgmustbe processed each minute. Responses to these inquiries . from
agents all over the country must be in the order of seconds,
so that minimum delay is introduced in handling a customer
inquiry over the telephone; further, these inquiries may
t Tclcrcgister Corp., Stamford, Conn.

M. L. Haselton and E. L.Schmidt, "Automatic inventory
system for air travel reservations," Elec. Eng.) vol. 73, pp. 641-646;
July, 1954.
1

AND

S. LEVINEt

occur at almost any time of day, particularly in systems
for nationwide carriers.
System analysis of these traffic, speed, and operational
requirements has led to the design of integrated systems
for airlines and railroads which include a number of data
processors, each of which provides access to agents using
specially designed input-output devices, either in the large
reservation offices adjacent to the data processor, or in
remotely located offices using specially designed communications equipment. The individual data processors are interconnected by means of high-speed data links so that
updating of regional processors by a central data processor
can be accomplished rapidly. These systems must be capable of continuous operation for at least 22 hours a day
so that reliability considerations are important factors in
the design of the system.
The shutdown of a data processor for maintenance
failure for more than a few minutes during peak traffic
periods can result in large queues, with resulting loss in
business. Requirements for uptime (i.e., the ratio of time
the system is actually available to the scheduled time) for
systems of this type are in the order of 99.5 per cent.
REMOTING OF INPUT-OUTPUT DEVICES

One of the first extensions of the Reservisor concept
was that of providing facilities for the connection of remote agents to the central information store. One method

Gaffney and Levine: Multiple Interconnected On-Line Data Proccssors

~
IS'

173

Sl'EI(£R-------~

A/52

T=mean processing time by (1) H=T+V+PS 1 +(I-P)S2
DCOT.

(2) E

time

transmission
message.

per

Mean
(3) Response R= H
Time

= call rate.
Sl = mean roll
C

= CH

E
call time of (4) P
calls delayed.
Mean
mean roll call time of (5) Delay D
calls not delayed.
Time

= proportion
layed.

+ j)

calls

= mean service time

de-

Ell

= ~ --

Mean
(6) R~sponse R
TIme

Fig. 2-Pan American Airways reservisor network.

l-E

=~ H

[2-EJ
-l-E

(DCOT
processing pI us roll call
and transmission times) .

effective load factor of
server (DCOT together
with communication links).
Fig. I-Queueing analysis.

for accomplishing this utilized an editor for ordinary teletype messages arranged in a standard format. Space sales
or cancellations can then be routed to the data processor
by existing communications facilities such as the 81-D-l
system of A. T. & T. Automatic broadcasting of stop and
resume-sale messages can be accomplished by the data
processor when indicated by the inventory level of a given
flight. This type of system has been installed recently for
Braniff Airways and is currently in use.
There exists, however, the requirement for direct connection of agent sets and the data processor to serve locations which generate appreciable traffic in order to minimize time delays between sales of space and their effect
on the central inventory.
One obvious solution to this problem is to connect each
remotely located group of agent sets to the data processor
by means of individual teletype lines. This procedure,
while economically applicable to centers of high-traffic
generation, is uneconomical of line costs for many remote
locations. It is necessary, therefore, to arrange for common utilization of a transmission line by a number of
remote locations along its length. This can be done on a
time-sharing basis.
In order to minimize transmission delays it is necessary
to provide way-station selection and roll-calling equipment
which operates in minimum 'time consistent with the bandwidth of the communication facility utilized. Such equipment has been developed in a joint Teleregister-Western
Union development program.
Fig. 1 shows a typical arrangement of remote locations
and central processor. The central processor is sequentially connected, by means of a master seeker, to a number

of remote line terminations called Distant Central Office
Transceivers, and to multiples which serve local groups
of keysets. Each of these inputs is served in turn and
remains connected to the processor until a reply has been
generated and transmitted .
Each Distant Central Office Transceiver (DCOT) is
the termination of a teletype line along which are located
remote stations, each equipped with a Distant Remote
Transceiver (DRT). Each. DRT, in turn, connects sequentially through a seeker to a number of agent sets at its
location.
In the sequence of operations, a remote agent set bids
for access to its DRT. The DRT then bids, by means of a
roll-call procedure to be described presently, for access to
its DCOT. The DCOT, in turn, bids for access to the
data processor. When this chain of events has been completed, the agent set sends its message over the line and
receives its reply from the processor. It is then automatically disconnected from the line and another transaction
from a different agent set proceeds in the same way. Fig. 1
also shows representative equations for queueing time for a
DRT under such a set of conditions. Eq. (6) shows mean
response time at a DRT for an idealized system. Actual
response time for any given fraction of the calls for consantholding time at the processor is obtained from charts.
It is possible to predict, for example, the fraction of calls
which will be delayed one holding time or five holding
times. In actual practice with current equipment utilizing
a 75-word-per-minute teletype line, the mean response time·
is of the order of several seconds.
The roll-call operation which connects the DRT's successively to the DCOT is of the Round Robin type, that
is, each DRT not 'having traffic when queried generates
the call letter of the following DRT. Single letter addresses
are utilized for economy of line time. Parity check of all
messages is provided.
A remoting system utilizing these techniques was installed early this year for Pan American Airways and is
shown in Fig. 2. Thecent.ral data storage is located in
Long Island City, N.Y. The communications system serves
26 cities arranged along four transmission lines.

174

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Fig. 4-Railroad Reservisor transmission network.
Fig. 3-Integrated airline reservation system.
INTEGRATED AIRLINES RESERVATION SYSTEM

Where traffic volumes in regional areas are high, greater
economy can be served by the use of regional data processors connected together to form an integrated system. Such
a system is shown in Fig. 3. The regional availability
processors, in this system, do not store an actual seatcount type of inventory, but do store several levels of
availability status (such as available, not available, in cushion, etc.) for each flight or flight segment. Automatic provision is made for routing sell-cancel calls over the tie line
to the central processor which does store the actual inventory on all flights and flight segments. Since there are s~v
eral availability requests for each sell-cancel transaction,
the availability processors act as filters to reduce the tieline traffic. The availability processors are updated automatically by the central inventory processor whenever the
inventory level of a given flight segment reaches a value
which dictates a different availability status.
The tie lines utilize voice channels which are capable of
transmission at the rate of 750 bits per second. Tie-line
terminal equipment incorporates buffer stores of the magnetic-core type to permit optimum-line loading.
Each of the availability processors serves local keysets
as well as a number of remote lines which feed traffic
from cities in its regional area.
INTEGRATED RAILROAD RESERVATION SYSTEM

The requirements of a railroad reservation system differ
considerably from those of an airline system, due principally to the requirements for storage of information on
a large number of types of reserved space. The economies
of regional storage of data can still be effected, however,
since sales of space on certain trains will predominate in
the region from which the trains depart.
In the Teleregister system designed to serve the New
I-laven, the New York Central, and the Atchison, Topeka,
and Santa Fe Railroads, two identical data processors are
located in New York City and Chicago, respectively. These

DUP
TIE
LINE

Fig. 5-Railroad Reservisor tie-line terminal.

processors are connected by a 75-word-per-minute tie line.
This tie line enables either center to direct an agent set
message to the other center and receive a reply, to receive a
message and send a reply, to receive a printer message,
and to send a printer message. The total inventory is split
between the two processors in such a way as to minimize
tie-line traffic. Agent set calls are automatically addressed
to the appropriate data processor as determined by the particular space involved in the transaction.
A map of this system is shown in Fig. 4. As in the case
of the airline system, each processor serves a number of
remote lines along each of which are located cities having
a group of agent sets. Each processor also serves agent sets
in its local reservation rooms and ticket offices.
A block diagram of the tie-line equipment is shown in
Fig. 5. A data-processing center has local agent sets and
long-line terminals serving agent sets at remote cities. Any
of the local agent sets or long-line terminals has access to
either the local Data-Handling Unit or the remote DataHandling Unit. The determination of which Data-Handling Unit is to be used is automatic, and is determined by
the addressing code generated by the request for specific
train routes when a particular plate is inserted into an
agent set. When the remote Data-Handling Unit is re-

Gaffney and Lev.ine: Multiple Interconnected On-Line Data Processors
qui red, the agent set or long-line terminal bids through
the tie-line seeker for the send selector. When an input is
connected, this seeker bids against the printer-switching
equipment and the tie-line register for the use of the send
control. When this unit is available, it controls the pulsing
out of the message in 5-element,· 7.42 unit code to the remote tie-line terminal. A distinctive character precedes the
message to identify it as an agent set query. The receiving
control at the remote central station recognizes the character and causes the receive selector to steer the incoming
message to the tie-line register where it is stored in relays.
Upon completion of the outpulsing, the send control in
the originating station becomes available for inputs from
the printer-switching equipment or from its own tie-line
register. The receive control at the terminal station is now
available for printer messages or a reply to one of its own
agent-set queries.
Upon completion of the message the remote central station tie-line register bids against the local agent sets and
long-line terminals for the Data-Handling Unit. When it
has received and stored its reply, it bids against the tie-line
seeker and printer-switching equipment for the use of the
send control. When connected, the send control pulses out
the reply preceded by a distinctive character to identify it
as such. The receiving control at the original station recognizes this switching character and thereupon switches
the reply through the tie-line seeker to the original agent
set or long-line terminal.
Certain agent set calls and some Data-Handling Unit
operations cause printer messages to be perforated by the
Data-Handling Unit 600-wpm, 5-unit punches. These
printer messages can be directed to local Receive-Only
printers, printers on the long lines connected to the center,
or to local or remote printers associated with the remote
Data-Handling Unit. The perforated messages have sufficient addressing characters to select system, line, station
and printer. When the printer-switching equipment sees
an address associated with the remote Data-Handling Unit,
it bids against the tie-line seeker and tie-line register for
the send control. When given access, it pulses out its message preceded by an appropriate switching character. This
switching character directs the receiving control to feed the
incoming message into the FRXD. This FRXD or RT is
a receiving reperforator and a transmitter distributor. The
machine is built so that a loop of tape can be stored between the punch and distributor. This loop forms as the
message comes in. When the message is complete, the
FRXD is connected to the printer-switching equipment
which selects the designated local printer or long-line terminal associated with a remote printer and completes the
routing of the message. The tie-line equipment is designed
such that two agent set messages, one from each end, can
be simultaneously in progress. By the same token a printer
message from one end can be in progress at the same time
as an agent set query from the other end.
The tie-line equipment is equipped with odd parity
checking and various time outs to prevent service delays.

175

DHU
-SKR

INPUT
MULT·

DRUM
STORE

Fig. 6-Dual data processor operation.
RELIABILITY CONSIDERATIONS

In an on-line system of interconnected Data-Handling
Units, reliability of the various components of the over-all
system becomes a prime consideration. The system must
be designed and maintained so as to secure long periods of
trouble-free operation and rapid correction of failures
which do occur. Means must be provided wherever possible to permit independent operation of the various units,
so that the system may be operated with somewhat reduced performance characteristics, even when one of the
component subsystems has failed.
Of great importance, of course, in maintaining this reliability is the use of carefully chosen electronic and electromechanical components of high quality. To effect this
requires an active standards program and careful life testing of components intended to be added to the standard
list. In the equipment described, for example, vacuum
tubes expressly designed for long life have been chosen.
These tubes, when operated under derated conditions in
the equipment, have demonstrated average lifetimes of
well over 50,000 hours. Bifurcated contact relays of the
telephone type are used for electromechanical switching
and are capable of several million operations before any
need for adjustment.
To further assure reliability of the central Data-Handling Units, portions of the equipment such as the electronic
data processor are duplicated. Reliability can be increased
by a large factor through the use of this technique. For
example, if the probability of failure of a given unit in a
specific time interval is one part in a thousand, the probability of failure of both of two identical such units during
the same specific time interval is of the order of one part
in a million.
A block diagram of such a duplicate processor is shown in
Fig. 6. Calls coming in to the Data-Handling Unit seeker
shown at the right-hand side of the diagram are stored in
an input register and fed simultaneously to two DataHandling Units. The process of a given call as it progresses through each of the units is checked by means of a
cross-checking circuit, and the call rejected, should a check
not be obtained. Where the call is one which requires writing on the drum to change the stored information, this is
done simultaneously on a check track fed by one of th('

176

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Fig. 7.

Data-Handling Units and on the appropriate inventory
track fed by the other Data-Handling Unit and selected
by a track-selection pyramid. The changed information is
then read back from both tracks through the Data-Handling Units and compared by the checking circuit for certification. Should verification not be obtained, the call is
printed out so that corrective action may be taken.
In the design of some systems of this type, facilities are
provided for automatic switchover in case of failure of one
of the systems to the remaining system, which can operate
alone. Indication of this type of operation is provided to
maintenance personnel.
To provide further reliability, the Data-Handling Units
are operated in a temperature-controlled environment and
with fully attended maintenance. Maintenance and adjustment periods of approximately two hours per day provide
opportunity for the employment of marginal-checking
techniques for weeding out components for which failure is incipient.
Open rack type of construction is used for the equipment to enable access to all of the components, in order
that repair time may be minimized. The equipment is provided with a maintenance console which indicates to maintenance personnel the progress of calls through the DataHandling Unit and which provides a facility for the introduction of test calls in order that failures may rapidly
be located. A typical system arrangement is shown in
Fig. 7.
Of equal importance in the over-all problem of system
reliability is the minimization of human error by operators
who have access to the system and the facility to change the
stored in·formation. It is' of paramount importance that
agent sets be well designed froth the human engineering
standpoint, with a view to simplicity, ease of operation
and protection against human error. Fig. 8 shows an
agent set designed for airlines use in accordance with these
principles. The use of a plate inserted into the agent set
provides an almost fool-proof method of addressing to a
particular location on the magnetic drum. It provides the

Fig. 8.

Fig. 9.

facility for automatically coding information which would
otherwise require a number of button depressions. The
addresses to which the plate directs the inquiry are clearly
shown in printed English on the plate itself, which can
also be used to display other information, such as fares.
The insertion of the plate in one position together with
the use of eight matching keys located immediately below
the lower edge of the plate enables simultaneous access to
the store for availability information on each of eight
flights, legs, or flight segments. Depression of the appropriate button then restricts sell or cancel action to the par.ticular flight or leg involved.
Facilities for inserting, by means of clearly marked buttons, information on the number of seats required, the
data and the action required (such as sell, cancel, waitEst,
etc.,) are provided. This design has been shown by ex-

Gaffney and Levine: Multiple Interconnected On-Line Data Processors
perience to require very little operator training and to
result in few operator errors.
The same general approach was employed in the design
of a similar set for use with the Railroad Reservisor system. This is shown in Fig. 9. Here it is necessary to provide additional buttons due to the requirement for listing
a number of types of space, such as seats, bedrooms, etc.,
but the design of the equipment is basically similar to that
of the airlines agent set. In the case of the Railroad Reservisor, the reply-back information from the central data
processor is also somewhat more complex, and it has been

Discussion
W. A. Morgan (International Business
Machines, Owego, N.Y.): Do you buffer
your inputs and outputs or allow each
agent set to tie up the data-handling unit
during processing time?
Mr. Gaffney: Each agent set contains
its own buffer register. The call is set up
in the agent set by the operator and a
master seeker permits each agent set to
have access to the data-handling unit in
turn. The data-handling unit has its own
input-output buffer and processes agent
calls at its own high-speed rate.
E. Ziolkowski (Datamatic): In interrogating the system, how is the problem
handled when two requests are made simultaneously for only one available reservation? Can conflict occur?
Mr. Gaffney: There is no conflict. The
data-handling unit processes the calls sequentially so that only the first one having
access to the equipment will obtain the
reservation; the second call will be rejected.
Claude Kagan (Western Electric Co.):
What error detection and correction facilities are provided in the Railroad Reservisor
tie-line system?
Mr. Levine: In this system, the primary error detection facility is an oddeven parity check on each character. In
addition, there is an over-all message character count. Error detection facilities are
available on the data-handling unit so that
nonallowable codes will be rej ected. Error
correction facilities are not provided; however, in the event of a detected error, the
agent set receives an error signal and no
change in the inventory is made. The call
is then repeated.
Mr. Harris (Stromberg Carlson):
Please describe the physical facilities over

177

found best to display this information in printed form
rather than with lights, as in the case of the airline set. In
the illustration shown, the printer is shown mounted adjacent to the agent set. The reply-back information is ejected
from the printer on prepared forms, one copy of which
can be given to the customer.
It is believed that the systems described above represent tangible forward steps in the development of on-line
data-processing systems, which enable these systems to be
greatly extended geographically, with consequent increased
utility.

which your high-speed-750 bits a second
-transmission takes place. For example,
carrier, long-distance telephone, leased
wire.
Mr. Levine: The physical plant facility
to be used for transmission of 750 bits
per second data will be a private wire,
leased, telephone quality voice channel. The
terminating device will be a data modulator-demodulator unit to be furnished by the
communications common carriers. Our data
terminal equipment will include necessary
buffer registers and will serialize the data
in the form of a series of dc pulses fed
to the data modulator-demodulator.
Mr. Harris: How many stations does
the Pan-Am system serve? By station, I
mean one group of individual agents' equipment.
Mr. Levine: In the Pan-Am system at
present we have 26 cities served by the
system, and a total of 120 agent sets.
P. F. Radue (Automatic Electric Co.):
What technique is used to aSsure correct
transmission of information where parity
check leaves gaps?
Mr. Levine: The technique for error
detection that we expect to use will include
both vertical and horizontal parity checks.
The vertical parity check is a single parity
check on each character while the horizontal check is performed on the entire message. This technique is expected to reduce
the number of undetected errors to a
negligible value. In addition, as discussed
above with reference to the railroad system
tie line, the data processor will reject any
nonallowed codes. The message will be retained in a buffer register at the transmitting end until an acknowledgment signal
is received. The message will be repeated
if errors are detected. If the acknowledgment is not received, a print-out will occur
after a suitable time out.

F. A. Reynolds: What is the effect of
errors on increasing the traffic on the
lines? How do errors break down between
operator errors and facility errors?
Mr. Levine: Both of these questions
are difficult to answer precisely in that
systems of the type discussed have not been
in operation long enough to provide a
good statistical sample. The first part, on
the effect of errors on increasing traffic,
may be answered as follows. With proper
operation and maintenance of both lines
and terminal equipment, the number of retransmissions due to error should not exceed approximately 1 per cent. This type
of performance has been achieved. However, during initial breaking periods this
rate has been somewhat higher. The retransmissions are not expected to increase
traffic sufficiently to cause difficulty. The
second part of the question concerning distribution of operator and facility errors
may be answered as follows. Error checking facilities are included in the operator's
equipment to reject improper input data;
however, certain operator errors will be
undetected except through audit procedures.
The number of undetected operator errors
should be low.
B. Hasbrouck (Atlantic Refining Co.):
If space is sold out, can all interested regional processors reject such requests locally? If space is still available, can any
requests be answered on the regional level
or must each such request go individually
to the proper central for servicing?
Mr. Levine: Yes, the regional processor
rej ects requests locally if space is sold out
or permits acceptance of sales if space is
available. Agent sets receive their answers
directly from the regional processor which
forwards the transaction to the central
processor.

178

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Reservations Communications Utilizing a General
Purpose Digital Computer
R. A. MeAVOYi'

'
T

HIS paper describes a communications system developed for the purpose of establishing communications between a general purpose digital computer and
a large number of employees whose work requires frequent and immediate access to the services of the computer. The communications system is, to a degree, "general purpose" in nature; however, in the specific application to be described, the system will be used by airline
employees in communicating with a computer which is engaged primarily in processing airline reservations data.
An over-all view of the system may be obtained by reference to Fig. 1. The rectangular area at the top of the
drawing represents the common location of those portions
of the system symbolized there, including the central computer. In the lower right-hand portion of the drawing, a
smaller area represents a typical location remote from the
computer. The remaining portion of the drawing contains
a brief legend for later reference; it may be disregarded
at this time.
Referring again to the upper portion of the drawing,
the group of small circles to the left represents a large
quantity of units called "agent sets." The agent set is the
primary communications device through which the individual employee transmits and receives information. This
device will be described quite completely later. At this
point, it is only important to note that messages to be
transmitted are formed by a combination of actions including the pressing of buttons and the operation of other controls. Messages received by the agent set are displayed in
the form of illuminated signs or signals.
To the right of the agent sets, the block labeled HSPS
represents a group of units called high-speed programmer
scanners, which perform the following functions.

1---------------------------- 1
I

CENTRAL

COMP.UTER

SITE

looooooo~

II
I. .

gggg~oo~

*gooSoEJ~og:::::::::
~~~~~~

OOOOO?oo----

L _____ ~-----------~----------J

t---TOOTH..
,.MOTE LOCATIONS

1--

--1

1
UFC' •

I
·1

UNIVAC INPUT-OUTPUT ADAPTER.

UNIVAC HIGH SPEED PROGR~MMER -SCANNER.

AND READ - PUNCH UNIT.

BELL
SYSTEM

I
I

- ' UNIVAC .FILE-COMPUTER, MODEL 1. INCLUDING FIVE G.S, DRUMS

1·0 ADAP. -

UNIVAC LOW SPEED PROGRAMMER _ SCANNER.

A. T. & T. CO STATION CONTROL UNIT, SERIAL-PARALLEL
CONVERTER, PARALlEL:- SERIAL CONVERTER.

I
I

AG;:~ S~TS
ANY OHE OF 20

L~~~c~~.-l

Fig. 1.

ceeds immediately to serve the next agent set after
translating the reply for the first.
4) They perform parity check and character count
check on received messages. Upon recognition of
error, the programmer scanner causes repetition of
the reply. If the second reply is error-free, normal
operation is resumed. If the error persists, the programmer scanner will cause an appropriate sign to be
illuminated on the agent set. Depending upon the nature of the input message, the agent set will display
a sign reading ERROR or a sign reading RESET.

Continuing to the right on the diagram, the next box
represents a unit called "input-output adapter." This unit
relays messages between the high-speed programmer scanners and a section of the computer which we shall call the
"message box." The loading and unloading of the "message
box" by the input-output adapter is independent of the
processing functions of the computer.
1) They provide service, sequentially, to agent sets
The input-output adapter loads the message box and
which have complete messages ready for transmis- then notifies the computer to that effect. The computer
SlOn.
subsequently processes the message, deposits a reply in the
2) They establish the sequence in which the characters message box, and notifies the input-output adapter. The
of the message are transmitted and cooperate in the adapter immediately relays the reply back down the line
transmission of messages at the rate of 200 charac- to the agent set.
ters per second.
The input-output adapter also checks parity and charac3) They receive reply messages from the computer and ter count, notifying the computer when an error is obtranslate the replies into selective illumination of served. Error notification to the computer is used to make
signs or signals on the agent set, causing the illumina- appropriate modification to the program.
tion to persist until retired by a specific manual opThe entire communications process just described oceration at the agent set. This latter activity does not cupies a period of time ranging from 90 to 125 msec detotally occupy the programmer scanner which pro- pending upon the nature of the input message. This figure
does not include processing time in the computer which
t Supt., Communications-Data Processing, Eastern Air Lines
.
.
Fl
(
will range f rom a bout 100 msec to shghtly
more than one
·
I nc., Mlaml, a.

McAvoy: Reservations C0111!lnunications Utilizing a General Purpose Computer
second. Most input messages will be processed by the file
computer in less than 250 msec.
Referring again to Fig. 1, the large box labeled UFC
is, of course, a representation of the Univac file computer.
Although only one box is shown on the diagram, the computer system will be composed of approximately a dozen
large equipment cabinets. The system will include five
magnetic drums having a storage capacity of 900,000
alpha-numeric characters. The storage system is capable of
expansion to thirty-three drums having a storage capacity
of nearly 6,000,000 characters.
At this point it is appropriate to call attention to the fact
that so far we have been referring to a communications
system which is wholly contained on our premises. The
greatest distance intervening between an agent set and the
file computer is not likely to exceed 300 feet. Let us now
direct our attention to the communications system which
serves agent sets at remote locations.
Referring now to the lower right-hand portion of Fig. 1
we note that the area represents "anyone of 20 remote
locations." These locations may be as far distant from the
computer location as one chooses to think, whether it be
one mile, a hundred miles, a thousand miles, or more. The
only requirement is that the remote location be capable of
being connected to the computer location by a dependable
telegraph circuit.
The limit of twenty remote locations is established by
choice. Up to forty remote locations may be served by one
telegraph network of the type referred to herein.
If we follow the path between the remote agent sets and
the Univac File Computer we note that the labels in two of
the boxes are somewhat familiar. At the top end of the
path next to the computer we find an input-output adapter.
This unit performs essentially the same functions as the
other unit of the same name. The essential difference is
that it contains a few relays which are required to compensate for the lower speed of communication.
At the other end of the path, we see a box labeled
LSPS which represents one unit named "low-speed programmer scanner." This unit performs essentially the same
function as the high-speed programmer scanner including
parity checks and character count. In the case of this unit
there is not only a difference in speed of operation but,
as the diagram indicates, one unit serves a maximum of
eight agent sets. Although the unit is capable of operating
speeds up to 20 characters per second, it is limited by the
speed of the telegraphic equipment which is, in this case,
10 characters per second.
A feature of the low-speed programmer scanner is the
ability to operate two such units as a single unit with a
capacity to serve as many as sixteen agent sets. Doubling
of units in this manner does not change the appearance
to the telegraph network.
As indicated on the diagram, the telegraph network is
leased from the Bell System. The Bell System in this case
is represented by the A.T.&T. Company which operates
most of the intercity <;:ircuits of the Bell System. The basic

179

system is called "83-B-l Selective Calling System;" in
this application it is modified for use with the Univac File
Computer.
A complete description of this system is beyond the
scope of this paper. However, the principal characteristics
of its operation will be evident in the following description
of a cycle of operation.
The elements of the 83-B-l system in this application
are a "control station" located at the central computer site
and a number of "station control units"-one at each remote location.
The "control station" initiates a continuous series of
events by transmitting a particular set of two characters
which are received by all of the station control units. This
set of two characters is referred to as a "start pattern."
One, and only one, of the station control units recognizes
this particular set of two characters and responds in one
of two ways. If the associated low-speed programmer
scanner has not indicated a need for service, the station
control unit transmits one particular character which when
received at the" control station" causes that station to transmit a different set of two characters; i.e., a different start
pattern, thus directing a different remote station to respond.
If the low-speed programmer scanner has indicated a
need for service, the station control unit transmits characters submitted to it by the low-speed programmer scanner.
The characters submitted by the LSPS are, of course, those
which are represented by the position of the controls on
the particular agent set which is then being served. Should
the "control station" fail to receive a reply of either type,
audible and visual alarms will be set off at the control station location.
At the completion of this transmission the telegraph
network becomes idle until the control section of the File
Computer tells the input-output adapter that a reply message is ready in the "message box." The control station
transmits the characters submitted to it by the input-output
adapter, thus completing the two-day communication.
I f a second agent set at the remote location is awaiting
service at this time, its message will be transmitted after a
delay ranging from roughly 50 to 150 msec. The amount
of the delay is determined primarily by the relative positions of the two agent sets in the fixed service sequence. As
the number of sets requiring service increases, this delay
decreases.
When each of the active agent sets at a particular location has been served once, the low-speed programmer
scanner submits for transmission a particular set of two
characters, which are then transmitted by the station control unit. The control station at the central computer location responds by transmitting a start pattern for a different remote station.
When all of the remote stations have been served, the
control station begins a new cycle without delay and the
process continues without interruption until the hour arrives for the start of daily maintenance procedures at the
computer center.

180

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

The telegraph system in this application will be operated at the nominal speed of 100 words per minute or
approximately 10 characters per second. The system includes protective features which supply visual and/or
audible alarms in the event of various types of failures
which may occur.
Having completed the general description of the communications system, attention may now be directed to a
more detailed examination of a most remarkable piece of
communications equipment-the agent set.
Fig. 2 is a photograph of an agent set. The predominant
feature of this unit is a photographic projection system in
which a light beam is passed through a single frame of a
conventional 35-mm film projecting the image of the film
on a viewing screen on the upper panel of the agent set.
Each frame of film is precisely mounted in a holder two
inches square. A cartridge containing thirty holders, i.e.,
thirty pictures, fits loosely in a tunnel, the opening of
which is visible at the lower right-hand corner of the agent
set.
A rack of teeth along the edge of the film cartridge
engages a pinion gear controlled by a knurled wheel visible
along the right edge of the lower panel. The top of the
film cartridge contains an index which is visible to the
operator of the set.
Therefore, a selection of thirty images is immediately
available to the operator. Additional groups of thirty images may be referred to quickly by removing one film cartridge and inserting another.
The cartridges incorporate a simple locking mechanism
which prevents accidental removal of individual film
holders. The potential difficulty associated with spilling and
subsequent faulty rearrangement of the file is thus avoided.
On each individual film, a narrow strip along the right
edge and a narrow strip along the top are reserved for
digital coding in a five-channel code. These strip areas are
appropriately divided into three discrete areas each of
which in turn is divided into five areas. Each of the five
areas is made black or clear white to represent binary "I" or
binary "0." Thus, there is created the means for specifically
identifying anyone of a thousand images, assuming the
identification system is based on the decimal system.
After a film slide has been selected by reference to the
index and by operation of the knurled wheel, a slide lever
is moved to the left, causing the projection lamp to be
lighted and the selected film to intersect the resultant light beam. In addition to the image which appears
on the screen, a pattern of light described by the coding
along two edges of the film is impressed on fifteen lightsensitive elements located behind the upper panel bordering
the viewing screen. The condition of these elements is subsequently sensed by the programmer scanner to transmit a
specific set of three digits to the computer.
All other information transmitted from the agent set
is obtained by sensing the five contacts of the buttons
which have been depressed. Each button is one of a particular set; the button supplies the decimal information in

Fig. 2.

five-level code while the identity of the set of buttons is
retained by virtue of the position of the digit in the total
message. Mechanical interlocks prevent the selection of
more than one button of a set.
A system of electrical interlocks prevents the agent set
from attempting to establish communication if any of the
required elements of a message have not been entered.
Before referring to the next illustration, note that there
are two sets of buttons along the left edge of the viewing
screen and one set of buttons along the lower edge.
Refer now to Fig. 3, which is an example of an image
which might appear on the screen of the agent set in our
application of airline reservations.
An employee wishing to inquire about the availability of
two reservations from N ew York to Houston would have
selected this slide. The employee would then enter the set
of buttons meaning FROM and push the particular button to the left of the words NEW YORK, and next, the
TO button opposite the word HOUSTON and other buttons specifying month, day, number of seats, and type of
transaction.
No further manual operation is required to code this
message, gain access to the communications circuit, transmit the message, consult the records, prepare a reply,
transmit the reply, receive it, decode it, and display the
answer in front of the agent. Yet the entire operation takes
place in less time than it takes to describe it.
The answer to an inquiry of the type illustrated covers
the availability of all flights represented in the horizontal

McAvoy: Reservations Communications Utilizing a General Purpose Computer

Fig. 3.

columns corresponding to the FROM, TO buttons selected.
Referring back now to our picture of the agent set,
Fig. 2, attention should be directed to a long rectangular
area near the bottom edge of the upper panel. In this area
are located the twenty-eight individual signs or signals
which may be illuminated to provide a direct answer or to
signify an answer to an inquiry.
The twelve lamps to the left, arranged in three rows and
four columns, are individually controlled .and provide back
illumination to words or phrases which, in theairIine application, constitute the range of replies to inquiries concerning estimated time of arrival or departure of the flight
specified in the input message.
The sixteen lamps to the right are arranged in eight
pairs, each pair corresponding with one of the buttons im-

Discussion
Mr. Johnson (Adalia Limited): What
is your specification on error rate, and how
much redundancy have you .found it necessary to add to achieve this specification
over leased communications lines? What
error correction and detection facilities are
provided in the remote agent set system?
Answer: The agent set and its associated programmer scanner cooperate to reject attempted transactions which omit
needed data. For example, if an agent
should attempt to book space on a flight
without specifying the number of seats, an
indicator on the agent set would be lighted
reading RE-ENTER. The message would

181

mediately above it and thus corresponding with the flight
which appears in that column on the viewing panel. Each
pair of lights may have one of three acceptable conditions
of illumination which represent one of three conditions of
availability; i.e., available, not available but expected to
become available, not available and not expected to become available. The fourth possible condition, which is
both lights not lighted, is not an acceptable condition and
is interpreted as evidence of malfunction.
N ear the center of the bottom panel, just to the right of
the four columns of buttons, nine lamps provide back illumination for an equal number of words or phrases which
supply verification or instructions to the operator.
In addition to the words ERROR and RESET previously referred to, there is one other instruction word,
WAIT. Illumination of the WAIT sign informs the operator that transmission has started. This sign remains on until
a reply is received.
The set is used by the airline employee to adjust the
central inventory when reservations are made or cancelled
or an auxiliary inventory may be adjusted to reflect unsatisfied demand. Each such transaction is verified by illumination of an appropriate sign in this bank.
Unfortunately, time does not permit description of the
intended expansion of the system. Much of it can be deduced from the communications capabilities described.
Widely separated File Computers can communicate with
one another through their "message boxes" and much
simpler teletype networks than the 83-B-1.
An accessory now under development will establish an
interlocking relationship between a specific punched card
record and a specific agent set transaction.
An accessory available immediately will provide discrete
identification of the agent set, thus opening the door to
measurement of employee effectiveness and other useful
benefits.
Credit for the development of this system is due to the
many individuals in Remington Rand Univac, A.T.&.T.
Company, and Eastern Air' Lines, Inc. who cooperated
with one another so magnificently to convert ideas to
reality.

not proceed further than the associated programmer scanner, thus avoiding waste of
valuable communications and data-processing time.
We use a five-channel code in transmitting input and output messages between
the agent sets and the input-output adapters.
The four low-order channels are used to
form an "excess-three" binary code which
is identical with the four low-order channels of the Univac seven-level code. The
fifth level is a checking level in which a
bihary "1" is stored or not stored so as
to form a character having an odd number
of bits. A parity check is performed on
the receiving end of each interunit communication.

Upon recognition of a parity error, appropriate action is taken depending upon
the stage at which the error becomes apparent. Generally, the agent set signals
RE-ENTER when no inventory adjustment
was likely, or ERROR when there was a
possibility of inventory adjustment.
Question: If error is indicated 'On the
agent set in a remote area, is the agent
able to determine from the nature of the
error indication whether the error occurred in the equipment in his remote area
or in the central c'Omputer?
Answer: The agent set has two backilluminated signs reading RE-ENTER and
ERROR, respectively. In normal operation,
the agent's entry of a transaction is fol-

182

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

lowed by illumination of a sign reading
WAIT. This informs the agent that his
message has been transmitted from the
programmer scanner on to the next stage
of communication, whatever that might be.
However, if the agent omits some necessary portion of the message, the WAIT
sign does not light; instead, the REENTER light comes on immediately and
thus the agent knows that he has erred.
A RE-ENTER light following a WAIT
light is indication to the agent that one of
several conditions has occurred; the system has detected a parity error at some
stage of the process which cannot cause
erroneous adj ustment of inventory, or the
agent has attempted some nonvalid operation such as entering the date JUNE 31,
FEBRUARY 30, entering a date November 21 while using a film slide for a flight
schedule which expires November 15, etc.
An ERROR sign will be lighted only when
the following conditions coincide: the transaction involved is one which is intended to
adjust inventory and the message has
reached the computer; the reply reaching
the programmer scanner contains either a
parity error or an insufficient number of
characters, and this is the second erroneous
reply received at the programmer scanner
(it having requested a repeat when it received the first such reply).
Communications circuits provided by
the communications companies will be
equipped with alarms and signals which will
provide information regarding cetrain types
of circuit failure. These alarms and signals
will be available to the agent or, in the case
of large installations, to the supervisory
personnel.
Question: In the midst of all this wonderful development, is there some way to
get Eastern and other airlines to answer
their telephone when one calls for a reservation?
Answer: We believe that he development and application of equipment here
under discussion is one way to come more
near the fulfillment of the objective of the
management and employees of Eastern;
that is, to serve our actual and potential
customers to their complete satisfaction.
I t may be of interest to you to know that
this objective is continually presented to
all employees because every paycheck has
clearly printed on it, in very large red
letters, "The customer pays our wages."
Question: When a cancellation is received on a flight which had been filled,
what is done to insure that the seat is first
made available to those on the waiting list
rather than to the next agent who happens
to request the fight?
Answer: This question, and the answer,
exemplify one of the reasons we chos~ to
employ a general purpose computer. The
computer will store not only an inventory
of seats, it will also store the number of
unfilled requests for seats, i.e., the waiting
list. As each cancellation is received, the
computer determines whether there are any
unfilled requests; if there are unfilled requests, the cancelled space is reported out
immediately on a punched card which is
then routed to the position where the waiting list is held.

Question: Is there some provision for
using the file computer on other problems
during the night or at times when the
computer is not ordinarily busy?
Answer: At the present time we are
not planning such use during the busier
hours of the day because we have yet to
learn how much spare time will be available. However, assuming available computer time, there is no reason why such
use should not be made. The equipment is
capable of such operation. We plan to use
it at night in slack hours to develop statistical data from the inventory during the
same hours when it is still serving agent
sets.
Question: If a flight leg can occur on
several different routings, does the agent
have to check several plates?
Is not the agent set itself able to store
information received? Also information requested?
Answer: Some definition of terms is
necessary to avoid misunderstanding. A
"leg" is -defined as that portion of a flight
from take-off to the next point of landing.
A "segment" is defined as that portion of a
flight which lies between the boarding point
and the deplanting point of a particular
passenger. A "flight" is the operation of an
aircraft from the point where it is identified by a particular flight number to the
point where it is no longer identified by
that same number. Within these definitions,
the question is best answered by stating
that the agent never has to check more
than one film slide to obtain the desired
information on any flight leg or on any
flight segment. If a passenger's itinerary
involves more than one flight, it may be
necessary to refer to a maximum of one
slide per flight.
So long as the keys on an agent set
are depressed, the information so represented is available to the programmer
scanner. And, of course, the film slides
store a great quantity of information.
When an answer has been displayed on an
agent set, that answer is continuously available until the agent "clears" the transaction
key or operates the lever to restore the
film slide in the magazine.
Question: How long does it take for
the two-character start pattern to reach a
remote station?
And, can a busy remote station's agent
sets be served more than once in a given
recognition cycle?
Answer: The time will be dependent
upon .the traffic situation at any given time.
Under a "no traffic" condition each remote
station will be polled once every five seconds, making the average access time two
and one-half seconds. This answer refers
to the 83-B-l ten-character per second system, with twelve remote stations.
The agent sets at a given remote station can be served only once during a
given "recognition cycle." When more than
one set is served on the same poll, the average access time is correspondingly reduced.
Question: Is the information from the
local agent sets transferred to UFC through
the HSPS and TO adapt in a serial or
parallel mode?

And, where are the messages from the
remote agent sets buffered? What is the capacity of the buffers and how many buffers
are there?
Answer: In serial.
For the purpose of this question, the
remote agent sets should be considered to
be grouped so that all agent sets served by
particular input-output adapter are in a
common group. Within such a group, agent
sets are served sequentially; thus, each
agent set stores its own message until it
is served. Each input-output adapter stores
the message it is then handling on to a
portion of one track of the high-speed
drum in the UFC. When the UFC tests
the input-output adapter and receives a
reply that there is a message waiting, the
UFC processes the message, stores the
reply in an unused portion of the same
track, instructs the input-output adapter to
take the reply and begins testing other
units to see if there is work waiting to be
processed. At the outset, we will be using
two input-output adapters (therefore, two
buffers) for agent set work. Other tracks
on the high-speed drum will act as inputoutput buffers for the electric typewriters
and the read-punch unit.
Question: Is there any provision for
expanding the system to handle reservation
names as input-printed output-ticketing,
et cetera?
And, how many agent sets, total, can
one system handle and still provide satisfactory service?
Answer: Initially, we plan to have the
read-punch unit produce punched cards containing the reservation details. These cards
will be matched with similar cards first
written by the agent and then keypunched.
This process will detect errors of many
kinds. Although there has been considerable
reflection and discussion regarding extending into the ticketing area, it would be an
exaggeration to say that such provision
has been made.
The answer is entirely dependent upon
the average handling time of the transactions at each of their various stages, and
the speed of service which is considered
"satisfactory." At the present time there
is not sufficient information available regarding the "mix" (relative quantities of
various types of transactions), nor is there
available information of sufficient accuracy
regarding the handling times. Because of
this uncertainty, we have (we believe)
deliberately underloaded the system for our
initial application in order that we may
safely collect the information necessary to
plan intelligent expansion.
Question: What developments are there
for storing passenger identification for preparing lists to be checked against later
ticket purchase?
Answer: The answer to a previous
question partially answers this one; that
is, the comparison of computer produced
punched cards with the manually written
and keypunched cards. Ticket purchases
will be related to the keypunched cards
(now audited for comparison with inventory) for verification. When reservations
are made at the time of ticket purchase,
the computer produced card (produced

Payne: Stock Transaction Records on the Datatron 205
when the ticketing agent operated the agent
set) will be held for comparison with a
keypunched card produced following a
telephone call from the ticket agent to a
voice recorder at the reservations office.
Question: What computer do you use?
How long has it been used in this operation?
Answer: We plan to use the Remington
Rand Univac, Modell File Computer, and
at this date (March 10) the program debugging has progressed to the point where
it is about 95 per cent complete. Agent set
sales have been entered through prototype
equipment into the particular Model 1 File
Computer which will be installed in our
N ew York Office this summer. All the
transactions mentioned have been tested and
have produced the results anticipated and
related here.
Question: Is this system operational?
If so, what is the operating reliability experience?
Answer: See answer immediately preceding. We are asking for a reliability
exceeding 99.7 per cent of the scheduled
operating time.
Question: \iVhat is the present state of
the development of the system? What, if
any, portions are in operation?
Answer: See answers immediately preceding.
Question: What steps do you take if
there is an interruption between the agent's
set and remote location?
Also, is the information contained by
the central computer retained elsewhere for
reference in the event of system failure?

Answer: The action to be taken will
depend to some degree upon the particular
circumstances, such as the duration of the
interruption, the proximity of other agent
sets and their operating condition, the nature of the interruption, etc. In general, a
location dependent upon agent set service
will use other communications facilities
such as telephone and/or teletype when the
agent set service is not available.
Each time that an agent set operation
results in adjustment of the inventory, the
computer produces a punched card record
of the transaction (as stated in answer to
previous question). This punched card record also contains a record of the entire
inventory for the flight as it exists after
the adjustment which produced the record.
It is planned to process these records at
fifteen-minute intervals on conventional
punched card processing equipment so that
it will be possible to revert to a completely
manual system within a few minutes if
necessary. This protection will be removed
when experience proves that it is no longer
necessary.
Question: What were some of the considerations that led you to choose a general purpose computer rather than a special
purpose computer?
Answer: The answers to previous questions have partially answered this one. Perhaps the most important single consideration was recognition of the fact that we
didn't really know in su/fi,cient detail just
what we wanted the computer to do. The
electronic data-processing science and our
need to apply it were developing at the

183

same time. We had to get a program under
way without being able to supply detailed
specifications. It was apparent that it would
take us several years to reach the point
where we would know what to specify; it
also was apparent that we would need the
equipment by that time. We, therefore, decided to select as the central data processor
a general purpose computer which would
meet the broad requirements of large capacity random-access storage and something faster than slow-speed processing.
(We would like to have had a much faster
data processor.) With this much behind us
we could concentrate on the detailed specifications for the equipment which at that
time appeared to be more special purpose'
i.e., the agent set and associated communi~
cations equipment. As it has turned out
the agent set and related equipment are ~
great deal more general purpose than any
of us anticipated. Within the next decade
there will be a tremendous increase in the
use of these devices.
Question: What is the required up time
per day for this system and what provisions are made for handling the load in
the event of a major electronic or mechanical failure?
Answer: The required up time is
twenty-two hours per day, seven days per
week. As previously indicated, punched
card records will be available to provide the
basis for reverting to a manual system of
control. The manual system would be comparable to our present system, modified
only to fit the new type of record and to
take advantage of then existing fa<::ilities.

Stock Transaction Records on the Datatron 205
A. H. PAYNEt

EVERAL users of stock market transaction records
maintain teams of ticker watchers to compile price
files. Six to eight such teams are kept by wire services and some newspapers. Each of the two major oddlot brokers and Teleregister Corporation, with its quotation
board service, are among others who monitor the ticker
visually. Not all of these cover more than one exchange
but some do.
The use of a computer to monitor the ticker automatically was accomplished by Melpar for Standard and Poor's
Corporation in order to compute their "Standard 500"
indexes. Fig. 1 shows schematically the flow of information from the standard Western Union 6-channel ticker
code through a converter to a punched paper tape which is
suitable for input to the Datatron 205 computer. Examples

t Melpar, Inc., Boston, Mass.

of the paper tape involved are shown in Fig. 2. In Fig. 3 is
shown the correspondence between ticker characters and
digit-pairs within the computer. The 4000 words of main
storage within the computer are used as diagrammed in
Fig. 4. Outputs from the computer are the hourly indexes,
daily small-group indexes, and price files for starting next
day.
OPERATING SPEED

At maximum ticker speed the computer is occupied
about 55 minutes of each hour, processing, computing indexes, printing rejected prices, and punching the price file
periodically for safety. The level of market activity varies
considerably and on less active days, the computer may be
free over half the time. Eventually this extra time may be
used for other jobs, but during our development phase we
have preferred to use it for increased reliability.

184

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

500 char/min

1
I

Model101.4
Reperf ora tor

I
I

TS·!;.

Western Union
ticker service
(6-channel code)

Ticker
Printer

S~··

lSU

t ts216

ttl

3s'" 41.. 21t' 1$.'1.5.1

2s2!1 2s1

Printed ticker

Used for
visual checking

1---1

6-channel punched paper tape

I (1 line/char)

............. ..... .......... ......

.... ..........
::: :::.: :: .........................
....... ~.! ... ~: •.... ~~.~ •••.......... ; .. ~ .. ~ .. ~;~~~~
. :.. ::. .:.: :. :..: . .:........
.. . . ...

ct.S

6-channel punched ticker

Reader

................•.............."
.•......... •.••....... ........
.,' ••... .....••• •
.....................•.•.••...•••...•....••.••••.•••...........
..
...
...
.
..
...
.....
.
•••
•••
•
••
•••••
•••
•••••
•
• • ••
•••• •
•
••••
•

--1--

.~..

Melpar
Converter

Programs and
Price Files

1
Punch

Ticker data converted to a 3-channel code and grouped in
Datatron words. The first character in each group is the sign
which is always positive. The three channels below the sprocket
channel are used with the following code:

. ...
. . .. .. .. .

1000 char/min

Standard Datatron 7-channel
tape with data on 3 channels
2 lines / char

Datatron 205 computer

to Standard & Poor's
via teletype

Fig. 2-Punched tape coding.

Extract transactions of those stocks on
Standard & Poor's list
Test for reasonable variation from last price
Print-out rejections
Update high, low, and last price file
Compute indexes hourly [industrials, rails, 1
utilities, composite)
Compute high and low indexes daily
Compute 97 group indexes daily

I Hourly and daily indexes

U pda ted price
\ files on paper
1 tape

Fig. 1.
RELIABILITY

Reliability in the system is attained by hardware duplication, to be discussed later, and by programmed checks.
Of the dozens of checking features incorporated into the
program, the most important is the check for price reasonableness. Each new price is compared with the last previous
price and, if the difference is more than approximately 2
per cent, the price is not immediately accepted but is
printed out along with the old price for visual inspection.
Such a "rejection" may be caused by:
1) An error on the ticker. In this case the price is
properly rejected and a correction will appear later.
2) An unusual configuration on the ticker which looks
like a transaction, for example, the report of a dividend in cents.
3) A legitimate price change greater than 2 per cent.
The new price is inserted as a special operation.
4) A price-file error. Same as 3) .
5) Malfunction of the tape-converter system. The converter is replaced by the spare.
This check therefore serves not only to minimize degra-

A
B
C
D
E
F
G
H
I

L
M
N

0
P

R
S

Upper Case
30
T
23
U
16
V
22
W
20
X
26
Y
13
Z
05
&
14
Pr
32
0
¢
36
11
07
06
03
15
35
12
24

01
34
17
31
27
25
21
33
04
10
02

0
1
2
3
4
5
6
7
8
9

£
$
-

Lower Case
5
42
8"
.:!
70
4
:I.
63
8
56
b
62
0
60
c
66
53
45
54
76
47
43
75

41
74
57
71

8"
1
"2

67
65
61
73
44
50

Fig. 3-Correspondence between printed characters
and computer digit pairs.

dation of accuracy of the answers, but as a diagnostic
feature. In this connection, it is desirable to distinguish
between a system error which results in degradation of
the answers and one which is incapacitating.
DEGRADATION

The computer almost never makes errors which have
the effect of degradation; if it is not functioning properly,
the program will quickly hang up. Therefore, the primary
source of degradation is faulty data input caused by converter malfunction. However, this effect is not serious for
several reasons: 1) the aforementioned rejection of gross
errors, 2) errors in the input data show up as rejections
and alert the operator to switch to the second converter,
3) with 500 stocks in the index an error of no more than
2 per cent in a few of them is not significant, and 4) there is

Payne: Stock Transaction Records on the Datatron 205

3999-

~~~~Il

l.ll~L.L

High

Last

Price file

I
I

Capitalization factors and
grouping information

Low

in dollars and eighths

3500j

I = Interval: maximum permissible time between the availability of raw data and the completion of its processing.
P=Processing time: time required for the computer to
process the data collected in interval I for which answers
are required at the end of interval I.
e = Change-over time: time required to switch from the given
job to another one, perform some useful work, and return
to the given job.
S=Sensitivity index: a measure of the time sensitivity of a
given job on a given computer:

p+e

3000-

S=--

I+C'

2500

Raw data storage

185

O~S~l

M = Mean error-free running time: mean time interval which
a system will operate without an error.
R = Reliability index: a measure of the reliability of a particular
system for a particular job.
M

R=----·
M+PS

2000-

i
I

Program

Fig. 5-Formulas.

1500-

Price file index
by stock symbol

1000-

0500

Program

say that the processing time cannot be greater than the
interval time.
The closer S approaches to unity, the more time-sensitive
the job is.
For special purpose systems, the change-over time may
be said to be prohibitively large, therefore S is close to unity
regardless of P and 1.
The relation
M

0000Fig. 4-Computer storage chart.

a chance that an error introduced during the hour will be
replaced by a legitimate transaction before the end of the
hour when the indexes are computed. (While true of last
price, this is not the case for high and low.)
Thus, an acceptable level of accuracy is maintained without operating the converters in parallel, but by merely
having a spare ready to insert when one shows signs of
failing. The possibility of an incapacitating failure has
led us to seek a measure of the reliability of a system in
terms of on-line or time-sensitive work.

R=--M+PS
implies that, even if M is small, if the term P S can be
made small in comparison to M, then a high reliability can
be attained. This assumes an efficient error detection and
recovery procedure so that available system time can be _
used to reprocess the data.
R may be regarded as the probability that the job will
get done.
If an R is computed for the separate components of a
system, then the over-all R is given by the product of R's
for the components.
If Rl and R2 are the reliability indexes of two parallel
systems performing the same job, then
Rtotal

TIME SENSITIVITY

There is a difference in the quality of urgency one feels
about the computation of a table of elliptic functions as
opposed to the computation of the trajectory of an enemy
missile, for example. But this difference is not solely one
of the existence or nonexistence of deadlines-the table
of elliptic functions may have a publication deadline. There
is a question of cost involved along with some subjective
factors which can probably be converted to dollars and
cents. We would like to have a formula, a technique of
obtaining a quantitative measure of the fitness of a particular system to perform a given job. Such a formula to be
comprehensive would be very complicated. The simple approximation of Fig. 5 illustrates what might be done.
To say that S must be less than or equal to unity is to

= Rl

+ R2 -

Rl R 2.

OPERATING EXPERIENCE

This operation was entered into by Melpar primarily as
a research project and for this reason we have had very
little "typical" operation as our control programs and procedures have evolved. One result of this is that the human
operators reach the point of boredom without having attained enough automatic proficiency to eliminate errors. In
fact, most errors in the system have been traceable either
to a human error or to faulty recovery from an equipment
error not very serious in itself, thereby reinforcing our
belief that human participation in such a system should
be kept to a minimum. However, we have devoted considerable effort toward making it difficult for an operator to make
an incapacitating error. Most such safety features take the

186

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

form of programming tricks peculiar to the Datatron computer.
A second result of the lack of "typical" operation is that
the figures tabulated in Fig. 6, continuing the reliability
measurement example, are based partly on extrapolations
and estimates. We have done worse in the past, especially
during the first few weeks of operation, but are doing better
now, and expect to do much better in the future than the
figures indicate.
COMMENTS ON FIG.

Question: Can you indicate specifically
the human links within the described system?
Answer: The converter output is a
punched tape which is wound and transported to the photoelectric reader of the
computer by hand. The indexes are transmitted from our Boston Laboratory to
Standard and Poor's offices in New York
via ordinary teletype which involves a
manual keyboard entry.

p
Reperforator
Converter
Human
Linkage
Computer
Composite

Single
Unit

1 1 1000 0.999
0.9 0.9 50 0.984
0.4 0.4 40
0.9 0.9 60

Cumulative R
per month

Dupli- Single Duplicated Unit cated

--------0.862 0.9997 0.089 0.956

0.548 0.996
0.987 0.9998 0.141 0.969
0.9664 0.9945 0.006 0.44

M
R=--M+PS

In estimating M, only those errors which resulted in
delivery of late or incorrect indexes as a computer output
have been counted. Accurate indexes can be extrapolated
from previous ones using a few market leaders as weighting factors. This provides a level of backup to our system.
The reperforator is a well-engineered heavy-duty device.
However, its large M is due in part to the fact that it is
buffered from the rest of the system by the converter, an
editing device.
In most cases of human error, the error was precipitated
by an equipment malfunction which might have been
overcome with little interference by perfect human response.
The figures used for human-linkage P and S contain
a larger portion of pure estimation than the others.
Most incapacitating converter errors arise from the fact
that an improperly edited data word may take the form of
a control word for the computer. Tape-reading control on
the Datatron 205 is exercised via special characters on the
tape. An additional checking feature is being built into
the converter to combat this.

Discussion

R per hour

Fig. 6-Reliability of components. The figures shown for composite
R may be interpreted as probabilities that there will be no delay
of answers or production of wrong answers during the indicated
period by the system described here. Another level of backup
for the "Standard 500" is provided by extrapolation and manual
techniques.

CONCLUSION

This application has demonstrated the feasibility of
processing stock market data as reported on the ticker tape.
I t has also afforded another example of the use of generalpurpose digital computers in on-line jobs. The sensitivity
and reliability indexes, while not precise measures nor
always applicable, are valuable devices for use in the analysis of system performance.
It is possible to translate computer operating speed into
reliability through its effect on sensitivity.
The best applications are those which have associated
with them enough off-line processing to support an additional computer which also provides backup to the on-line
system.

Question: Have you made any cost
evaluation between the described system
and all-manual operations? What is the
relative occurrence of error in the indexes
delivered to Standard and Poor's with the
computer system as with the previous
manual methods?
Answer: This operation has not been
done manually. Standard and Poor's did a
90-stock index before going to the present
operation. This has been a research project
for us and our conclusion is that the job
can be done far more economically and

accurately on a computer than by manual
methods.
Question: How do you check the accuracy of your price files?
Answer: There are several programmed
checks, such as the test for reasonableness
already mentioned, designed to preserve the
integrity of the price files. If the file is
read out of memory and then back in
again, the file is protected by the formation of a check sum. If the check sum fails
to check, a price reasonableness test is run
against an earlier file known to be accurate
in order to pick out which price is wrong.

PROCEEDINGS OF THE EASTERN COMPUTER CONFEREiVCE

187

A Small, Low-Cost Business Computer
ALEX B. CHURCHILLt

HE Monrobot IX is a desk-size electronic digital
computer expressly designed for on-line business
applications of those types which are basically repetitive. We include in this category such applications as
invoicing, prepayroll computation, and production planning. In all of these applications, an operator must be able
to receive problem solutions promptly after insertion of
data into the machine. The Monrobot IX produces printed
solutions to problems in a fraction of a second to a few
seconds, depending upon the particular application.
Fig. 1 is an over-all picture of the computer in its desk.
Input and output is by the electric typewriter, and the computer itself is entirely contained within the single pedestal
of the desk. The power required is less than 750 watts at
115 volts ac.

program key is labeled to indicate the type of invoice
line for which the machine has been programmed. For
example, whenever the operator totals out an invoice, she
depresses the proper key and the computer, causing the
word "Total" and its dollar amount to be automatically
printed out by the typewriter in the proper columns. Simultaneously, the accumulation to total accounts receivable is
made internally, and the register being used for subtotals is cleared in preparation for the next invoice.

·e
8

•

Fig. 2.

18.

'"/(II/IIlillg. "itli"g • "((ONnling • Jala proussing
CALCULA.TlNG MACHINE COMPANY. INC.
c..u Offic. • Onaae. New JerMJ'
DATE 1210211957
NO

•

KINGS FABRICS
•

•
•
•
•
•
•

SH'",O,

~~A~~~L R:

mA:~AO

•

130
60
25

3/8
7/12

ITEMS
ITEMS
ITEMS

1.15
12.50
89.75

Ixfl~

YO
OZ
EA

149.93
757.29
2243.75

SUB TOTAL
01 SCOUNT
SUB TOTAL
F TAX ON
. S TAX ON
POSTAGE AND INSURANCE
TOTAL

2094.51
1885.06

PTI"#l.ON
5
20
40

142.43
605.83
1346.25

2094.51
209.45
1885.06
125.67
56.55
22.38
2089.66

•
•
•
•
•

Fig..3.
Fig. 1.

All operator controls are at the typewriter keyboard.
Fig. 2 is a close-up of that keyboard and helps to illustrate
operator requirements as to training and ability. The typewriter itself is a completely standard machine. The nonstandard assembly located at the front of the typewriter
is the program-selection keyboard. In this particular illustration the computer is programmed for invoicing. Each
• Monroe Calculating Machine Co., Orange, N.J.

The capabilities of the machine can best be seen by looking at Fig. 3, which is a complete sample application. This
figure illustrates a completed invoice in which multiplication by two different fractions are involved, and taxes
are applied to two different subtotals. In this particular
example, the date and invoice number, including the alphabetic characters, are automatically typed as the result of
operator depression of the date and number-program key.
N arne and address are normally typed. The operator then
selects the proper program to extend quantity times unit
price less discount in which the fraction "eighths" occurs

188

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

in the quantity column~ She enters the value 130, manually
tabs, enters the numer:ator, 3, and the oblique dash, whereupon the computer causes the denominator "8" to be typed
and the carriage to be tabbed into the description column.
The start signal for the computer is obtained by operator
depression of the manual tab or oblique dash key of the
typewriter.
As soon as the operator completes the item description, she again manually tabs and the computer is ready
to accept entry of the unit price. It is not necessary for
the operator to align decimal points. As she makes her
entry of unit price, the accumulation of the partial product within the computer occurs as each digit key is depressed. The computer may be programmed equally well to
handle decimal or fractional parts of a cent in' the price
column. After completing the entry of unit price, the
operator again tabs, the identifying letters Y D are automatically typed, indicating that pricing was on a per yard
basis, and the product of quantity times unit price is
rounded off to the nearest whole cent, stored, and printed
out in the gross column. The decimal point is automatically
aligned. The typewriter tabs automatically and the machine
is ready to accept entry of the discount percentage. The
operator enters the discount value and tabs; the discount
is applied, the answer is rounded off, accumulated to the
subtotal and printed out. An experienced operator can
complete that line, including manual typing of the word
"Item," in nine seconds. The time required for computation of the net extension after the discount entry has been
made and accumulate it to the subtotal is less than sixtenths of a second.
The next two lines of this invoice illustration use the
same basic program with a few modifications. All that the
operator has to do for the remainder of this invoice is to
make the proper selections of programs in sequence, and
at the appropriate times enter the number ten to effect a
discount of 10 per cent on the subtotal, the 6 and 3 per
cent tax rates, and the dollar amount for postage and insurance. All other information, both numeric and alphabetic, is automatically printed out by the computer. If the
state or federal tax is a constant percentage, then it too
could be automatically typed and computed.
Not indicated in the illustration is the fact that accumulations are being made of total sales, discounts, federal
taxes, state taxes, postage and accounts receivable, all of
which may be printed out whenever desired by selecting
the appropriate program. Any other desired accumulation
can be programmed.
A good operator can complete this entire invoice as
shown, excluding the date and number line and typing of
the name and address, in less than 78 seconds. In a competetive run between the Monrobot IX and an experienced
desk-calculator operator, the computer cut almost 70 per
cent off the time required by the desk-calculator-typewriter
combination to perform the identical job.

It has been found that very little time is required to
train an operator. Within one hour a Monrobot IX operator can outproduce a skilled typist and desk-calculator
operator team. Her speed and accuracy will continue to
rise and reach peak performance in less than a week.
This machine may readily be programmed for virtually
any invoicing application, including step-rate utility billing
and tax billing.
In the case of utility billing, the only operator entries
required are the two meter readings. The quantity being
billed and the dollar amount of the billing are both computed and printed within five seconds in the case of a rate
structure having three steps.
We have said something about the field of application
of the Monrobot IX. We would like to point out some
of the features of the computer system design. Fig. 4
shows a block diagram of the computer and indicates the
control and information paths. Program control is achieved
by means of stepping switches in conjunction with a plug
board. The computer is capable of the four common
arithmetic operations, decimal shift right and decimal
shift left. Other commands exist for automatic typewriter
control and alphabetic printout.

\r---...L-----,
MULT I PL I CAND I DIVISOR
REGISTER

CONTROL
DATA FLOW

.-----------.-..- - - - -

Fig. 4--Monrobot IX system.

Word size is equivalent to 18 decimal digits. Information
is coded in straight binary form. Storage registers can be
split in any desired manner by proper programming; thus,
for example, for some applications the machine can be
considered as having 42 six-digit registers, or 28 ninedigit registers.
Fig. 5 is an over-all view of the completed computer less
typewriter and program-selection keyboard. The magnetic
drum, which can be seen at the front of the assembly,
rotates at a modest 2500 rpm. The one information track
and the three clock tracks occupy less than one third of
the drum surface. The extra width is unused.
The electronic unit is shown expanded as though for
servicing. The main circuit section, which is visible at
the center of the illustration, and the programming section,

Churchill: A Small, Low-Cost Business Computer

Fig.S.

Fig. 6.

visible at the rear, folds into the frame to make a compact unit. The machine is divided into four basic subassemblies to simplify construction and repair.
Our application of printed-wiring techniques is shown
in Fig. 6. Each tube board contains three flip-flops, three
inverters, and their associated circuits, although other
combinations are possible without modification of the
basic printed wiring. Each diode card can readily accommodate 60 diodes. There are 9 printed tube circuit boards and
15 diode boards in this machine.
Monrobot IX uses approximately 1000 diodes and 71
tubes, of which 23 are flip-flops and 24 are inverters.
Minimization was of tubes rather than diodes since we
are able to use a type of diode costing 23 cents each. Logical levels are plus and minus 3 volts.
A four-stage counter, not shown in the block diagram,

189

serves as buffer storage between typewriter and computer
and as storage location-selection control. Multiplication
and division are by repetitive addition or subtraction.
Conversion of a number from pure binary to decimal
form for read-out is achieved by dividing that number by
the appropriate power of ten. A count of the successful
subtractions before the remainder goes negative yields the
desired decimal digit. The next lower order decimal digit
is obtained by decimally shifting the remainder and repeating the iterative subtraction.
The computer is fast enough to be able to read a number
out to the typewriter at the rate of twelve characters per
second, which corresponds to the maximum rated speed of
the typewriter.
We mentioned that there was only one information track
on the magnetic drum. The two fast access loops, the
product/dividend register and the multiplicand/divisor
register, are interlaced together with the storage registers
in such a manner that only one record circuit and one
playback circuit are required for the handling of all information. Fast access loops are regenerated continually,
whereas storage registers remain untouched except on the
occurrence of a store command. Pulse density on the
information track is approximately 75 bits per inch, and
pulse-repetition rate is about 80 kc.
Negative numbers are not encountered in this machine
because the subtract operation has been modified to what
has been called the diminish operation. The result of this
operation is zero whenever the subtrahend is greater than
the minuend. Under any other conditions the operation is
a normal subtraction. This feature is particularly useful
in handling such problems as step-rate billing and payroll
computation in that it eliminates the need for branch programs. The diminish operation and its field of application
has recently appeared in the literature. 1
To summarize, the Monrobot IX is an on-line business
machine that is well suited to several basic business functions in which format and computation are repetitive; for
example, invoicing. The machine is sufficiently versatile to
be able to compute answers involving fractions such as are
encountered in lumber billing. The machine can be applied to any currency in the world, including that of the
British Sterling. Problem solutions are printed within a
fraction of a second to a few seconds after entry of input
data, depending upon the particular application.
Training time for an operator is virtually negligible provided the operator commences training with the ability to
type.
One of the chief advantages of this machine is to be
found in the form of a by-product, that is, accumulations
of group totals, such as total accounts receivable, total
federal taxes, total quantities, and so forth, which are
readily available simply by the push of a button.
1 R. W. Murphy, "A positive integer arithmetic for data processing." IBM 1. Res. and Dev.; April, 1957.

190

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE
Discussion

The answers were provided by W. Burkhart, Chief of Electronics Division, Monroe
Calculating Machine Co., Inc., Orange, N.J.
Question: While entering unit price,
what does the operator do if she makes an
error?
Mr. Burkhart: The method of correcting errors depends on the particular program used. At worst, the error will have
affected daily balances, in which an error
program would correct it.
Question: What is the memory capacity of your computer?
Mr. Burkhart: The memory capacity
is fourteen registers of eighteen decimal
digits each. Through programming, these
may be split. The average arithmetic speed

for multiplication and division is 1.6 seconds for a five-digit multiplier or quotient.
Actual addition time is 2.5 milliseconds,
while average access time is 10 milliseconds.
The actual speed generally is governed by
the speed of input and output on the electric typewriter.
Question: How many program steps
are available? What arithmetic operations
can be performed? Can the machine be
used for scientific use?
Mr. Burkhart: Fifty-two program steps
are available on each program. Eight basic
programs are available, and through the
use of extra program selection keys, as
many as thirty-two programs are possible.
The machine can be used for scientific use
if the programs involved are within the
capacity of the machine.

Question: With so many VT in a confined area, what means is used to control
temperature?
Mr. Burkhart: There are only seventyeight tubes in the machine, and the total
power dissipation is about 600 watts, so
that a small fan suffices to keep the interior
cool. Since the vacuum tubes themselves are
near unwired portions of the printed circuits, there has been no component deterioration due to heat.
Question: What provisions are made by
circuit design or computer logic to prevent
errors in computation to be printed and
invalidate correct work already processed?
Mr. Burkhart: There is no means for
preventing the printing of errors due to
operator mistake or computation error.

A Self..Checking System for High-Speed Transmission
of Magnetic-Tape Digital Data
E.

T. CASEyt

EVERAL years ago it became evident that high-speed
communications facilities were required as part of
many future data-processing systems. The need had
appeared in the planning phases of several systems and
certainly, as the speeds of computer operation increased,
it would be desirable to centralize computing systems to take
maximum advantage of them. Situations indicating its
need are:
1) Real time surveillance and control systems of military significance.
2) Faster computers which are able to do the total
data processing for a large business so the data from
the many sources must be brought to the computer
site.
3) The necessity for prompt sending of data to a central
location to permit over-all control, even if the development of small internally programmed computers
permits many geographically separated computer
installations.
4) There is a gross difference between the new and the
old data-processing speeds.
Important practical considerations in the selection of a
data-transmission facility as part of a data-processing system are that the media to hold the transmitted and received
data, and its code and format must be compatible with the
t Remington Rand Univac, St. Paul, Minn.

rest of the data-processing system. Although an "on-line"
data code and format significantly different than that on the
input and output media are possible, the present state of the
art indicates that those of the data-processing media be retained with minimum alteration for the data transmission.
The quantity of data to be transmitted in future systems,
when the facility for rapid, accurate data transmission is
widely available, naturally is unknown now. The situation
may be compared to that in earth moving. The number
of millions of yards of dirt that needed to be moved when
shovels and wheelbarrows were the only tools was much
different than the number that "needs to be moved" today,
now that huge power shovels, long conveyor belts, and
large tractor-scrapers are available. Even today, data
transfer by some concerns involves 5 to 30 million characters per day; while this accomplishment with techniques
used may be likened to the tour de force of the Egyptian
in pyramid building, it does indicate that ten to one hundred
times as much is not out of line for future needs.
The "state of the art" digital transmission speeds may
be compared with theoretically possible data rates by noting that a fairly low quality phone channel of 1700 cps
bandwidth and 22-db signal-to-noise ratio should, by the
B = W log2 (1 + SIN) formula yield an error-free transmission rate of approximately 12,000 bits per second, but
present practicabilities offer more like 750 bits per second.
Compared to teletype and telegraph service of 30 informa-

Casey: A Self-Checking System for Transmission of Digital Data
tion bits per second in 60-wpm service, this is a significant
increase (25x). With a 7-bit code, 750 bits per second
could yield 80 to 100 characters per second. A 30 million
character-per-day transmission load, at 100 characters per
second, would need this kind of line for 300,000 seconds
per day, or three of these lines would be needed to transmit
this much information per day. A large industrial complex
may have only 3000 to 300,000 characters per day to send
from each location, so many lines could be involved but
for only a few hours per day.
One of the most important aspects of digital-data transmission is the accuracy-the accuracy needed by the user
and the accuracy given by the digital-data-transmission
service. The present indications concerning the phone-line
error characteristic is that on the average, between one
in 10,000 bits to one in 100,000 bits will be in error. These
data are derived mostly from teletype and telegraph experiences. The redundancy planned for digital-data communications and the extensive "intelligence" built into the terminal equipment will detect these errors and can ask for a
repeat of the message; so even if the error rate on the line
should remain as indicated, the error rate in the records
submitted to the user will be dramatically less. The checking features of the digital-data-transmission systems can
be evaluated thoroughly only after detailed analysis involving as yet unavailable detail error characteristics of the
line and modem equipment, or by months of actual "online" t~sting. Then, with good records kept 6f those errors'
found by other checking, as with a large computer, and
traced to the records submitted by the digital-data-transmission equipment to the data-processing center, the accuracy can be evaluated. To indicate in terms of presentday media, punched cards and magnetic tape, the accuracies
expected in checked digital data transmission, one hole
punched wrong in one of 80,000 eighty or ninety-column
cards and not caught by the checking facilities, would yield
a probability of error of one in 40,000,000. This quantity
corresponds to sending eight cards per minute, twentyfour hours per day for one week. This is about 1/1000 of
the rate indicated as the error rate on the line. Similarly,
for a magnetic tape-transmission system at ninety characters per second for 168 hours (one week), one bit in error
in the submitted record would indicate a probability of undetected error of one in 400,000,000 or 1/10,000 the rate
indicated above as the" on-line" error rate.
The MTM Transrecorder, which transmits data from
one magnetic tape via voice-band facilities to another, is an
example of a facility for providing high-speed, accurate,
automatic, digital-data transmission. The source media
can be any Univac tape from 200-foot perfect tape reels
to 1500 or 2400-foot reels with random bad areas or splices.
Data are recorded in C10 code, in blockette or high-speed
printer format on both source and receiving tapes, and so
is compatible with other system equipment.
The philosophy of error correction in the Transrecorder
is to introduce into the transmitted message sufficient re-

191

dundancy to detect errors and have the receiver check the
received message, and if it is erroneous, request a retransmission. This is accomplished by:
1) Dividing the total message of 100 to 2000 blocks
of information into submessages each consisting of
an integral number-1 to 16--of blocks of information. A block consists of 720 characters which in
the high-speed printer format, used with the Transrecorder, is divided into six blockettes of 120 characters each.
2) Sending one submessage from transmitter to receiver
and awaiting reply.
3) Checking the incoming data at the receiver during
submessage reception and checking the recorded data
corresponding to this submessage to insure that they
satisfy the proper checks, which are:
Character parity.
Correct character countjblockette.
Long parity check during reception.
Proper number of blockettes/submessage.
4) Receiver sending a reply to the transmitter indicating
that the last submessage was received and recorded
correctly, or that it is to be retransmitted.
5) Continue sending submessages until the end of the
total message.
This is error correction by error detection at the receiver
and then use of feedback from receiver to transmitter to
advise the transmitter that the data have either been received
satisfactorily, or was received with errors and is to be
retransmitted. It is to be differentiated from error correction by error correcting codes as discussed by Shannon,
Fano, and others, which would permit reconstruction of the
information from mutilated incoming signals without the
necessity for reverse transmission.
From the time the operators initiate transmission until
end of rewind, at both transmitting and receiving locations,
the operation is automatic.
Some of the automatic features of the Transrecorder
are as follows:
1) After mounting the reels of tape on the tape handlers
and initiating operation, the control automatically advances both source and receiving reels far enough to
insure that good magnetic tape will be under the
heads. The receiving station simultaneously erases
the leader and tape which feature permits reuse of
tapes without requiring pre-erasure. The transmitting
tape advances ten feet prior to initiating "read" to
avoid noise due to clips and leader-tape junctions,
and the receiving advances fifteen feet before recording to permit adequate tolerances between the
various tape handler-control combinations of a dataprocessing system.
2) The data read from the tape at the transmitting station are checked for character parity and for 120
characters per blockette. This data is stored in a 120character magnetic core buffer.

192

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

3) In transmitting the data from the buffer to the
7) The "end of message" indication on the transmitting
tape is five feet of erased tape beyond the last
modem and line, the character parity is again checked
and a "horizontal parity" character, the 121st charblockette of the message. The transmitting ~ontrol
acter sent for each blockette of information, is genunit senses this and after insuring that the last
erated to indicate the sum modulo 2 of the "ones"
information sent was received and recorded correctly
for each information level.
at the receiving site, it sends an "end of message"
4) The control unit sends before each submessage, a
signal. Then the receiving and transmitting MTU's
submessage preamble consisting of a train of alterrewind the tapes onto the original reels and indicate
nate one-zero pulses. This alerts the receiver that
the completed message condition to the attendant.
a submessage is to be sent, prepares the phone line
The Transrecorder consists of basically three units, a
for propagation in the transmitter to receiver direc- Control Unit (CU), a Magnetic Tape Unit (MTU), and a
tion, i.e., reverses possible echo suppressors and modulator-demodulator (Modem) to take the "squarestabilizes compandors enroute, and establishes proper wave" voltages from the CU and apply signals to the line
receiving-clock phasing. This receiving clock estab- at the transmitting site, and to receive the signals from the
line and change to "digital signal" form at the receiving end.
lishes what we might call "bit synchronization."
5) The preamble is followed by one or more special The Univac Modem was designed for use on customer
characters which effect a "character synchroniza- owned lines and to permit gaining experience and insuring
tion," which lets the receiving control unit know compatibility between the facility "Modem" service and the
the first bit of each of the incoming serialized char- CU of the Transrecorder. When the Transrecorder is
acters. This special character both provides this used on public communications facilities, the facility's
"character synchronization" and a time-buffering modem will be in a separate cabinet to the left of the CU.
action which permits proper fitting of the tape-unit Any Transrecorder installation can be used to transmit
advance and bad-area traversal times to the required data or to receive by switch selection on the control panel
bit and character synchronization during the submes- at the top of the CU so an installation at a regional office or
sage. Each "on-line" blockette of 121 characters is remote factory can first submit a reel of tape data to the
preceded and succeeded by a special character to per- Data Center and later switch to the receive mode and remit the receiving station to effect a character count ceive data, instructions, shipping schedules, etc., from the
check on the incoming data. It also permits a long center.
With the exception of the Modem, the total CU is tranparity character to be generated from the first 120
characters of the incoming blockette. This is then sistorized. TheMTU involves both tube circuitry for tapetransport control and reading and writing and transistor
compared with the 121st incoming character.
6) At the end of each submessage or block group, which circuitry for logical control functions.
The maintenance requirements due to plug-ins and commay be chosen by a switch to be from about 5000
information bits to about 80,000 information bits ponents have been very low in installations using this
long, the transmitting station stops transmitting and type of construction; these plug-in circuit designs are used
awaits a reply from the receiving station. If all the also in the perforated paper tape to magnetic tape coninformation sent since the last "answer back" was verter, in the magnetic tape to perforated paper tape conreceived and recorded correctly, the receiving station verter, and in an aid converter and recorder. Provision
sends back a train of alternate 1-0 signals long is made for convenient preventative maintenance by alterenough to reverse the direction of propagation on ing supply voltages to selected racks of machine in the test
the line and to register the "resume" order in the mode and observing limits. Supply variations of -!- 25 per
transmitting site's control circuits. If errors were cent and more on entire racks are permissible when all
detected in the submessage as it was received, or if components are within limits, and the maintenance-panel
after three read tries the recorded submessage at indicators, the selective alteration of supply voltages, and
the receiving site fails to check properly, the re- the plug-in construction allows detailed analysis and
ceiving MTU repositions the receiving tape to the prompt correction of a fault if it occurs during scheduled
beginning of the submessage improperly received operation.
or recorded, and sends back a train of alternate
The Transrecorder operates on 115-volt 60-cps phase
2 "I" 's-2 "0" 'so This is sent long enough to stabilize power, the total average power requirements being less
the line transmission in the reverse direction and than 3 kw, and it is capable of operating in 90°F ambient
register the "retransmit" signal in the control cir- temperatures.
The author wishes to credit C. W. Fritze, B. L. Meyer,
cuitry. If the answer back was "resume," the transand
R. Goossens with the majority of the control and logimitter proceeds to send the next submessage, if "recal
features
of the Transrecorder and the continuing effort
transmit," it repositions to the beginning of the last
to bring this development to its present stage.
submessage and proceeds to retransmit it.

Casey: A Self-Checking System for Transmission of Digital Data
Discussion
Question: What percentage of the total
number of bits transmitted are redundancy
bits for error detection?
Answer: In the Univac data automation system, each character is composed of
six information bits and one odd parity
bit. This error checking feature is retained
in the high-speed digital data transmission
system where it is often referred to as the
"vertical parity" check. Hence in each
block of 720 characters, 720 of the 720 x 7
bits are "vertical parity" redundancy bits.
To each blockette of information, an additional "horizontal parity" character is
added, giving 42 additional bits per block.
Hence, about 15 per cent of the transmission consists of error checking bits. In
addition, timing bits and special spacing
characters introduced for purposes other
than error checking are checked against
the a priori knowledge that they should be
present at particular intervals, and so also
serve as "error checking" bits. But, this
latter feature is a sort of bonus, since their
primary purposes are for timing and interblockette and interblock spacing.
Question: Will the equipment handle
both metal and plastic tape?
Answer: Yes.
Question: Is the equipment now available?
Answer: The equipment is undergoing
laboratory and system testing and is not
available for immediate delivery. For delivery and similar information, please contact the Communication Department, Remington Rand, 315 Fourth Ave., New York,
N.Y.
Question: Will the equipment transmit
data over standard telephone lines?
Answer: Yes, and this was an important consideration in the design. The goal
was that any phone line over which satisfactory voice communication could be obtained should be suitable for digital-data
transmission.
Question: Can the blockette generally
be arranged by computers of other manufacturers?
Answer: This question has several
ramifications. The logical structure of a
blockette, i.e., 120 characters per group,
odd parity characters, particular bits significance in each character, etc., could be
prepared by any computer. The problem
would arise in the magnetic head structure, writing densities, writing mmf, readback signals, track orientations with respect to tape edge, head gap staggering
for various tracks, etc. Hence for most
practical purposes, the Transrecorder could
be expected to accept only tapes prepared
on Univac equipment.
Question: What is the form of the
transmission on the line, the signal representations of the "0" 'sand "I" 's?
Answer: This probably will vary with
each different manufacturer's Modem, or
each different communication company.
The Univac Modem uses 100 per cent amplitude modulation of a tone carrier of
about 1500 cps, "1"'s represented as full

amplitude and "0" 's as zero amplitude
signal. The signal levels on the line can be
varied but are generally considered as
about dbm into a 600-ohm balanced phone
line. A binary or two-state FM system is
being developed by others, with one tone
near the lower edge of the phone channel
pass band for one digital state and a tone
near the upper edge of the band for the
other; the tones are put on the line one
at a time. The binary FM appears to have
some advantages in increased SIN, for
ease of implementation of gain control,
and for bit detection implementation, since
a comparison should be more reliably made
between the power in the two tones on a
variable loss link than can the determination of whether the incoming power level
on such a line exceeds a preset level. Other
systems use mutiple tones, or different
phases of a "continuous" tone, all presently
known ones having the tones in the audio
phone band at the modulator output and at
the demodulator input. In the normal
trunking facilities where a given channel
may be subj ected to frequency or time
multiplexing techniques, the power spectrum and type of modulation may be very
different at enroute points than at the Modem terminals. It should be mentioned that
the Transrecorder (less Univac Modem
which is furnished for use only where a
communications company facility, with its
own Modems, is not available) is not interested in the "on-line" signal representation, if the Transrecorder-Modem interconnection signals are appropriate.
Question: What reasons led to the selection of a serial rather than parallel mode
of line transmission of the bits comprising
the characters?
Answer: Primarily, this decision was
based on ease and simplicity of instrumen-'
tation and the consequent economy. Also,
with the normal type frequency separation
techniques, the percentage loss of total useful bandwidth due to "guard bands" makes
for less efficient bandwidth utilization in
multiple tone systems, and the ratio of peak
power to average or rms power on the
channel increases when more than one tone
at a time is impressed on the channel. Since
generally, actual channels have a peak
power limitation as well as an rms power
limitation, the rms signal that can be impressed is higher for single tone modulation than for multiple tone. The development of new frequency separation techniques and the increase in duration of each
signal interval as the number of simultaneous bits per signal interval is increased,
which increase minimizes the effects of
certain kinds of noise, suggests a multiple
tone system. But, the economy of Modem
implementation seems still to be in favor
of serial transmission.
Question: On a low-grade phone circuit of about 12oo-cps bandwidth, did you
mean that theoretically this should handle
12,000 bits/second?
Answer: The example given was for a
1700-cps channel bandwidth having a 22-db
SIN ratio (S = 159.N and the transmission channel degraded only by additive

193

noise) and this facility then should, per
Shannon's formula B
W log2 (l + SIN),
give a "long time average" bit rate of in
excess of 12,000 bits/second. The actual accomplishment of this rate of transmission
on the above phone channel awaits the development of considerably more $ophisticated methods than we have at present.
Question: What is the transmission
medium employed?
Answer: When the Modem is furnished
by the communication facility, the medium
is of no interest to the Transrecorder
proper; but in normal installations it is
expected that it will consist of a 2 or 4
wire, one-half or full duplex phone channel facility. When the Univac Modem is
employed, a 2 or 4 wire, 600-ohm nominal
characteristic impedance, preferably balanced to ground, with inputs to the phone
line from the Modem of between +3 and
-6 dbm, and outputs from the line to the
Modem of between +3 dbm and approximately - 25 dbm, is required. After the 2
or 4 wire lines leave the vicinity of the
'Modem, especially if long-distance transmission is involved, it is expected that frequency or time multiplexing techniques
will be used, and the channel may go on
open wire lines, coaxial cables, microwave
links, or similar trunking facilities, but will
reappear on 2 or 4 wire lines in the vicinity of the Univac Modem at the remote
location.
Question: Are blocks and blockettes so
recorded on the sending tape or does the
control unit do the subdivision and control
the tape feed?
Answer: In the system implemented,
the information on the magnetic tape at the
sending end is divided into blocks and
blockettes (so-called high-speed printer format) and is reproduced in the same fashion
on the receiving end tape.
Question: What is the bit rate over the
line in the existing system?
Answer: The existing equipment is
working at 750 bits/second and at 800 bits/
second. The change from one to the other
involves a change in transmitting cIock
generator, and receiving clock recovery
circuit plug-ins. Since the maximum transmission speeds are so intimately associated
with the transmission channels, it is
planned that the transmitting Modem will
establish the bit rate by furnishing a
"transmitting clock" signal to the Transrecorder and the receiving Modem will recover a "receiving clock" from the incoming signal and supply it, with the received data, to the receiving Transrecorder.
As accurate transmission at higher speeds
is accomplished, due either to more sophisticated methods on given channels or the
installation of wider band facilities, the
Transrecorder then can very conveniently
utilize the higher speeds.
Question: Is it necessary to have two
sets of Modems on long-distance transmission, one furnished by Univac and the
other by the telephone company?
Answer: No. The Univac Modem is
a self-contained, panel-mounted unit easily
removed from the control unit in situa-

194

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

tions where the communications facility
furnishes the Modem and associated
"clocks." In the continental U.S. it is anticipated that the great majority of the installations will not have the Univac Modem.
Question: What is the information
transfer rate? How much time is required
to transmit, check and rewind, say, 100
blocks of 720 characters each?
Answer: This information transfer
rate depends on many factors among which
are "on-line" bit rate, tape leader length,
tape handler speeds, writing density, on
line error rate, number of blocks per block
group, number and length of bad spots,
"end of message" designation, line reversal
time, etc. An illustrative example representative within 20 per cent for all except
the most unusual cases would be that with
an 800 bits/second "on line" bit rate and a
fair quality line, hence few requirements
for retransmission. It would require about
7.8 seconds per block or 780 seconds plus
approximately 20 seconds rewind time
~ 800 seconds for 100 blocks or for about
500,000 bits of information.
Question: How long does the receiving
unit require to reread the blockette and
compare to stored information, and is the
data transmission stopped during this time?
Answer: The information is read and
checked at the receiving end in block
groups rather than per blockette, so all the
information sent during a block group is
checked in one continuous operation. The
information recorded on the tape is checked
against a priori known criteria rather than

against stored information. The check consists of insuring that each character of the
block group has satisfactory character parity, that there are 120 characters per blockette and 6 blockettes per block. The time
required to check a block group then depends on the number of blocks per block
group. As an example, if 4 blocks per
block group are chosen and a bit rate of
800 bits/second is used, the time to transmit and record the block group is about 32
seconds; whereas the time required to
check the block group recorded at the receiving end and reposition the tape would
be about 1.75 seconds, or about 5 per cent
of the time is used for checking. During
this checking time, transmission from the
transmitting end is stopped pending receipt
of a resume or retransmit signal from the
receiving end.
Question: Why do you use five feet of
blank tape to detect the end of message
rather than use a specific code?
Answer: With the exception of the odd
parity bit redundancy deliberately introduced for error detection, the Univac code
is a very low redundancy, or highly efficient, coding scheme, so it is not possible to
use a specific single character code to detect reliably a mark or signal of such important logical consequence as an "end of
message" signal. Even to limit detection of,
and action on, such a single character code
to the intervals such as end of a block of
information known a priori possibly to contain it is not sufficiently reliable for so important a logical operation. Hence, it would
require instead an entire blockette or block

of a very unusual code pattern to reliably
establish an "end of message" signal at the
transmiting tape, and then either its accurate transmission to the receiving end, or
the sending of a less redundant signal to
the receiver and the regeneration and recording there of a similar coded blockette
or block. Further, such implementation
would require considerably more instrumentation at both Transrecorders, as well
as imposing such an "end of message" coding on all source data devices. The relative case of implementation and reliability
of generation and detection of an "end of
message" indication with a short erased
section after the last useful data resulted
in this choice.
Question: To what extent does the time
required to reverse line echo suppressors
affect transmission time; for example, percentage increase per packet of data?
Answer: This relationship also is a
rather complex one in the general case.
The higher the bit rate and the fewer the
number of blocks per block group, the
larger is the percentage of the total time
assigned to echo suppressor stabilization
(line reversal). Also in the "answer back"
mode, two line reversal times per block
group are involved. If a one block per
block group mode is selected, and 800 bits/
second is the "on-line" data rate, approximately 8 seconds are required to transmit
the data. The maximum echo suppressor
stabilization may approach 0.3 second or
0.6 second for the block group cycle, so a
6 to 8 per cent time increase may be involved. At significantly higher bit rates
this effect would be more important.

Communication between Remotely Located
Digital Computers
G. F. GRONDINt

AND

INTRODUCTION

HE usefulness of complex data-processing centers
can be increased by rapid and accurate communica.,.
.
tion between remote locations. The problems encountered in the data transfer are not new to the communicator; however, the familiar characteristics of the
communications link assume increased significance when
the digital nature of the data, the high information rate and
the required degree of accuracy are considered. The stringent requirements demand that the communication system
place special emphasis on providing maximum utilization
t Collins Radio Co., Burbank, Calif.

F. P. FORBATHt

of channel capacity, on minimizing the raw error rate, and
on using special coding techniques to achieve unprecedented error detection.
The reliability achieved even by near-optimum communications systems falls short of the accuracy demanded. In
spite of the communication-link limitations, the desired
degree of accuracy is attainable by error-detection tech;..
niques and data repetition. The burden of error control as
well as the task of providing compatibility between the
various data sources and the transmission equipment falls
on special converters (input-output devices). Their design
is dominated as much by the inherent limitations and peculiarities of the communication system as by the characteristics of the data source. One such special converter

Grondin and Forbath: Communication between Remotely Located Digital Computers
intended for high-speed punched card transmission over
voice-quality circuits is described, and it illustrates how a
particular combination of parameters meets this specific
requirement.
GENERAL

Since common wire-line facilities represent a vast available network, economical data transmission depends on
efficient utilization of the voice channel. Unlike speech, the
inherent redundancy of digital data is extremely low and
a single error may cause misinterpretation. Therefore,
three important properties the transmission equipment
must have are 1) efficient utilization of bandwidth, 2) minimum binary error rate in presence of noise, and 3) low
undetected error probability. The first two are related to
the binary communication system while the third is
achieved by redundancy and coding techniques.
The system's basic error rate or susceptibility affects
the information rate. It determines the percentage of data
which needs to be retransmitted or corrected and the
amount of redundancy that must be added to detect erroneous data. Although theoretically any desired accuracy
can be attained, the complexity and cost of doing so are
directly related to this factor, and may be prohibitive.
Any error-detection method should meet system requirements with minimum redundancy, simplicity of coding, and
freedom from systematic errors.
KINEPLEX

A communication system known commercially as Kineplex, which uses "predicted wave" techniques, is particularly well suited to digital data transmission. Its theory of
operation and performance characteristics over radio circuits have been described in several papers. 1 - 3
. Kineplex lends itself to frequency, time and phase multiplexing for spectrum conservation; near zero crosstalk
between adjacent channels is effected by synchronous keying and sampling of infinite-:Q detection filters. The detection method provides perfect integration of the signal over
the pulse duration while noise which lacks phase coherence
is increased only on a rms basis. Phase-shift coding permits two independent bits of information to be encoded
on each pulse by resolution of phase into quadrature components. Thus, predicted wave detection yields a gain in
signal to noise ratio accompanied by a lowering of usable
signal threshold and a narrowing of the required bandwidth
WIRE-LINE ApPLICATION

The above techniques have been applied III the design
of the TE-206 Kineplex Data System (Fig. 1), a general1. M. L. Doelz, E. T. Heald, and D. L. Martin, "Binary data
transmission technique for linear systems," Proc. IRE, vol. 45,
pp. 656-661; May, 1957.
2 A. A. Collins and M. L. Doelz, "Predicted Wave Signalling,"
Collins Radio Co., Burbank, Calif.; June 22, 1955.
3 R. R. Mosier and R. G. Claybaugh, "Kineplex, a bandwidth
efficient binary transmission system," AlEE Trans., to be published.

195

purpose, high-speed binary data transmission system for
voice quality circuits. It features efficient bandwidth utilization, low susceptibility to noise, adaptability to use with
a wide variety of inputs, and parallel data transmission.
Its proven superior performance is derived from the phaseshift keying and the ideal detection techniques summarized
and referenced above.

TE-206

KINOE:T~EX
SYSTEM

CHANNELS

I
.......I11III....... :

I
I
I

Fig. 1.

Specifically, it accommodates 2400 bits per second within
a 2200-cyc1e minimum bandwidth. It provides eight parallel input channels and can therefore accept 8-bit characters
at a rate of 300 per second. Each of the four tones, spaced
440 cycles apart, carries information from two input
channels; the actual tone frequencies are determined by
the line characteristic. To accommodate a majority of
known facilities, tone frequencies of 935 cps, 1375 cps,
1815 cps and 2255 cps were selected for the TE-206. The
3.3-msec pulse length was selected to be several times
longer than the expected duration of impulse noise, longer
than the incremental delay distortion across the band of
unequalized voice circuits, and yet short enough to provide
frequency-error tolerance for carrier systems.
Since data can be handled in parallel by the transmission
channels, the necessity of parallel to series conversion is
avoided, and the cost and complexity of associated converters are reduced.
KINECARD (FIG.

The wide use of the punched card as a versatile and reliable source document has produced the need to duplicate
its information content at remote locations. The Kinecard
converter system permits continuous and accurate transmission of scientific and business data from punched cards
over common voice facilities. It illustrates how the various
design parameters can be combined to maximize performance within the bounds of economic feasibility.
Punched cards are processed at a nominal rate of 100
cards per minute. This makes possible on-line use of IBM
523 Gang Summary Punches for local reading and remote
punching of cards. Data are accepted from the card reader,
indexing markers and check characters are added, and the
information is presented as synchronous 8-bit characters
suitable for Kineplex transmission equipment; at the remote end, the data are stored until required by the punch,
its validity is checked, cards are punched and erroneous

196

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE
Transmission system detected
bi terror rate
Per cent cards in error
(off-set)
N umber of erroneous
cards undetected

10-3

10-6

60%

1 in 20

1 in 200,000

Fig. 3.

Fig. 2.

ones are offset. Operation is continuous without provision
for answer-back or automatic fills.
Other than establishing a communication link prior to a
card run, no special operational procedures are needed.
Interlocks prevent initiation of a run unless Kineplex,
Kinecard and IBM punch are ready. The converter controls parallel the punch controls and the system can be
operated from either.
DESIGN CONSIDERATIONS

It may be fairly stated that if cost and complexity are
not considered, just about any combination of operating
features may be provided. Features which were considered
in Kinecard were code translation, format control, card
verification, automatic error correction, and interchangeability of terminal devices. In its present form, Kinecard
is a special-purpose device having reasonable efficiency
and adequate error detection for wire-line applications.
The punched-card code contains twelve elements per
card column to accommodate about 50 alphabetical, numerical and special characters. The 12-bit coding could be
translated to a 6-bit code thus doubling the information
rate of the transmission system. However, most card readers present the data row by row, 80 bits at a time, such
that characters represented by each column can not be fully
interpreted until a whole card has been read and stored.
Transmitting the card as on a row-by-row basis eliminates
extensive storage and code-translating circuitry.
The card reading and punching operations are not veri-

fied even tho.ugh 2 or 3 machine errors per 10,000 cards are
possible: . Since these errors are not introduced nor aggravated by 'the communication equipment, their detection
should be by routine accounting-type cross checks.
The error-detection scheme takes into account the nature
of the noise over wire lines and the related error probabilities introduced by the Kineplex equipment in deriving
its phase reference. The impulse noise which may affect
all channels and the possibility of occurrence of adjacent
bit errors are countered by deriving two separate lateral
parity-check bits on each channel.
Assuming random-error distribution, the number of
erroneous cards, detected and undetected, is tabulated as
a function of system error probability (Fig. 3).
Operational tests are planned to determine the effectiveness of the error detection. If additional protection is required there is ample time between each card transmission
to add more check bits.
OPERATION

Reference to the transmitted card format (Fig. 4) will
help clarify operation of the converter.
Several control signals are derived from the card reader
to indicate the start of the card-reading cycle and to identify each row of information.
A reader-card start impulse initiates the emission of
several "start of card" characters which serve to index the
remote punch-control equipment.
As each row of information becomes available from the
reader it is transferred into eight 10-bit shift registers. A
row-start character precedes each row-transmission cycle
which consists of reading out all eight registers in parallel
with synchronous pulses derived from Kineplex. The register is emptied before the next row is presented by the card
reader.
At the end of the twelfth row the parity checks are inserted. Two parity-check characters are obtained from alternate data characters; two bits per row are derived. Each bit
is formed by adding the number of punches and complementing to an even multiple of two. Fig. 5 is a simplified
block diagram of the converter.
Since the reader undergoes speed variations, synchronization is achieved by inserting no-information characters
between rows and between cards as required.
At the receiving terminal, card- and row-start markers
are identified and they control the assembly of the incoming data into a magnetic core memory. The memory ca-

Joel: Communication Switching Systems as Real-Time Computers

197

Fig. 5.

The converter is completely transistorized including the
punch-magnet drivers. Construction is modular and consists of printed-circuit cards many of which are identical.
A complete transmit and receive terminal is contained in
a 5-,% foot cabinet.
CONCLUSIONS

Reliable communication between digital computers· and
on-line use of business machines utilizing wire-line and
radio facilities can be accomplished with· efficient datatransmission systems and special-purpose converters.
Fig. 4.
Economic considerations, clarifications of use requirements, lack of common language, format and data-rate
pacity is 72 characters. The punch operation is started standardization, incompatability of equipments and operawhen sufficient data has been stored in the memory to tional inexperience are some of the limiting factors.
assure that the punch will not overtake the incoming inforThe Kineplex data-transmission system provides a
mation even if it is running at its fastest tolerance of 107 common signaling method for use with a variety of existcards per minute.
ing and future input devices which want to take advantage
An SO-bit storage register assembles a full row from of available voice channels.
the memory. The punched card is reproduced row by row",,The Kinecard converter increases the speed of transIf the parity checks indicate an error, the punched card is mission of punched-card data to the usual operating speed
offset in the stacker.
of punching machines thereby permitting on-line use.

Communication Switching Systems as
Real-Time Computers
A. E. JOELt

INTRODUCTION-THE NATURE OF
COMMUNICATION SERVICE

AUTOMATIC communication switching systems were
the first practical and mass-produced data processing systems. Initially, they were designed with
electromechanical elements, such as selector switches and
relays. However, as will be covered in more detail in another paper,! electronic data processing techniques are
rapidly being applied to these systemsr

.fi.

t Bell Telephone Labs., Inc., Whippany, N.J.

R. W. Ketchledge, "An introduction to the Bell System's first
electronic switching office," this issue, p. 204.
1

Not only were automatic communication systems the
first mass-produced data processing systems, they were
also the first "real-time" computers. What makes them
"real-time" computers? They must serve on demand of the
customers "quickly when wanted" and "all of the time."
A service request by a communications customer is a perishable commodity and the longer the delay in serving, the
more likely the chance that the request will be withdrawn.
Therefore, in communication switching the time between
a service request and its fulfillment must be kept small,
if we are to provide the service for which we are granted
public franchise and to capitalize to the maximum extent
by making the supply of service approach the demand as

198

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

closely as possible. Public reaction to inability to serve
is an important factor in providing service when requested.
If service cannot be offered when requested, there is
either abandonme~t or delay. The amount of delay, in this
case, may determine whether the system is truly real time
or not. Real time is relative. A one-second delay on a call
which lasts for 180 seconds is not excessive. Perhaps d
10-second delay is excessive on a short distance call but
not on a longer distance call.
Equipment failures are inevitable in a system which
must serve "all the time." Even with only one source demanding service on request, duplicate or multiple facilities
are provided to greatly reduce the chance of expected failure to serve on demand. Improved maintenance facilities
and device reliability also reduce outage time of equipment
which has failed and thereby reduce the amount of multiple
facilities required to maintain service.
In the telephone business we also speak of service orders
when someone requests telephone service, that is, installation of an instrument. In so-called "common carrier" type
of communication system, the service must be offered to
all who' desire it. The ability to add new telephones on
demand imposes many problems on this real-time system,
since in "real time" there is a limit to the number of telephones and calls therefrom which a system may serve with
a given amount of equipment. Each new telephone served
will require additional equipment capacity in this real-time
system.
Therefore, we have many sources of demand. Each
source originates requests more or less at random and
independent of others, but over a long period of time patterns do exist which influence the engineering of these
systems. Another paper2 covers these traffic aspects in
considerably more detail. Each new source also becomes a
"sink" since communication traffic is usually two-way and,
therefore, facilities to interconnect all sources must be
provided. One of the differences between a communications
system and a computer using the same data processing
techniques is this necessity for interconnecting sources as
well as processing data for each. Many of the real-time
computers have a plurality of sources, but these sources
seldom need to be interconnected or interrelated directly
through the computer.
SIGNALING

In recent years we have tended to separate the switching
problem into two parts and deal with each separately. These
two parts are: I) the receipt and recording of service requests and the control of the establishment of the desired
connections, and 2) the arrangement of switching devices
to permit any desired pattern of interconnection. The first
part is a digital data processing problem. Thinking of a
communication switching system as a real-time system
brings traffic and other aspects which have long been with
2 J.
A. Bader, "Traffic aspects of communications switching
systems," this issue, p. 208.

us in this industry to the attention of those designing
complex digital computers:]:pe second part· of the system
is an inherent characteristic";b{ most communication switching systems and will not be dealt with here. 3 The analogy
between telephone switching systems and computers has
been covered previously4,5 and will be assumed as background for this paper.
The devices with which these problems were first solved
were several orders of magnitude slower than those presently being employed in digital computers or contemplated
for electronic switching. The earlier and most familiar dial
telephone systerr)$, such as the step-by-:step system, combine these two parts inexorably. Nevertheless, their study
in the light ofytlJe,pew electronic techniques teaches us
many things a~,~ilt' the nature of the real time problem.
For one thing, the speedo£ the devices in the central office
was fast enough to keep up with the maximum rate at
which human beings could spontaneously actuate call devices, such as dials and keysets, to place information
directly into the system.
Assuming ,each channel receives information from only
one person, the rate at which information enters the system is thus limited by his sending rate. Common carrier
systems usually involve very large numbers of sources;
there are at present some 62,000,000 telephones in the
United States. To select one of a large number in a single
operation would require a very large calling device, such
as a dial with 62,000,000 holes, and tremendous dexterity
on the part of the user. It would also require complex or
time-consuming signal generating and receiving equipment.
Considerable time might also be required to select the
desired telephone. A practical way to accomplish a selection among a large number is to use a sequence of digits
to sift through all the possibilities. This means that several digits must be sent and a compromise made as to the
number base of the system, the complexity of the calling
device for encoding the signal, and the number of digits
to be transmitted. It is well known that the base 10 has
become universal because it is most readily used by the
public and the number of digits, up to recently, have not
been excessive. The calling device is simple-a 10-hole
dial or 10-button keyset.
Here human engineering comes into the picture. Obviously, the base 10 is used because it is best known. However, one of the early methods was to use the first letters
of central office names arranged in base 10 as a mnemonic
aid. Recent psychological studies indicate that such an
aid may not be as useful and helpful as we once believed.
In any event, such aids should not be ignored since they
may provide better service by reducing dialing errors.
Reduction of dialing errors allows the equipment to operate
3 C. Y. Lee, "Analysis of switching networks," Bell Sys. Tech.
i., vol. 34, pp. 1287-1315; November, 1955.
4 W. D. Lewis, "Electronic computers and telephone switching,"
Bell Labs. Rec., vol. 32, pp. 321-325; September, 1954.
5 W. D. Lewis, "Electronic computers and telephone switching,"
PROC. IRE, vol. 41, pp. 1242-1244; October, 1953.

Joel: Communication Switching Systems as Real-Time Computers
more efficiently by eliminating waste usage. Also, customers dial more rapidly, thereby reducing holding time
of the call receiving circuit.
With the call information broken up into a number of
digits it is not always possible for the signal receiving
equipment in the switching machine to act on one digit at
a time. Most systems require that a number of digits, such
as the first 3 which represent the central office name, be
received, before any action may be taken. This introduces
the need for storage of digits in the central office until
sufficient digits of the number are available for processing.
These receiving and storage circuits must be provided in
sufficient quantity to care for the maximum number of
sources sending simultaneously. Interconnecting means
must be provided to associate these circuits with the calling lines. Circuits which are called "registers," "senders,"
or "directors" are used only during the period when the
customer is sending selection information into the machine.
These circuits may be dropped from the line after this
phase of each call to be reused by other customers. In this
way they are provided and used more efficiently than if
one stayed associated with each call until its completion.
But still one of these call receiving and storage circuits is
utilized for each call being dialed. Circuits of this type
may serve both operators and customers, or separate
groups may be provided for each class of input. This is
a designer's choice which is determined by the degree of
difference in the logic between a register used for operator
calls and one used by customers. For example, if operators
use 100button keysets instead of dials the difference is
sufficient to warrant a separate group of registers to work
with keysets rather than providing a single group in which
all registers are capable of recording either dial pulses or
keyset pUlses.
The holding time varies with the customer and the type
of call. Some customers take longer to start dialing. Dialing time varies. The number of expected digits may vary;
for example, local calls require 7 digits and long distance
calls 10 digits. (This is equivalent to variable word length
in computers.) The expected number of digits may be
made known to the receiving circuits in several ways:
1) by the coding of the number dialed assignments, 2) by
allowing time after each digit is dialed for another digit to
be started; if no new digit is started after 2 or 4 seconds,
it may be assumed that sufficient digits have been received,
or 3) by an end of dialing signal which eliminates the need
for timing for further digits and is most useful where
coding conflicts or large differences occur in the number
of expected digits.
To reduce the call receiving circuit holding time still
further and to improve service, overlap features are sometimes provided so that when the circuit has received sufficient information to interpret the general destination of
the call, this information is processed before awaiting the
receipt of the complete called number. For example, while
the fourth and fifth digits are being received the first three
are interpreted and a connection is established to the de-

199

sired office. While the sixth and seventh digits are being
received the information interpreted from the fourth and
fifth digits is being transmitted to the office of the called
number. By introducing overlap operation, a system is
better able to serve in real time with a minimum of delays.
It means that the receiving circuits must be more complex
and generally capable of sending as well as receiving.
Overlap may also be employed within the control so that
a single control unit may act simultaneously on different
phases of the different calls. Again this improvement in
real-time service requires more complex circuitry. In telephone system design this complexity must be matched
against the additional delays that might be encountered
without the feature or the cost of additional control circuits to provide equivalent capacity.
The logic of the call receiving circuits or program devised for interpreting call data must be capable of dealing
with much more than the normally expected call inputs of
7 or 10 digits. There are the types of service calls with
"0" or 3 digit codes. All possible codes are not used or
usable. If one of these were dialed, the system must be
capable of so informing the customer, and more important,
of disposing of the call to free the common circuits for use
in more productive work. The system must also handle
such nonproductive situations as when the handset is
knocked off the "hook" momentarily or even for extended
periods, or incomplete dialing, or not hanging up at the
end of a call. Finally, it must be prepared to receive a
hang-up, disconnect, or abandonment of the call at any
time and be ready to serve that input again as quickly as
possible thereafter. This is usually the customer's way of
indicating an error as well as a change of mind. In short,
the logic or program for the system must be prepared for
any eventuality and always have a plan of action. There
can be no dead ends or stops in logic circuits or program
operation if the next expected piece of information is not
forthcoming.
To accommodate these factors, a series of tones and announcements are used to inform the calling customer of
the status of his call. A dial tone, busy tone, and ringing
tone all let the customer know of the progress being made
to serve his request. They may also serve to notify the customer that the machine is ready to serve him, and thus discourage him from sending in information when the call
receiving equipment is not ready. Recorded announcements
are used to inform of expected duration of delays and for
intercepting on incorrectly dialed or unassigned central
office codes and telephone number series.
Most communication switching systems are designed
with call storage means provided on a traffic or whenneeded basis at the central office. Call storage can also be
provided at each source.
Certain tape teletype systems are of the type which include source storage. 6 But here traffic is delayed to insure
6 W. M. Bacon and G. A. Locke, "A full automatic private line
teletypewriter switching system," AlEE Trans., vol. 70, pt. 1, pp.
473-480; 1951.

200

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

a high occupancy of the communication channels. Telephone systems where the complete number is preset before
sending it are also of this type. 7
MULTIPLE INPUTS

The fact that communication switching systems serve a
plurality of inputs has been referred to several times in the
foregoing discussion. The traffic problems which this
raises are probably the most important real-time aspect of
communication switching systems. There are other factors
that influence design, which might be of interest in other
multiple input computers. Needless to say, when there is
a plurality of inputs to be served in real time, there may,
by necessity, be a plurality of control circuits provided to
keep up with the input information processing. (It is
possible where traffic is light or by introducing higher
speed control devices a single control circuit will suffice.)
Each input is two-way. It is designated by a number. Since
there are few, if any, restrictions on the requests from any
input with respect to what numbers may be called, the
control circuits must be capable of reaching any number.
The more inputs there are, the more digits must be received, the more complex the selecting functions each control circuit must perform, and the longer the time it may
take to do them. Furthermore, with more inputs, there will
be more control circuits and a greater interaction among
them. Naturally, this greater complexity increases the unit
cost.
In a single input system the call receiving portion is presumably always ready to serve. When the receiving circuits
are provided on a traffic basis in mUltiple input systems, it
is possible that one will not be available without some delay.
Premature dialing due to these short delays in assigning
call receiving equipment is one of the most difficult traffic
problems in real-time telephone systems. This is the doorstep of the system. Once inside, delays can be controlled
by providing adequate storage if the delays are tolerable
and the control is alert to abandonment of calls by the
customer. It is posiible to develop conditions known as
"snowballing" of troubles. Under these conditions, serious
overloads, or the "nervous breakdowns" of communication
switching systems that you may have heard about, can be
produced. When a call receiving circuit is available it may
be connected to an input which has already started to dial.
A wrong number or partially dialed call will result since the
system does not have the correct call information. This
means wasted usage of the limited call receiving circuits
when they are most needed and consequently system capacity is reduced during overload.
Other overload conditions peculiar to multi machine operation will be described in the next section.
There are other factors which must be considered with
multiple input systems. We have already mentioned some
7 W. A. Malthaner and H. E. Vaughan, "An experimental electronically controlled automatic switching system," Bell Sys. Tech.
J., vol. 31, pp. 443-468; May, 1952.

differences, for example, between operator and customer
inputs. There are many different types of customer inputs,
e.g., coin sources and business or residence sources with
flat or message rate. Furthermore, the customers may be
on single or multiparty lines, or have a subswitching office
such as a PBX. Each of these classes of input may require
some different treatment on calls originating from and
even terminating to these lines. These differences as well
as differences in signaling languages and the necessity of
anticipating abnormal actions make complex the logic of
the control circuits or the programs for these systems.
Each request must be interpreted in accordance with the
class of the input and the desired output. Some noncoin
customers, for example, may be able to reach numbers not
dialable by coin customers.
Despite all these differences in class of inputs, certain
standards, particularly electrical, are set so that the same
digital receiving equipment will operate with most classes
of inputs. Standards are set on signaling limits and also
for transmission so that the established communication
path, regardless of length, will be satisfactory. Furthermore, where there is a large number of sources it is important that these be standardized as much as possible to
insure low initial and repair costs. For this reason, it is
important that the many needs of the customer be satisfied
by combinations of standard instrumentation.
MULTIMACHINE OPERATION

It is well known that the automatic communication central office was the first successful large scale digital information processing machine, and that such machines are
interconnected over great distances and communicate with
one another. The network of these machines is the largest
and most extensive digital processing equipment ever to
be assembled or likely to be assembled. Machines therefore have inputs not only from customers but also from
other machines. There are a number of interesting engineering problems which arise in networks of such machines.
As was the case for inputs from customers, these inputs
also transmit digital information, 'and call receivers are
connected to them when digital information is sent for
processing. Since the interconnecting links are sometimes
quite long and costly, it is usually economical to use them
in two directions which means that a request for service
may originate at either end. Once the channel is seized at
one end, precautions must be taken to prevent seizure at
the other end and to connect a call receiver. Guard means
are also provided to prevent immediate reseizure of the
channel or to delay digital transmission when an interoffice channel is reused after release from a previous connection.
When signaling between machines in this real-time system, the information to be transmitted is usually available
by the time the path between offices is seized. It is desirable to devise signaling methods for this application which

Joel: Communication Switching Systems as Real-Time Computers
are faster than those serving customers, since only a small
part of the total time required by the customer to send
digital information to the machine is signaling time. As
the art has progressed, new and improved signaling techniques have been devised. Rather than thwart progress,
the new signaling methods have been adopted with new
systems. Therefore, when switching machines of different
vintages are placed together in a network they do not
necessarily have a common language. When it is determined to which machine a call is to be routed, the method
of signaling is also determined. Means must be provided
in the originating office to send at least one of the languages which the machine to which it connects is capable
of receiving. The smaller the number of languages which
an office must send or receive, the simpler its circuitry and
the lower its cost.
Signaling distances between machines are generally
greater than from the customer to the machine. Therefore, it is sometimes necessary to place intermediate signaling equipment in the intermachine paths to regenerate or
amplify the signals. Regeneration usually involves storage
which delays the retransmission of the information.
Also, when paths or connections are established between
machines, a new signaling problem develops. It is necessary to send signals over the connection in the direction
opposite to that of the original digital transmission to indicate .when the called party answers so that call changing
may start, and also when the called party disconnects to
indicate that the connection should be taken down. In general, these interoffice or intermachine signals are known as
"supervisory signals" and the planning for these signals is
as important as the planning for the digital transmission.
As a real-time problem it is even more critical, for the
longer the time intervals in the various unguarded periods
for seizure, release, and reseizure, the greater is the chance
of malfunctioning of the system.
In a complex network of offices or machines it may not
be economical to provide paths from every office to every
other office. For this reason it was realized early that intermediate switching machines should be established to
provide more efficient trunking. They could also be used
to regenerate the signals and translate the languages in
both directions.
Normal traffic between machines is sometimes handled
by means of alternate routing so that when all direct paths
are busy the originating machine reroutes the call to an
intermediate machine which also has paths to the desired
terminating machine. This means that both the numericals
of the called number and the desired central office code
must be sent ot the intermediate office. The signaling language between the originating and intermediate office may
be different from that which would have been used on a
direct connection to the called office. The control circuits
must take all this into account in processing the call without greatly increasing the holding time. Calls routed through
intermediate offices take longer to complete, but this

201

is minimized by employing overlap in the signaling.
Once intermediate offices are set up for traffic reasons
they can be used to concentrate other complex switching
functions which are required for only a small percentage
of the calls. Thus a hierarchy of information processing
machines has been established. More recently, automatic
long distance switching equipment has been placed directly
at the disposal of many customers and in the near future a
large percentage of the nation's telephones will be able to
call one another completely automatically by "direct distance dialing" through combinations of local and long
distance switching systems.
There are overload situations which occur where the
machines talk with one another. An example of this is on
calls coming into an office from other offices. The incoming call receiving circuits here seem to act fast since there
are usually no delays waiting for the calling customer. He
has already given all the required information by this time.
The holding time of these circuits is short; therefore, not
many are required. But at the called customer's number,
particularly a private branch exchange (PBX), a bottleneck may develop even though the PBX has sufficient
lines to handle most of the traffic. The more traffic to this
number, the more hunting required to find an idle path.
Hence, delays are encountered by the incoming call receiving equipment which causes delays in call sending
equipment in offices all over town that are trying to reach
this office. With the call sending circuits so tied up, they
cannot be used to complete calls to other offices. Again, an
overload reduces call carrying capacity when it is most
needed.
..
There are several ways of dealing with these overload
situations. In electronic switching systems now being devised1 where all lines are supervised on a time division
basis, it may be accomplished by abundant call storage since
such devices are relatively inexpensive. In electromechanical systems the overloads may be partially alleviated by
introducing a speed-up in the time allowed for certain realtime functions when the office or some part thereof is
working near maximum capacity. For example, the time
allowed to determine whether the call is partially dialed or
incomplete may be cut from 10 seconds to 2.5 seconds.
Another design feature is to eliminate certain safeguards
or trouble detection features which utilize control circuit
time on normal calls during these overload periods.
The transmission of digital information must be as accurate as possible. This is particularly true between machines as contrasted to human sources since they usually
have higher occupancy and transmit over greater distances.
Here it is not uncommon to insert in the digital signaling
paths compensatory electrical networks to insure that the
signaling paths are more nearly identical in electrical characteristics. Such networks are automatically inserted on
each call based on the call information. In a similar manner, transmission compensation including amplifiers may
be automatically added.

202

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE
SYSTEM CHANGES

Another important characteristic 6f communication
switching is that the system is dynamic, ever changing. As
the system is changed it must continue to function, in realtime. A communication switching system cannot be taken
out of service for maintenance, to add new service features, to add facilities to handle an increase in the number
of inputs served, or to obtain information for the benefit
of those administering the system.
When a general purpose computer is designed, compromises are usually made to produce a design which will
satisfy to a maximum degree the greatest number of potential customers. Thereafter, modifications may be made
for specific applications where the basic machine is not
satisfactory. In the application of communication switching systems the needs for each community to be served
are different. One machine design must fit the needs of a
heavy traffic area in a large metropolis where there are
many calls but with shorter holding times. This means
more control or digital processing equipment in proportion
to the switching network facilities. Another office might
be in a small isolated area with fewer long distance facilities and little call handling capacity. The machine design
may also be influenced by the method of charging for service, particularly the extent to which coin, message rate or
flat rate noncoin services are offered.
The central office switching machine must be designed
so that it can be manufactured, installed, maintained, and
expanded over its entire range of applications. There can
be little compromise between the machine's requirements
and those 'of the customers. We cannot, for example, reduce the amount of equipment by taking more time-adding another eight-hour shift. Furthermore, when we determine that additional equipment is required, we cannot
shut down the machine while it is being added. "Real
time" here also means "all of the time." The number of
combinations of the services available to customers and
administrative and service features available to the communications companies mapped into the characteristic
needs of each community gives a very large number of
design variables for which each installation must be manufactured and engineered.
Imagine adding an additional arithmetic unit to a computer while it is working. In most communication switching systems there are usually new additions periodically.
Engineering of each machine installation is continuous.
We install enough capacity to deal with the real-time needs
for a limited period, say one or two years. It is necessary
to obtain a compromise between the over-all investment in
idle equipment and the cost of engineering and installing
to suit each change in demand. This means that from the
start the equipment must be designed and the installations
planned for growth.

ing devices so that the traffic characteristics and the degree to which the equipment is handling the load offered
may be measured and compared, and engineering steps
may be taken to insure a satisfactory grade of service.
Since the system operates in real time these data· cannot
be recorded at any other time except when the calls are
taking place. The measurements must be made without
impairing the service.
The results of traffic measurements may show the need
to rearrange the available plant on a seasonal basis in addition to providing new equipment. Digital data processing of these traffic measurement records is helping the
traffic engineer to keep up with the real-time central office
machines by shortening the interval between measurement
and valuation.
In designing a complex digital data processing and
switching system it is difficult to predict by theory or analysis the way it will respond to the real-time input conditions. Therefore, it is not uncommon to set up existing
digital machines to simulate the functioning of the real
time machine under typical and overload conditions. In
this way, "real time" can be slowed up and studied microscopically. 8 With the advent of general purpose digital
computers some communication switching system simulation problems have been solved with these tools. The difficulty of programming these nonmathematical problems
and the frequent lack of sufficient memory capacity have
impeded the use of digital computers for this application.
SYSTEM MAINTENANCE AND RELIABILITY

Earlier it was said that in communication switching
systems real time also means all of the time. This means
that the detection, location, and repair of system faults
must go on with the system still giving service, preferably
without impairment. This means that the system must contain redundancy so that it is not dependent upon one and
only one of a certain element to handle the traffic. When
more than one element is required for traffic reasons, then
a failure of one merely reduces the capacity. By providing
more elements than required for normal traffic two objectives may be fulfilled. First, overload peaks are better accommodated. Second, taking an element out of service for
maintenance will not upset the normal traffic capacity.
Of course, once a system element fails and is removed
from service it is important that the trouble location and
repair time be kept small enough relative to the failure
rate so that the multiple failures occur infrequently and
are of short duration. To achieve this, the circuits are frequently designed with self-checking and even error-correcting features. Trouble recorders are provided to indicate the state of the circuits when error checks detect
troubles. In the more complex systems the trouble records
may be automatically analyzed to diagnose the trouble and
to indicate its location.

TRAFFIC MEASUREMENT

To carry out an intelligent and efficient engineering
program, the system must be fitted with built-in measur-

8 G. R. Frost, W. Keister, and A. E. Ritchie, "A throwdown
machine for telephone traffic studies," Bell Sys. Tech. !'J vol. 32,
pp. 292-359; March, 1953.

Joel: Communication SUJitching Systems as Real-Time Computers
Some types of circuits, particularly those associated with
signaling and transmission, cannot readily detect their own
troubles so that auxiliary routine test circuits are provided
which, once started, automatically connect to these circuits
one at a time and put them through their paces, usually
with marginally acceptable signals. Automatic recording
is also provided with such test circuits.
To avoid calls from being lost due to malfunctioning
equipment, second trial features are sometimes provided
which allow a different combination of system elements to
be employed on a call, if it does not progress satisfactorily
on a first attempt. Failure of a second trial usually results
in a tone connection which indicates to the customer to
try again.
The above is an example in a real-time system where
trouble not only cuts down the capacity of the system but,
to insure some service to each call, additional work load
is taken on when it can be least afforded. Features such as
this may be automatically circumvented in periods of overload, but are provided to avoid failure of calls to be completed in nonbusy periods. In this way the modern offices
with low failure rates may be left unattended bymaintenance forces during nonbusy periods with reasonable assurance of providing good service to all customers. Even
when the office is left unattended, remote alarm indications
are given at a manned control center so that steps may be
taken to repair serious troubles before they do affect traffic.
SYSTEM POWER

What has been said about service continuity in case of
equipment failure applies equally well to power equipment.
Communication switching systems in central locations normally use power that is available commercially. This power
source is backed up by local diesel generators which usually start automatically if commercial power fails. Being
a real-time system, a communication switching system
cannot have power off even during the short period required to start a diesel engine. If this happened, all the
communication paths established through the office would
be temporarily or permanently lost, depending upon
whether or not the memory associated with that part of
the office maintaining the switched paths was "volatile."

203

Such service would not be considered satisfactory, especially where connections would not be reestablished after
power was available from the diesel generator. Therefore,
power storage in the form of wet batteries has been an
essential part of a communication switching office. The
capacity provided in these batteries is sufficient to cover
the maximum expected period between the loss of commercial power and the full operation of the reserve power
source.
Other power considerations in these systems call for
dispersion of power so that trouble on anyone feeder will
not put the entire machine out of service. Such considerations apply also to other intramachine cabling.
OTHER FORMS OF SYSTEMS

Another form of communication switching system which
functions in real time is that used for network broadcasting of radio and television. 9 - 11 Here the real-time switching is cued to the actual time specified by the customer.
A "program" of such times and desired connections are
the inputs to the system. There can be no delay or information is lost. Perhaps with the increase in data transmission this type of "real" real-time system will find greater
application.
CONCLUSION

It has been shown that communication switching systems are a form of real-time digital computers. Some of
the engineering considerations in designing such systems
have been presented. These requirements are difficult to
meet because of their complexity, the dynamics of the
service demands, numbers of inputs and their characteristics, and the need to provide absolute continuity of service.
Despite this, communication switching systems have been
designed, built, and installed, and are operating as the
world's largest aggregation of real-time digital data processing machines.
9 C. A. Collins and L. H. Hofman, "Switching control at television operating centers," Bell Labs. Rec., vol. 35, pp. 10-14; January, 1957.
10 A. L. Stillwell and A. D. Fowler, "Switching at tv operating
centers" Bell Labs. Rec., vol. 34, pp. 366-369; October, 1956.
11 P: B. Murphy, "Program switching and pre-selection," Bell
Labs. Rec., vol. 20, pp. 142-148; February, 1942.

204

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

An Introduction to the Bell System's First
Electronic Switching Office
R. W. KETCHLEDGEt

FULLY electronic telephone central office is· being
developed for experimental Bell System service.
Both the electronic devices themselves and their
system organization represent major changes in the art
of telephone switching. The electronic switching system
consists of electronic voice frequency switches controlled
by electronic memory and logic. However, the system is
not designed on the basis of direct substitution of electronic circuits for corresponding relay circuits of an existing switching system. Rather, the electronic circuits are
organized in ways that exploit the advantages of the electronic technology.l
One of the obvious differences between the electronic
and the electromechanical switching technologies is simply
speed of operation. The electromechanical devices such as
relays operate in times measured in milliseconds, often
many milliseconds. The operate times of electronic devices
such as transistors are measured in microseconds, often
small fractions of a microsecond. Thus, the relative speeds
differ by a ratio of well over 1000. This dramatic difference in speed permits a given amount of electronic equipment to handle much more telephone traffic than a corresponding amount of relay equipment. The high-speed
operation also permits the system designer to organize the
electronic system quite differently and often much more
efficiently.
Comparison of existing types of electromechanical systems with the electronic system now under development
shows that in the electronic system the various equipment
units are much more specialized. This functional concentration or specialization of function is well illustrated by
the case of memory. In relay systems the memory function
'is widely dispersed throughout the system. For example,
in a modern relay switching system, memory functions are
performed in all of the various kinds of circuits. In the
electronic system much of the memory function is concentrated in two memories, one temporary and one permanent, which perform most of the memory functions for the
system.
In a similar manner most of the logical operations of
the electronic system are performed in a single functional
unit. Again this contrasts sharply with electromechanical
systems whose relay contacts are used for logical operations in all parts of the system.
t Bell Telephone Labs., Inc., Whippany, N.J,
A. E. Joel, "Electronics in telephone switching systems," Bell
Sys. Tech. !'J vol. 35, pp. 991-1018; September, 1956.
1

Functional concentration permits the electronic system
to be organized into a group of major components, each of
which performs some single major system function. These
equipment units can thus be designed to perform their
function very efficiently and, because of the high speeds,
perform it for the entire system. A further result is the
simplification of the interconnections between the various
equipment units. The number of wires connecting these
units together is measured in tens rather than the hundreds
or even thousands of wires often used in relay systems.
The electronic system has a further advantage of understandability by virtue of the simplified relationships among
the functional units.
Most of the signals that flow between units can be described as a combination of an action and an address. The
address gives the location at which the action is to be performed. For example, the temporary memory might receive an order to write or read at a particular memory
site. Alternatively, the switching network might receive a
set of signals representing the action of "connect a voice
path" and the addresses or identities of the terminals to
be joined.
SYSTEM OPERATION

A block diagram of this electronic switching system is
shown in Fig. 1. The memory and logic units are separated
from the voice switches and gain access to lines and trunks
only through the scanner and selector. The scanner and
selector are multiposition diode switches which may be
directed to particular terminals for collecting or transmitting information. The function of the scanner and selector
is to permit the fast control circuits to be time shared
among the telephone customers. Information gathered by
the scanner is processed by the controls and results in
orders to the switching networks or to trunks via the selector. At any instant the system is usually engaged in
processing only a part of a single call. Simultaneous actions involving more than a single call rarely occur.
In order to meet the real-time demands of the telephone
customers, some system actions are given higher priorities
than others. For example, it is more important to count a
dial pulse than to detect a call origination because of the
transient nature of the dial pulse. Thus, in each 5-msec
interval, the system goes through its more urgent tasks
first and then, if it has time, completes its less pressing
commitments. The cycle time for any single logic memory,
or scanner function is 2.5 ILsec. While this provides 2000
operations in eacn 5-msec interval, many of the tasks take

Ketchledge: Bell System's First Electronic Switching Office

205

Fig. 3-Gas diode switch.

Fig. I-Electronic switching· system block diagram.

Fig. 2-Gas diode electronic crosspoint.

a number of operations to complete. Further, the number
of tasks varies sharply with the telephone traffic. In a
typical interval the system would first scan about half of
the lines which are dialing in order to gather and record
any new dial pulses. Next would come selector actions involving, perhaps, pulses being sent out on trunks to other
offices. Then, any traffic awaiting network action might be
completed. If this exhausts the interval because, for example, a .large number of dial pulses occurred, then the
system immediately goes back and repeats these high pri.ority tasks in the next interval. This defers the lower
priority tasks, but not for long since the probability that
two successive intervals will both be overloaded is low. In
the s~cond interval the other dialing lines are scanned, and
so forth. Then, lower priority tasks such as regeneration of
the barrier grid stores, scanning of all lines for call origination, etc., are completed.
These actions result in all lines being scanned ten times
a second for call origination. If a line is found which is
drawing current but which the memory reports was idle
on the last look, the controls recognize a service request
and record the situation in the memory. The presence of
this line number in the memory results in the line being
scanned 100 times a second during dialing. The higher
scan rate is required to insure detection of all dial pulses.
SWITCHING NETWORK

The switching network and the associated concentrator
network provide the voice frequency paths for interconnecting telephone lines with each other and with trunks
and various signals (ringing, dial tone, etc.). The switching element is a cold-cathode gas tube as shown in Fig. 2.

Fig. 4--Cabinet arrangement for the switching network.

It is a neon-filled diode utilizing a hollow cathode to obtain negative resistancejn the conducting condition. 2 This
tends to compensate for transmission losses of transformers and other elements in the talking path. These gas
tubes are arranged into switches of the type shown on
Fig. 3. Application of one half of the breakdown voltage
on an input and an output wire causes the gas tube joining these wires to fire and connects the wires for speech
transmission. Only one side of the transmission circuit is
switched, the other side being grounded. Large numbers
of these switches are connected together to form the complete network. A typical connection would be through one
tube in a concentrator switch, then through six tubes in
the switching network, and finally through one tube in a
concentrator switch to the other telephone. Fig. 4 shows
the physical form of the switching network. The gas tube
crosspoints and the control circuits are assembled in
plug-in packages which are then inserted in the cabinet.
This permits easy maintenance and growth.
LOGIC

In general, the circuitry used for the processing of control information in the electronic switching system can be
characterized as asynchronous and dc coupled, using germanium-alloy junction transistors as the active elements.
2 W. A. Depp and M. A. Townsend, "Cold cathode tubes for
audio frequency signaling," Bell Sys. Tech. J., vol. 32, pp. 13711391 ; November, 1953.

206

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Fig. 5-Barrier grid store block diagram.

It is constructed of smaJ1 general purpose circuit packages
which are interconnected in accordance with the system
control operations to be performed.
Logical operations are performed by conventional semiconductor diode AND and OR gates. To permit standardization of the design of these gate circuits, transistor gate
amplifiers are placed in chains of diode logic to maintain
appropriate signal levels. Bit storage associated with the
logic circuitry consists of groups of double-ended flipflops. These transistor flip-flops are coupled to the diode
logic circuitry and to external circuitry composed of relays, gas tubes, or magnetic cores through specially designed ampli.fiers. Other miscellaneous circuit packages include inverters, emitter followers, pulse stretchers, and
cable pulsers.
Most of the logic circuit packages are based upon
grounded emitter transistor configurations in which the
transistors are held out of saturation by self-biasing or
diode feedback arrangements. This permits high-speed
pulse operation at reasonable gains (O.5-~sec rise and fall
times with current gains of approximately 15 in the gate
amplifier) .
The logic circuit packages are physically laid out on
printed wire boards with all components crimped around
the edge and dip soldered. The circuit boards are all jack
mounted in the equipment. Both cards and jacks are coded
to prevent improper interchange of packages. A shoe is
provided on each card using transistors, the removal of
which permits the transistors to be tested without unsoldering their leads.
THE TEMPORARY MEMORY

Two general types of memory are provided, temporary
memory and permanent memory. The temporary memory
is used to record data which must be changed in the course
of processing a telephone call. Conversely, the information recorded in the permanent memory is not changed
during a call, although numerous references to this permanent information may be required. The distinction between
these two forms of memory is therefore based not on
reading but rather on their writing characteristics. In the
temoparary memory, information may be recorded in little
over 1 ~sec whereas several minutes are required to change
information stored in the permanent memory.

Fig. 6-An early model of the barrier grid store.

The temporary memory device used in the electronic
switching system is the barrier grid tube. 3 This is an electrostatic storage tube wherein binary bits of information
are recorded as electrostatic charges. An electron beam
and electrostatic deflection plates provide access to the individual storage areas on a mica target. Writing is controlled by manipulation of the electric field at the mica
surface.
The complete barrier grid store, shown in Fig. 5, consists of the barrier grid tube, deflection circuits to position
the electron beam, circuits to turn the electron beam on
and off and to pulse the target, an amplifier for the output
signal, and control circuits to cause the various operations
to occur in the correct sequences and at the right times.
The barrier grid stores now operating are capable of storing 16,384 bits in a 128 by 128 array. In addition, the
functioning times are quite short: 0.4 ~sec to deflect
the beam; 0.7 ~sec to erase, of which only 0.3 ~sec is required to read; and 0.7 .~sec to write. The typical cycle is
to read and then write at a given address, and this requires
less than 2 ~sec. Fig. 6 shows an early model of a store
having these characteristics.
THE PERMANENT MEMORY

In the electronic switching system there is a need for
millions of bits of storage with access for reading in no
3 M.
E. Hines, M. Chruney, and J. A. McCarthy, "Digital
memory in the barrier-grid storage tubes" Bell Sys. Tech. I.,
vol. 34, pp. 1241-1264; November, 1955.
'

Ketchledge: Bell System's First Electronic Switching Office

207

Fig. 7.-Television flying spot scanner.

more than 2.5 !J.sec. This information is of a semipermanent character and, for reliability, must not be lost by any
failure of an electronic element. Storage of a photographic
character was developed to meet this need. A group of
photographic plates can store vast amounts of information
as transparent or opaque spots and, with optical interrogation, an electrical malfunction cannot cause a loss of
the stored information. Fig. 7 shows an example of a type
of photographic storage, the flying spot scanner used in
television. This represents a method for converting the
transparency of a selected area on a photographic emulsion into an electrical signal. As shown in Fig. 8, the flying spot store uses similar techniques to read a number of
photographic areas simultaneously.
In the photographic memory under development, over
forty photographic areas 1.5 inches square are used to hold
approximately three million bits of information. Each area
is an array, 256 by 256, and individual spots are approximately 0.006 inch in diameter. Each photographic area has
associated with it a lens to focus the spot of light from a
cathode-ray tube. By use of a multiplicity of lenses in
front of the same cathode-ray tube, a number of photographic areas can be interrogated simultaneously. Quick
random access to any of the 65,536 words is achieved by
electrostatic deflection of the cathode-ray beam.
In order to achieve practical photographic storage it is
necessary to find means for positioning the small spot of
light on the face of the cathode-ray tube to an accuracy
and reproducibility of less than l/1000 of an inch and in
times of the order of a millionth of a second. This problem has been solved successfully. The basis of the solution
is an optical-electrical feedback system which uses mechanical edges as the references for spot positioning. Fig. 9
is a photograph of an early model of the flying spot store.
CONTROL BY

STORED PROGRAM

The use of a stored program in place of wired logic is
made possible by high-speed, random access, large capacity,
permanent memory. The complexities of a telephone office
are such that hundreds of thousands of bits of program
are required. The real-time nature of the system requires
fast random access, even though the stored information
is changed infrequently. This substitution of memory for
logic makes the remaining wired logic become a general
purpose unit for interpretation and routing of order

Fig. 8-Flying spot store block diagram.

Fig. 9-An early model of the flying spot store.

words. Thus, the machine problems are solved not by
"logic" but by looking up the answers in the "back of the
book." The stored order words thus control the system sequences and decisions for both telephone calls and internal
trouble detection and location. The advantages of the use
of a stored program are reduced system compleixty, fewer
variatioris in manufactured units, reduction in wired options, simplified engineering of particular installations, and
great flexibility. This flexibility permits the addition of
new features and changes in control sequences with little
or no change in hardware. The electronic switching system can be programmed to perform a wide variety of complex tasks and to render telephone service in a wide variety
of ways. Changes in operation or provisions of new services are achieved by modification of stored information
rather than by changes in wired connections. This same
ease of modification is also useful for·changes in line

208

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

translation records which are also recorded in the permanent memory. These records give, for each directory number,the network terminal at which it appears and all class
of service information.
SPECIAL SERVICES

The application of electronic techniques to telephone
switching make possible new services; services which have
been impractical heretofore due to either economic or technical reasons. One example of what could be done is abbreviated dialing where one dials a preliminary "one" followed by a single digit. Each of the ten possible digits
represents a preselected, frequently called, telephone number. You might choose to have a 1 represent a nearby
friend; 2, your office; 3, a relative living thousands of

miles away, etc. The system would recognize the type of
call and, using its photographic plates, translate your IX
into the actual telephone number, whether it be a local call
or not. The translations would, of course, have been previously been recorded in the system. This and many other
services are made technically feasible by the use of a
stored program and the use of electronic memory.
CONCLUSION

The introduction of electronic techniques into telephone
switching represents a major change in the art. Both the
new types of devices and new types of telephone system
organization offer important advantages. Perhaps the most
important result will be the increased flexibility and new
services that this will make possible.

Traffic Aspects of Communications Switching Systems
JOSEPH A. BADERt

o DESIGN a communications switching system
which provides a satisfactory grade of service at
minimum cost, an understanding of the nature of
the offered traffic is necessary.
For many years, telephone traffic engineers have been
studying the problem of providing sufficient equipment to
meet time-varying demands at a given level of service. The
experience gained in these studies and the analytical results
derived may prove valuable in the planning and use of
modern digital data processing equipment for real-time
applications.

MAJOR COMPONENTS OF TRAFFIC VARIATION

In order to design a switching system to meet a specified
grade of service, an estimate of the average traffic offered
to the switching system must be made. The reliability of
this estimate depends on the magnitude of the components
of variation present.
Variations such as seasonal, day-to-day, and hour-tohour are not easily susceptible to an analytic approach.
They are usually determined empirically for each exchange
or area. Typical seasonal, daily, and hourly traffic patterns
are shown in Figs. 1 to 3. These patterns are usually stable
so that in choosing the average busy season, busy day,
busy hour traffic as a base for engineering, we can be
reasonably assured that the busy hour traffic offered during the rest of the year does not greatly exceed our engineered estimate. Of course, events occur such as eartht Bell Telephone Labs., Inc., New York, N.Y.

quakes, snow storms, disasters, etc., which provide a common cause for call origination. This results in traffic
"peaking" well above the engineered level with a resulting
degradation of service. Under such extreme conditions,
special overload-control procedures are usually initiated
which tend to spread the peaked demand over a longer
period of time.
The remaining component of variation is the instantaneous variation of the number of calls offered per unit time
during the busy hour. Fig. 4 shows the variation in the
number of calls offered per 24-second interval for a 10minute measurement period. From these data, a frequency
distribution giving the fraction of 24-second intervals containing % call originations was constructed (Fig. 5). If
calls arrive at random, then the distribution per unit time
should be Poissonian. To test the data for randomness of
call arrivals, a graph of the Poisson and the sample distribution was constructed (Fig. 5). By inspection, the
agreement appears to be quite close. It is evidence of this
kind which lends assurance to the assumption of randomcall input which is basic to most traffic theory. Accordingly,
we can compute the magnitude of the instantaneous variation of offered traffic as the first step in determining the
engineered capacity of the system.
HOLDING-TIME VARIATIONS

The next step is to investigate the length of time required to serve the offered calls. The service time, generally called the holding time, ranges from a fraction of a
second for certain switching equipments to several minutes

Bader: Traffic Aspects of Communications Switching Systems

209

100

2.4
2.2

2.0

'"",-""-

1.8

1.4

1.2
100

1.0
)0

.2
JAN

FEB

MAR

APR

MAY

JUNE

JULY

AUG

SEPT

OCT

NOV

DEC
SUCCESSIVE 24

S~CCND

lNTERVh!.

Fig. I-Seasonal variation in daily local calls in Boston, 1934.
Fig. 4-Variation in number of calls offered
per 24-second interval, Asbury Park, 1957.

~ 20
~

:r~

lI'-~ ~t--. ~

-.
.... 1--1

~ .I

I/"~ "\

c;;
w

2
10

3£
Q

9
~

4
2

JULY 8 9 10 II 12 13 15 16 17 18 19 20 22 23 24 25 26.27 29 30 31 AUG.I 2
DAYS

Fig. 2-Day-to-day busy hour variation in load, Newark, 1918.

Fig. 5-Comparison of Poisson theory to measured
number of calls offered.

~.~~-+--r1-~~~

.20

METROPOLITAN OFFICE. LONDON
- - - WAVERLY TO MULBERRY. NEWARK
COPENHAGEN. DENMARK

~~ 41---+---+--HH

2 1\

1\
\

'f/

•

~".

HOUR.S

Fig. 3-Hourly traffic distribution. (*) values shown are per cents
of total traffic handled in 12 MIDNIGHT-6 A.M. period.

.02

~~
2.

~~""-r--- -_..... -3

LENGTH OF CALLS IN MINUTES

Fig. 6.

for talking paths. To expedite theoretical treatment, one
of the following two assumptions is usually made with
regard to call lengths or holding times: 1) holding times
are distributed according to the exponential distribution, or
2) holding times are constant. The assumption of exponentially varying holding times is very well confirmed
where local-call conversation and control time are examined. Fig. 6 shows the distribution of such times as
measured in Newark, N.J., in a past traffic study. The
agreement between data and theory is obviously good.
Based on evidence of this kind, an exponentially varying

holding time is assumed for equipment held during the
entire conversation.
Constant holding time is exhibited by equipments whose
function is to perform rapid switching operations prior
to the beginning of conversations. For example, in one
type of telephone switching system, a so-called marker
performs the fUnction of connecting the subscriber's line
to an outgoing trunk. The marker then releases and is
immediately available to another subscriber. The time required for this function is essentially constant.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

210

TRAFFIC USAGE
Having determined the number of offered calls and the
server's holding time, we can define a third parameter
called traffic load, or usage. This is the product ot the number of calls and the average length of each call. Thus, if a
system is offered 1000 calls in the busy hour, each of average length 100 seconds, the offered load is 100,000 call
seconds. In dealing with a system which is offered a large
number of calls per hour, it is more convenient to deal
with call hours per hour. In our example, then, the offered
ioad would be 100,000/3600 or 27.8 call hours per hour.
The traffic unit of one call hour per hour has been named
the erlang in honor of A. K. Erlang, Danish mathematician, who pioneered in traffic theory. The offered load expressed in erlangs represents the average number of simultaneous calls that would be in progress if sufficient servers
were always available. We can see this from Fig. 7 which
shows the variation in the number of calls present on a
trunk group between two central offices during the busy
hour. During this hour, 246 calls were carried. Greatest number of calls in progress was 16 and this occurred
four times during the hour. Average number of simultaneous calls in progress was 9.5. This means that on the average during this hour, the trunk group carried 9.5 call hours.

them in the system. If % exceeds the number of servers
provided, newly arriving calls will fail in obtaining immediate service. The proportion of such failures is a commonly used criterion for the adequacy of service. Modification of the assumptions underlying the Poisson gives rise
to other formulations. In particular, the assumption regarding the behavior of calls failing to find an idle server
immediately is of primary interest.
DEFECTION RATIO
The behavior of calls which fail to find an idle server
immediately can be expressed in terms of the defection
ratio j. This is the ratio of the rate at which waiting calls
abandon before being served, to the rate at which they are
served. j, of course, can assume values from 0 to infinity.
However, in Telephone Traffic Engineering it is usual to
find only three different values of j assumed. These values
and their physical interpretation are as follows. 1) j = 0
corresponds to the case in which unserved calls wait indefinitely for service. In telephone traffic parlance, this is
called the "lost calls delayed" assumption. 2) j '= 1 corresponds to the case in which calls wait no longer than
their holding time and then abandon. I f an idle server becomes available, a call seizes the server and uses it for the
remaining part of its holding time. This is the "lost calls
held" assumption. 3) j = infinity corresponds to the case
in which calls are not willing to wait at all for servers and
immediately abandon. This is the "lost calls cleared" assumption.
Loss ENGINEERING

TIME IN MINUTES DURING A TYPICAL 'BUSY HOUR"

Fig. 7.
.16

f--~

~.12

IV
j'

41-·
1--1--1-o

l.---

".~
,'-....

OBSERVATIONS

R"::
I~

r\.

J
6

vPOISSON THEORY

1-11--1--

1-1-1-

~~
16

NUMBER OF CALLS IN PROGRESS

Fig. 8.-A comparison of theory with the
observatio~s in the hour above.

As indicated previously, the Poisson distribution correctly describes the variation in the number of randomcall arrivals per unit time. Under certain assumption, however, the Poisson also describes correctly the probability of
finding at a random instant a given number of calls on a
group of servers. Fig. 8 shows the Poisson with average
9.5 and an empirical distribution of the number of calls in
existence each 30-second interval taken from the data of
Fig 7. In this case, the Poisson closely predicts the proportion of time % calls that are present; this is also clearly
the fraction of calls arriving and finding % calls ahead of

A group of servers engineered solely on the basis of "expected proportion of calls which fail to receive immediate
service" is said to be engineered on a "loss" basis, and
formulas used to predict the proportion of calls failing to
find an idle server immediately are called "loss" formulas.
Loss formulas and their underlying assumptions are listed
in Fig. 9. The assumption of "infinite sources" is, of
course, never quite realized in practice but where the rate
of arrival of calls is nearly independent of the number of
calls momentarily being served, this assumption can be
used with confidence. The list in Fig. 9 is by no means
complete. However, the formulas tabulated are those most
widely used for engineering telephone switching equipment. Graphs of these loss formulas are shown in Figs. 10
to 12. By means of these curves, the traffic engineer is
able to solve a wide range of "loss-engineering" problems.
The following examples demonstrate the use of these
curves.
E%ample 1

How many trunks should be provided at P.O 1 serVice
if the load offered is 10 erlangs?
Solution: Past experience indicates that for trunk
groups with no provision made to reroute overflow calls,

Bader: Traffic Aspects of Communications Switching Systems

211

e = Number of Full Access Trunks
a = Load Submitted in Average Simultaneous Calls

= (N~mber ~~~~~.2I0ur)(AverageE~~~E~e i~Se~nds)
3600

Lost Calls
Assumption

Usual
Designation
of Formula

"Delayed"

Erlang "C"

"Held"

Poisson

"Cleared"

Erlang "B"

Frequency Distributions, j(x)
Probability of Delay, P
When x6) = 0.23 so that 23
per cent of the delayed cars are delayed more than 30
seconds. Multiplying by C(1.75) = 0.75 we have 19
per cent of all cars delayed more than 30 seconds.
Example 5
A counter in a department store is manned by three
clerks. The time required to serve a customer is distributed exponentially with an average of 3 minutes.
Customers receive a number indicating the order of their

213

Bader: Traffic Aspects of Communications Switching Systems
0.8

'):'
~ 0.7

lL.

v/!7

0.5

Z
:J

t--

a:
w

0.4

w 0.3

~ 0.2

«
II

~O.I
III

11)

I~/~

.f,'/

c/~

ti 10
'-J

.09

20.

30
40
50
60
70
C =' NUMBE,R OF TRUNKS

/
II

./.os

./2

II
II

~
o

.f.'4

/ II

t:I

f'",o!oJ.-t1--1-/~
v
V

~]~

w
U

iL 0.6
~

'S'

l----I-~

Iff I
/) /
It' /

VII '/ /

100
'2

Fig. 15.

.0'I

arrival so that service in the waiting line is very nearly
first come, first served. Observations indicate that during
the normal busy hour, 10 per cent of the delayed customers
are delayed at least 6 minutes in acquiring the services of
a clerk. If one clerk is added, what per cent of the customers would be delayed at least 6 minutes?
Solution: Using PC >6) and Fig. 13, determine the ofa
fered load a. Since - = 0.62, we have a
c

1.86. If one

clerk is added, the new occupancy is alpha

a
1.86
-= - c
4

0.47. From Fig. 13 we find that 1.40 per cent of the delayed
customers would be delayed at least 6 minutes. The addition of one clerk resulted in over a 7 to 1 improvement
in the per cent delayed at least 6 minutes.
OVERLOAD PERFORMANCE

So far we have considered the problem of engineering
equipment to accommodate an average busy hour load at
a given grade of "loss" or delay service. Related to this
problem is that of balancing efficiency against overload
capacity. A brief discussion of this problem might be of
interest.
The efficiency of a trunk group or average load per
trunk is defined to be the ratio of the load carried to the
number of trunks. Fig. 15 shows the relationship between
efficiency and group size at engineered losses of P.Ol and
P.OOL From the curves, it is clear that large groups are
more efficient than small groups. On the other hand, the

/
!!
/
II

}V/
~~

~oo

P".OI

t:r

,0/

;Y
/

110

PC'RCeNT Of"

120

IJO

140

ISO

eNGINeeReD LORD

Fig. 16-Relative ability of different sizes of
trunk groups to carry overloads.

higher the efficiency, the less margin available percentagewise for small overloads. Fig. 16 shows the relationship
between the increase in loss from P.01 against group size
when the offered load is over engineered level. The ideal
balance between efficiency and overload margin depends on
additional factors such as the purpose for which the system is to be used and the environment in which it is to
function.
CONCLUSION

Some of the fundamental traffic aspects of switching
systems and some formulas by which probabilities of delays and losses may be calculated have been displayed.
Working curves have been shown by which many trafficengineering problems can be solved. Examples are given
which illustrate the application of the curves in practical
situations.
BIBLIOGRAPHY

[1] Brockmeyer, E., Halstrom, H. L., and Jensen, A. The Life and
Works of A. K. Erlang. Copenhagen: Copenhagen Telephone
Company, 1948.
[2] Molina, E. C. "Application of the Theory of Probability to
Telephone Trunking Problems," Bell System Technical f ournal) Vol. 6 (1927), p. 461.
[3] Fry, T. C. Probability and Its Engineering Uses. New York:
D. Van Nostrand Company, 1928.
[4] Thorndyke, F. "Applications of Poisson's Probability Summation," Bell System Technical fournal) Vol. 5 (October, 1926),
p.604.

214

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

The Use of the IBM 704 in the Simulation of
Speech-Recognition Systems
G. L. SHULTzt

I NTRODU CTION

HE !ERM spe~ch-recognition devi.ce ~sually brings
to mmd a machme capable of duphcatmg the function of a human listener. Such a device not only
would have to be capable of receiving and classifying
acoustic stimuli, but also would have to be able to extract
from these stimuli the message the speaker intended. To
do this, a recognition machine must be familiar with the
language statistics, and indeed, the entire human environment. Many years of investigation will be required before
such a human replacement can be achieved.
However, more limited man-to-machine communications
systems can be defined, and acoustic (as distinguished
from speech) recognition devices can be built in the near
future. Articulation tests with nonsense syllables show
that listeners can agree upon, and classify, speech sounds
on the basis of the acoustic signal with little use of language redundancy. Surely, then, a set of measurements
exists by which a machine could likewise classify these
sounds. Since classification of speech sounds is a necessary
part of even the most comprehensive recognition system,
our efforts are first turned to this task.
The study by introspection how we, ourselves, classify
speech sounds is not very successful. We need to "look at"
speech. The sound spectrograph was developed by B~ll
Laboratories to produce visible speech. An example of 1ts
display is shown in Fig. 1. The more familiar voltage
amplitude vs time function is shown at the top of the figure. Directly below, and aligned in time, is the sound spectrogram of. the same utterance. The frequency is scaled
along the ordinate. Intensity at a given frequency and time
is depicted by the blackness of the mark. Rules for reading
these displays have been developed and reported in the
literature by Bell Telephone Laboratories, M.LT., Haskins,
and others. These rules are being presented to us in a
unified course conducted by Prof. Morris Halle at M.LT.
Note that the spectrogram is divided into segments of no
activity, horizontal bar structure, and areas of striation.
The horizontal bars, termed formants, characterize the
vowels, (r), 0), and nasal consonants(m), (n), ('fJ). The
irregular striated areas characterize the fricative or noiselike consonants (s), ( S ). A vertical blank area, followed
by a sharp vertical line, is the distinguishing property of
stops or plosives (t), (P), (k). The presence of a heavy
horizontal bar at the bottom of the spectrogram indicates

t IBM Corp., Yurktc.wn Heights, N.Y.

Fig. l-Comparison of amplitude vs time function
and sound spectrogram.

voicing. Shown here, the first word consists of an unvoiced fricative, a changing vowel sound, and a terminating unvoiced stop. The second word is voiced throughout
and begins and ends with a nasal sound characterized by
abrupt transitions in the formants of the middle vowel.
The third word is the same as the first. Although we have
only mentioned some rules for classifying vowels, fricatives, and stops, such rules have also been developed for
subdividing these three classes into the approximately 40
basic elements of speech.
These qualitative rules must be operationally or quantitatively defined in solving the. problem of mechanical recognition. For example, just what circuit would be able
to identify a striated area? Even after quantitative rules
have been defined, a set of physical properties will result
whose ranges of variation with context and speaker must
be determined. This calls for a statistical approach with
its consequent data-handling problems. Further, the number and complexity of these speech properties require a
versatile system of analysis. To accomplish this analysis,
special advantage is taken of techniques made possible by
the advent of the large-scale digital computer.
A computer can be programmed to duplicate any of the
measurements of speech signals now used in speech studies.

Shultz: Use of the IBM 704 1,n the Simulation of Speech-Recognition Systems
727 MAGNETIC
TAPE UNIT

•

~"'(MOP)_

•

215

COMPUTER

. .- /~.' )M)
\1.1.:.

o..~
.

. . -\~ ).0
Fig. 3-Editor and Coder.

~
Fig. 2-System of analysis.

The computer is especially well suited for the large-scale
reduction and analysis of data. The flexibility of the computer enables a recognition system to be evolved through
easily inserted program modifications. Further, the inertia
associated with the construction of specialized equipment
is eliminated.
When a set of speech properties has been found, and a
successful system based on this set has been duplicated
in the computer, then the system can be embodied in the
circuits of a speech-recognition machine.
SYSTEM OF ANALYSIS

U sing the computer as a central tool, we have built up
the system of analysis as outlined in Fig. 2. In order to
gather a large number of like speech events, a device is
required to edit these events from continuous speech.
Once a speech event has been selected, the acoustic wave
must be converted to a digital form satisfying the input
requirements of the computer.
The initial program routines are designed to aid in
determining which speech properties are most significant
with regard to recognition. First, the proposed property is
computed for a large number of speech events. Then, the
statistical distribution of these measurements is estimated
and listed by like speech event.
We have divided the measurement of speech properties
into two operations. The transformation block in Fig. 2
contains a basic program which yields a quantitative pattern of the acoustic signal. The many properties of this
pattern are then explored by a set of simpler features
programs.
AN ALOG-TO-DIGITAL CONVERTER

The input system of equipment required for the computer analysis consists of two machines, an Editor and a
Coder, each with dual-track, audio-tape devices. Fig. 3
is a photograph of the Editor (left) and Coder (right).
The Editor aids in the selection of speech events from con-

~~~~-.Jl
PULSE

W.IAIII.
I
1

I I
I

---.Jr-!--""--;1

I
I

~I~I---------

I
I

1
I

1"....,'-_ _ __

Fig. 4-Edit timing delays.

tinuous speech. The Coder converts these selected events to
digital form. Care has been taken to make this system
automatic so that the main concern of our work can be
the study of speech events rather than their preparation.
The desired speech event is selected by the Editor under
push-button control. Editing is accomplished by transferring the section of speech containing the speech events of
interest to an endless tape, or loop. Synchronizing pulses
are placed on the second tracks of both the input and
endless tapes as the speech is transferred. When the loop
is read, two electronic delays, as shown in Fig. 4, are
initiated by the loop synchronizing pulse. One delay is
adjusted to extend from the loop synchronizing pulse to
the start of the speech event. A second delay, oscilloscopesweep length, is then adjusted to the duration of the
speech event. This second delay also keys the selected portion into earphones, permitting simultaneous sight and
sound adjustments.
After the delays are properly set in the endless tape
operation, they are initiated once more by the synchronizing pulse on the second track of the input tape. This final
timing sequence properly writes an editing pulse on this
track. Hence, the result of the editing operation is an audio
tape with speech recorded on one track and editing pulses
located on the second track opposite the selected speech
events.
Fig. 5 shows the entire Editor-Coder operation in block
diagram form. It has been split here for convenience into
the audio and pulse tracks. As mentioned, the audio signal
is recorded once on the Editor, and during the editing op-

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

216

o f-----(i)

a.. -2

c01~~0r.. }
.------,

:~~

DIGITAL
AUDIO

10:

t::>-- ,......, T""" .....
,,

CUTPUT
-r~

lL""

'"'4

-6

~. -8

,,
~

cd -10

·e

INPUT""----

~ -12

~ -14

-16

-I~O

100

1,000

10,000

FREQUENCY (CPS)

Fig. 7-Frequency responses of Editor-Coder.

Fig. 5-Editer-Coder audio and timing.

ltJ

gr++-~~-+~Rr+-~~-+~

...

~~~~~~~~~~~~~

0..

~r-~H-~~~~+-H--a~~

::::::::::::::::::: ..
....................
..................

r-~~-++4~~~4-~~~

:::::::::::::::::::
AMPLITUDE
(TOP.20 OBI

EDITED SIGNAL

31:

ltJ·

=>:

!:::
...J'

0..:

ct:

...

....

...
... .... ...
... .... :
...
: ..
.. : .. ....
.. ..
.. ..
....

..
..
:

:::

.. :

....

..
..

....
..
... ..............
~J:~;~~ ~~.5 ~s
.,

COOED SIGNAL

Fig. 6-Selection accuracy.

eration a pulse is affixed on the second track of this tape
opposite the speech events selected. This tape is then
played back in the coder where the audio information is
continuously passed through a playback amplifier, an
equalizer, a compressor, and a low-pass filter before entering the analog-to-digital converter. Upon reading a pulse
from the second track of the tape, the converter, externally
triggered at the prescribed 727-tape character rate of 16
kc, converts and speech-signal amplitude at each sample
point to an II-bit binary number. Simultaneously, the editing pulse causes control circuits to bring the 727-tape unit
up to speed and properly write the converted speech signal.
One 727 -tape record is made for each speech event selected. At present, only the five most significant bits are
written on digital tape since this accuracy yields a sufficiently low quantizing noise value. The speeds (7~ or 15
ips) of the input tape in the Editor and the playback tape
of the Coder can be arranged to produce effective sampling rates of ~, 1, or 2 times the tape character rate.

Fig. 8-Diagram of filter bank with spectrum .

The selection accuracy of the Editor is shown in Fig. 6.
The 2-kc sine wave is shown at the top of the figure as it
was selected in the editing operation. Below is shown a
print out of this sine wave sampled at 16 kc. The accuracy
of selection is primarily limited by tape-speed variation
during delay 1. For a delay of 1 second from the synchronizing pulse to the start of the speech event, the start is
located with an accuracy of approximately 2.5 msec.
The audio system response is shown in Fig. 7. Here the
input signals were attenuated at a rate of 6 db per octave
above 1000 cycles. With this input and the high frequency
emphasis circuit inserted, the over-all audio response between half-power points is from 85 to 7500 cycles.
SYSTEM OF PROGRAMS

A set of programs has been written to implement the
general analysis system. These programs have been tested
together on a set of vowel phonemes.
The first transformation program was written to compute spectra. Essentially this program simulates a bank of
band-pass filters as depicted in Fig. 8. The output of each
filter is averaged for a certain period of time and this
average output is sampled periodically. At each sample
time the output amplitudes of the entire bank of filters are
plotted as a function of the center frequency of each filter.
The resulting graph is shown to the right of the figure.
In this graph, then, we have two of the constituents of the
spectrogram, namely, frequency and amplitude.
A series of these graphs at adjacent sample times would
display amplitude and frequency as a function of time. The

Shultz: Use of the IBM 704 in the Simulation of Speech-Recognition Systems

217

4.0oor-----------------------.
4,000, .

:::::~;::::::::.

3,500

Ui
Q.

2.500

3,OOOr-:::::

~ 2,500 ~

>~ 2,000

~ 2,000

3,000

I&J

::::.

:..

a. 2,500r-:

•
:

::>

•

::>

... 2,000

S
g:

1,500 :::::::: ......

::: ::::: :::::::::: ...

I,OOOr-::::::::::::::::::"

g:::I

........ .

.a:.

1,500r-~~~:::::::::::::

................
....

1,o00r-:::::::··
..............
................

...............
................

..............
................

:::::::::

3,500

........... . ..
................
...............
,. ........... .

500

...................
............. .
................
.................
............... .

S
II:

1,500

II.

1,000

.............
................ .

500 .

or-:::::::::::::::··

AMPLITUDE (DB)
[FILTER WIDTH a 184 cPs]
[FILTER WIDTH.II4 CPS]

...................
........... . .
.......................
...................
.................... ................
r.::::::::::::::
...................
................
................
.....................
............. . ..
500r-::::::::::.
..............
..
...............
..............
............. ..
................
.................
.................
.....................
................
.....................
........................
................... .....................
................
...................... ......
....................

Fig. ll-Spectra compared for two filter widths .

"
[FILTER .SHAPE

,.. AMPLITUDE

250

(DB)

~f\:p i NO EQUALIZATION;

...................
............. .
::::::::::::::::
..................
....................
:::::::::::::::::::: . ....................
::::::::::::::::::::.
150 !!!!!!!!W!!!!!!j~~:
....................
::::::::::::::::::::
.
................... ::::::::::::::
................
................. .
...
100 r-::::::::::::::: .•

NO AVERAGING]

200

Fig. 9-Spectra produced by original program.
~OOOr-----------------------------------__~

3,500r-:::::::
3,000 r-:: :::::: ..
2.5oor-:::::::::
.....
..................
...................
2.000r-::::::::::::::···
1,500r-::::::::::::::.
..............
. '::::::::::::::::::..
500 r-:~...................
~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~::.
IOOOr-::::::·
,

...................

AMPLITUDE (DB)
[FILTER WIDTH a 44 cps; AVERAGING IS FOR 20MS]

Fig. 12-Use of spectrum program for pitch finding.

................. ..
.................
.................
.................

.............. . .
.................
.................

................
........... . .
.............
..............
............
............
...............
.............
.............
................
................ . .
..................... .................
..................
.....................
.................. .
...................
.................
................
... .......... ..
.................
.
..............
..................
.
.
................... ...................
...................
....................
.................. ..

4,000..-------------------------------,
3,500
3,000
2,500

O~----------------------~------------~
AMPLITUDE (DB)

[FILTER SHAPE.

......•••••••••

A.....

;EQUALIZATION AND AVERAGIN\>]

FiS;. lO-Spectra produced by modified program.

2,000
1,500
1,000

500

...................
............. .
.............. . .
............
...............
,

...............
...................
~g~~~~i~~~i~~i:::··
...................

O~·~··~··~··~···~··~··~··~··~·---------------------------~

program was so written that the number of filters, the
center frequencies of each filter, the shape of each filter,
the averaging time, and the sampling rate could be modified by program parameter cards.
The flexibility of this arrangement was of great advantage in obtaining curves that provide a much-improved
display of formant structure. Spectra resulting from the
program as originally written are shown in Fig. 9. The
filter widths were 200 cps and the weighting function was
a rectangle which produced a high side-lobed frequency response. Note that formant structure is masked by the additional contributions of these side lobes. Furthermore,
formant position with time is not uniform since there was
no time averaging. Finally, the amplitudes decrease with
higher frequencies, making the third formant quite low.
By contrast, Fig. 10 shows a much-improved display.
Here, a cosine weighting function produced a filter frequency response with low side lobes. The formant structure is well defined for this 184-cps width. Time averaging
of 20 msec resulted in a smooth flow of formant position
with time. High-frequency emphasis of 6 db per octave
above 100(1 cps yielded a well-defined third formant. The

AMPLITUDE (DB)

Fig. 13-Formant tracking.

general equations computed by this. program are listed in
Appendix I.
Further investigations of filter widths were made using
this modified program. Fig. 11 compares a spectra obtained with filter widths of 114 cps and 184 cps. The ( :xl )
sound has a first and second formant so closely situated
in frequency as to become one broad peak for filter widths
of 184 cps. The 114-cps width resolves this broad peak
into the two formant peaks.
We are continuing to modify and improve this program.
Conceivably, an ultimate filter might be varied dynamically in accordance with the speaker's pitch so as to provide the best display for each speaker. Fig. 12 represents a
pitch-finding effort already made. To get the display shown,
we used the spectrum-analysis program described to simulate a filter bank of closely-spaced, narrow-width filters.
The first feature program was written to compute the
low-order moments of the spectra. This concise mathematical computation (Appendix II) permitted the system of

218

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

programs to be tested while our input equipment was being
constructed. Studies indicated that moments might provide
a means of distinguishing the front from back vowels.
A sufficient number of samples has yet to be tested.
In contrast to the mathematical statement of the moment
program, a formant-tracking program can be defined only
after a series of program modifications has been made
based on experience of a rudimentary program.
Such a "starting" program is now being written. Fig.
13 depicts the procedure followed in this program. First,
a peak is defined. Then, the peaks are examined further
to determine what peaks are considered formants. In order
to detect a "bar" structure, the peaks must be tracked with
time; that is, after a major peak is located in a given
spectrum, it must be confirmed in succeeding spectra before it is established as a formant. We can emphasize this
by pointing out that only after a threshold of formant
duration has been determined through experience, can
we ignore spurious indications such as the third peaks appearing here in the second and third spectra. By taking
full advantage of the 704, we hope to evolve a highlydetailed, formant-extraction method.
The final stage of our system is the statistical evaluation
program. For a collection of measurement values, the statistical evaluation program, as it is now written, can develop the frequency table, can sample mean and standard
deviation, cumulative probability function, and probability
density function. The term evaluation will have more
meaning as experience with this program grows. For
example, when we find ourselves consistently performing
further data reduction, and routine evaluation tasks, then
these tasks should be inserted in the statistical program.

ApPENDIX I

The spectrum-analysis program solves the following:

E·
10 10glO ') ~
\. N

N [( L:
s
L:
j=l

O"kW k

k=l

27rkJ- )2
sin --~
R

(t. ~kWk

cos

2T:/,- y]}.

The power-frequency characteristics are determined in this
equation through the simulation of the power outputs
from a bank of band-pass filters. The weighting functions
of the filters are products of the function, W k, which determines the filter shape, and the sinusoid, sin (2'/rkii) jR,
which determines the location of the pass band. S is the
segment, or summation, interval in input-time samples, and
(jk is the ac amplitude of the kth input-time sample of the
acoustica signal. The filter power output, Ai, for the ith
frequency, ii, is averaged for N segments of speech, which
corresponds to approximately two pitch periods (20
msec). These coefficients A i are further modified by E i,
the frequency equalization factor. R is the sampling rate
in samples per second. The equations used to compute the
two filter shapes discussed are:
Shape

Width*

R
2S~f'

Rectangular
Cosine

1/2 [1

+ cos (27rk/ S

- 7r)]

R
S~f

* Width between 3-db points of main lobe.
ApPENDIX II

CONCLUSION

I have assumed, here, the role of correspondent, reporting the result of the highly cooperative effort of my associates, Messrs. Welch, Wimpress, and Wilser. During the
past year, we have built the equipment and written a first
system of programs. This effort has been supported in part
by the Office of Naval Research.
I t is our belief that this system of analysis will provide
an efficient means of experimental study. Through this
study we hope to contribute to the understanding of speech.

Discussion
M. Martin (General Electric Co.):
What is the sampling rate used? How many
samples are used with each filter to determine the frequency spectrum?
Mr. Shultz: The audio is sampled at
the required 727-tape character rate of 16
kc. Since the two audio tape systems each
have two speeds, effective sampling rates
of g, 16, and 32 kc can be achieved. The
number of samples depends on sampling
rate and filter width. For the 44-cyc1e filter
for extracting pitch and for a sampling rate
of 8 msec, 160 samples are required.

The moment program solves the following:
N

L: (Qi -

Qo)p Ai

i=l

where M p is the pth moment about point Q0 in a spectrum
of Q equidistant coefficients, Ai, located at points Ni.

J. R. Barley (Du Pont) : Can you transfer the numbers 0 to 9 to digital tape?
Mr. Shultz: Yes. Speech events from
30 msec to 5 seconds can be edited from
continuous speech. A microphone input has
been provided on the CODER to allow
direct conversion of speech to digital tape.
L. S. Bearce (U.S. Naval Research
Lab., Washington. D.C.): Would you
comment in regard to the feasibility and
practicality of a speech recognition system that would operate in real time?
U sing your program in the 704, how
much increase, if any, in computing speed

and complexity do you think will be necessary?
Mr. Shultz: A bank of analog filters
would operate in real time. Once we have
completely investigated and specified a filter bank through this program, then it
might be desirable to shorten our analysis
time by building a bank of analog filters.
The spectrum analysis program produces spectra at a rate of 200 times real
time. For example, sound of 200-msec duration would require 40 seconds of computation. This delay is due largely to serial
computation of the power coefficient of
each frequency.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

219

An Automatic Voice Readout System
c. w. POPPEt

HE rapid development of electronic data processing
systems has created a need for improved manmachine coupling equipment. In systems where the
human operator is part of the control loop, the 'need for
optimizing this coupling is even greater. To date, most
display devices have utilized the human sense of sight. Additonal opportunities for improving the man-machine
coupling lie in the use of the other senses, particularly the
sense of hearing. The automatic voice readout systems
(A VRS) described here makes use of this sense of hearing.
This permits the operator the simultaneous use of his sense
of sight for other forms of data display.
!

TYPICAL ApPLICATION

One example of an application of the A VRS is in air
traffic control systems. Such systems are, in reality, closed
loop control systems and many of the well-known closed
loop techniques may be applied to the analysis of them.
Fig~ 1 shows a very simple example of the complete loop.

\ \

Fig.

l~CIosed

loop control system.

The position of the aircraft is obtained by the radar and
fed to the computing element. The computing element,
depending on the system, may range anywhere from a
sophisticated electronic system to the human controller.
The guidance commands which are outputs of the computing element must be communicated to the pilot for the
necessary aircraft position correction. As automatic computers such as form part of the FSQ-7, GPA-37, or
TSQ-13 systems become operational, the capability to
control large numbers of aircraft at high data rates requires automatic, high speed communications. Where these
systems are being used today with human operators relaying guidance commands, over-all system performance may
degenerate due to time lags or operator error. This is one
of the big reasons for the great interest and activity in
the development of automatic data links.
Camera and Instrument Corp., Syosset, N.Y.
*t Fairchild
Marine Div., Sperry Gyroscope Co., Roosevelt Field, N.Y.

AND

P. J. SUHR*
HISTORY

During the development of the Tactical Air Control
System, AN/TSQ-13, at Air Force Cambridge Research
Center, a project was initiated to develop a voice readout
unit for the digital data display [1 J, [2J. This unit gated
together previously recorded words to form the desired
voice message. As the readout unit progressed, it was felt
that as a computer readout it could also be used to transmit
the output of the ground-controlled-intercept-return-tobase computer to the aircraft in the absence of the digital
data link. One problem always present in the system tests
was the shortage of digital data link-equipped aircraft. It
was and still is obvious that it will be many years before
the majority of our operational aircraft are so equipped.
For this reason an automatic voice data link which requires no additional equipment in the aircraft can be used
to match the ground data processing equipment, now becoming operational, to the present day operational aircraft.
In addition, it will serve to provide automatic data link
capability for those aircraft which, for various reasons,
may never be equipped with digital data links, as well as a
back-up for the digital data link in the event of equipment
malfunction. It should be pointed out. here that the automatic voice data link does not significantly decrease the
bandwidth nor speed up the data rate over that of the
manual operator system, as do the digital links. It does,
however, provide the same automatic, error free advantages as the digital links while providing improved
intelligibility over the manual link. Moreover, the innate
redundancy in spoken languages provides the ability to
transmit through some fading and certain forms of jamming.
Returning briefly to the discussion of the history of the
automatic voice readout system, a breadboard unit was
constructed at AFCRC and flight tests were run at L. G.
Hanscom Field in mid-1955. Reports from pilots taking
part in these tests showed that the clear enunciation of the
words and the crisp concise message made for more inte1ligibleand efficient communication. During the last two
years improved techniques have been incorporated into
a model of this automatic voice readout system at Fairchild
Controls Corporation. It is this equipment which is discussed here.
DESCRIPTION

The inputs to the automatic voice readout system are the
output commands from the computer. In some cases these
are serial digital or parallel digital, while in other cases
they may be shaft position or analog voltage outputs. The
readout system (converts these inputs into a series of messages, the variable portions of which are controlled by the
inputs. These messages are then transmitted to the aircraft

220

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Fig. 2-Message structure.

Fig. 3-AVRS block diagram.
Fig. 4-Vocabulary storage unit.

using a standard voice channel transmitter such as the
GRC-27. Fig. 2 shows a typical message structure. It is
composed of words, one-half second or less in duration.
An unlimited number of message formats may be programmed and the variable portions controlled by the input
signals. In the example chosen, the aircraft address illustrated as red leader, the command heading shown as onenine-zero degrees and the altitude shown as 20,000 feet
are the variable portions of the message. In addition, one
or more message formats may be selected by the computer
outputs so that in addition to a command-heading-command altitude message, a range-bearing message, a landing
assignment message, or any other prescribed format may
be transmitted. The double beep tone transmitted at the
beginning of each message serves to alert the pilot and
to provide an indication of the message beginning. Fig. 3
shows a block diagram of the basic voice readout equipment. It is composed of five basic units.
The word storage unit is the heart of the system. It is
shown in the photograph of Fig. 4. It consists of a magnetic drum rotating at two revolutions per second and a set
of recording-playback heads. The drum is approximately
two and one-half inches in diameter, four inches long
and has a nine ten-thousandths thickness coating of magnetic oxide. The magnetic heads are spaced one and onehalf thousandths from the drum face and may be stacked,
ten tracks per inch. Such stacking allows about 35 tracks
in the full drum length. Storing one word per track provides a vocabulary of 35 words. The basic numerals,
"zero" through "nine," the programmed words such as
"vector," "angels," "range," "bearing," and twenty call
signs can be included in the 35-word vocabulary. It is to
be noted that vocabularies up to 100 words are possible on
a single drum. Such large vocabularies would be required
where large numbers of aircraft call signs were to be used.
All words stored in the vocabulary are recorded in phase
so that during anyone-half second interval all words are
being reproduced simultaneously at the magnetic heads.

This then allows the sequential switching between heads
to form the complete message.
The switching between heads to form the message is
performed by a relay pyramid. The five digit control signals required to control the pyramid are supplied from the
coding unit and the synchronizer. The output of the relay
pyramid drives the audio amplifier which in tUrn feeds the
transmitter or other output devices. The audio amplifier is
of standard design requiring only the equalization necessitated by the magnetic playback heads.
The synchronizer serves as the basic timing unit within
the system. Synchronization pulses are received every
drum revolution, or Yz second, and are used to advance
the synchronizer to the next word interval. The simplest
synchronizer consists of a six-bank, twelve-position
stepping switch. The incoming control signals from the
coding unit are thus sampled in sequence and the programmed words such as "vector" and "angels" are sampled at the proper interval.
The coding unit will vary considerably depending on
the types of computer outputs which must be accepted.
These must be converted into the five digit parallel code
used to drive the relay pyramid. Diode matrices are used
to provide the digital coding. Coded commutators may be
employed where the computer outputs are shaft positions.
An additional unit is generally required where several
computers are feeding the same voice data link. This
sampling unit which is not shown in the block diagram
must sample the various outputs and feed them one at a
time to the coding unit. It is conceivable that the computer
outputs could be sampled, coded and transmitted in sequence, but in all probability a priority system will be required. This will allow more efficient use of the command
channel since it will be necessary to transmit only when
a command is required. As an example, the original test
model voice data link was programmed to transmit only
when a heading change of ten degrees or more was re-

Kirsch et al.: Experiments in Processing Pictorial Information with a Digital Computer
quired. With such a system, it is believed that up to six
aircraft may be controlled over a single channel. It is worth
noting that a multichannel voice data link (i.e. a system
capable of transmitting several different channels simultaneously) could use a single word storage unit.
OPTICAL UNIT

One current program at Fairchild Controls Corporation
is development of an optical vocabulary storage unit. Such
a unit employing film strip for audio storage would provide an easy means of changing vocabularies by simply
inserting a different film disk. The size and weight of the
unit would also be reduced.
ApPLICATIONS

A large number of additional applications exist for the
automatic voice readout system. In air traffic control work,
it may be used to advantage with GCI, Return-to-Base and
GCA computers. Where voice communications must be
used but transmitted from a remote site, the commands
may be transmitted digitally from the command center to
an automatic voice data link located at the remote transmitter site. The voice messages are then transmitted to the
aircraft. The resultant system would greatly reduce the
number of land lines or voice channels required from
command center to remote transmitter.
Another important use could be in a multilingual area.
Vocabularies could be stored in several languages for use
in areas, for example, where NATO forces were stationed.

221

The complaint against the Frenchman's English or the
American's French could be eliminated.
The original concept of this idea is still an important
application. In many radar data processing systems,
cathode-ray-tube displays are used to read out computer
data. Other data can be obtained upon request and may be
displayed on an auxiliary status board. To obtain such
data, the operator generally must divert his attention from
his primary display. Often his dark adaptation is disturbed.
The reading out of this auxiliary data through a voice
readout would eliminate these undesirable points.
A combination data readout and communication application is the transmission of weather conditions from unmanned weather stations. Such stations, located on planned
flight paths, could transmit temperatures, barometric readings and other important weather data to enroute aircraft
either on interrogation or at periodic intervals.
It is believed that these applications represent only a
few of the possible applications of analog and digital to
audio conversion systems.
BIBLIOGRAPHY

[1] Krauss, B. F. "An Oral Readout System for Digital Data,"
Master's thesis, Massachusetts Institute of Technology, Cambridge, Mass., May 23, 1955.
[2] Pastoriza, J., and Poppe, C. '.'Audio Presentation of Aircraft
Data," Air Force Cambridge Research Center, Bedford, Mass.,
Systems Digital Group Technical Proposal No.4, June 21,
1954.
[3] Begun, S. J. Magnetic Recording. Murray Hill, N.J.: Murray
Hill Books, Inc., 1955.

Experiments In Processing Pictorial Information
with a Digital Computer
R. A. KIRSCHt, L. CAHNt, C. RA yt,

INTRODUCTION

N almost all digital data processing machine applications, the input data made available to the machine
are the result of some prior processing. This processing is done manually in many applications. Thus, such inputs as punched cards, magnetic tape, and punched paper
tape often are the result of a manual processing operation
in which a human being is required to inspect visually an
array of printed characters and to describe these data in
a form capable of being processed by machine. In recognition of the importance of automating such operations,
many investigations have been undertaken to devise autot National Bureau of Standards, Washington,

D.c.

AND

G. H. URBANt

matic character sensing equipment. Suppose, however, that
we attempt to view such efforts in proper perspective. We
find a more fundamental problem that has, heretofore,
failed to receive the attention that it warrants. The problem is one of making directly available to a computer pictorial information which would ordinarily be visually
processed by human beings before being fed to a data processing system. This pictorial information may range from
such highly stylized forms as printed characters, diagrams,
schematic drawings, emblems, and designs through less
stylized forms in cartoons and handwritten characters to
such highly amorphous forms as photographs of real objects, e.g., people, aerial views, and microscopic and telescopic images.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

222

In recagnitian af the impartance af pictarial saurces af
data far a data pracessing system, experiments were undertaken at the Natianal Bureau af Standards to' determine
whether autamatic pracessing techniques might be applied
to' .pictarial infarmatian in arder to' reduce the amaunt af
human interventian required during the input pracess. In
cansidering this problem, new areas of the application of
autamatic data pracessing techniques far pracessing pictarial infarmatian have appeared. It had nat been suspected
that autamatic data pracessing techniques were applicable
in same af these areas, even if human interventian were
allawed. The type af infarmatian with which these investigatians are cancerned ranges fram the stylized to' the
amarphaus farms previausly mentianed. In the NBS experiments described in this paper, the equipment used
cansists af the general-purpase digital camputer SEAC, to'
which are attached an input scanner far sensing pictures
and capying them intO' the camputer memary, and a
cathade-ray-tube autput display far repraducing pracessed
pictarial infarmatian fram the camputer memary.

II.

DESCRIPTION

THE EXPERIMENTAL EQUIPMENT

A. SEAC
The experiments described here were perfarmed an
SEAC during a periad when the capability af the camputer
far perf arming lagical data pracessing aperatians was being enhanced by the additian af several new features. The
state af SEAC at the time af mast af the experiments
described here was that af a 1500-ward memary camputer
with an average time of 250 microsecands far perf arming
a three-address instruction. Althaugh faster camputers exist, SEAC was faund to have one decided advantage aver
these machines, namely, its availability far experimental
use and madificatian an same frankly explaratary ventures.
The restrictian af having to' accaunt far every minute af
use an a mare pawerful machine wauld have been a seriaus deterrent to the praductian af the experimental results described here.
B. The Scanner
In arder to' feed pictarial informatian intO' SEAC, it was
cansidered adequate to' canstruct a simple mechanical drum
scanner which cauld digitalize the infarmatian in a picture
and feed it intO' SEAC in a few secands. The scanner is
shawn in Fig. 1. The phatagraph to' be scanned is maunted
an a drum abaut twa thirds of an inch in diameter. As the
drum rotates, a phatamultiplier and a source of illuminatian maunted an a lead screw pragresses alang and scans
the whale picture with a helical scan. The pitch af the lead
screw is such that the phatamultiplier assembly pragresses
0.25 mm alang the picture far each revalutian af the drum.
Between the drum and the photamultiplier, and in the image plane af the aptical system, there is an apaque mask
with a square aptical hale af such a size that a square area,
0.25 mm an a side af the picture, illuminates the phatamultiplier at each instant. A strabe disk maunted an the
same shaft as the drum praduces aptical pulses each 0.25

INFORMATION TO
PULSE DIGITALIZING
CIRCUITS

Fig. I-The scanner.

Digitalizing
Circuitry
(Two Valued)

I
Mechatr.ical
Linkjage

Discrimination
Threshold
Setting
(Manual)

PiCture
Information
In Binary
Form

Reference
-Signal

I
Strobe Disc
For Creating
SEAC Word
Format

Start and
Stop Signals
From SEAC

Fig. 2-The scanner connections to SEAC.

mm af drum ratatian. These aptical pulses are arranged
in the farmat af SEAC input wards, i.e.~ multiples af 44
binary digits. The time far the scanner to' scan ane phatagraph is 25 secands.
C. Method of Input

The scanner was first cannected to SEAC in Navember,
1956. As far as SEAC is cancerned, the scanner is just
anather input device, and it may be selected by the camputer interchangeably with such ather input and aut put
devices as a printer, magnetic tapes, etc. This is shawn in
Fig. 2. At any time during the aperatian af a pragram if
the input af phatagraphic infarmatian is called far, SEAC
starts the drum ratating. The analag signal fram the scanning phatamultiplier is campared with a dc reference signal
that has been manually determined with a patentiameter
setting. If the light reflected from the 0.25-mm square
being scanned is less than that needed to' praduce a signal
equal to' the reference signal, then when a strabe pulse
accurs, a binary 1 is fed to' SEAC. If a sufficiently white
spat is being scanned, a binary is fed to' SEAC.
The result af this aperatian is that in 25 secands (ar
less) SEAC can, up an demand, call far all (ar any part
of) a picture to' be fed intO' its memary. The whale picture
is 44 mm by 44 mm and is thus digitalized intO' 176 by 176
ar 30,976 binary digits, each binary digit representing the

Kirsch et al.: Experiments in Processing Pictorial Information with a Digital Computer

223

blackness of a unit square 0.25 mm by 0.25 mm in the
picture. The elementary squares cover the whole picture
and are nonoverlapping. The entire picture with one
binary digit per square occupies 704 words of SEAC
memory. One way of recognizing several different levels
of grayness is to use several scans of the picture made with
different manual settings of the discriminator threshold.
The mechanical precision of the equipment is such that on
successive scans of the same picture the scanner reproduces
its scan with a maximum discrepancy of less than 0.25
mm at any point in the picture.

D . Method of Output
As soon as the first picture was fed into SEAC, an uncomfortable fact became apparent. SEAC could store pictorial information in its memory, but the machine users
could not "see" the picture in the SEAC memory except
by the very time-consuming procedure of printing the contents of the computer memory on a typewriter and attempting to interpret the numerical information. Fortunately,
however, there was conveniently available a piece of equipment well suited to the task of producing pictorial output
from the SEAC memory. This equipment was capable
of decoding two ten-bit fields in a SEAC memory word
and producing two analog voltages corresponding to the
two sections of the word. There could be up to 96 such
words decoded. The rate of presentation of the output
signals was 23 kc.
The obvious course was to connect these two analog
voltage outputs to the horizontal and vertical inputs on
an oscilloscope and thus to plot points on the face of the
scope corresponding to the information presented by a
computer program for purposes of display. As it turned
out, the display equipment was more than adequately fast
for SEAC, since no program could generate useful display
information that would change at a 23-kc rate.
To display a picture that had been fed into the computer
from the scanner, it was then necessary to write a program
which derived a pair of coordinate numbers for each binary
1 in the picture in the SEAC memory in such a way that
these coordinates corresponded to those of the point in the
picture from which the binary 1 was generated. The program then displayed a spot on the output scope corresponding to the spot on the original picture.
As an example of the use of the output display routine,
Fig. 3 (a) shows a picture fed into the scanner and Fig.
3 (b) shows the same picture reproduced from the output
display. To produce this picture, an artifice was used
which allows the visual effect of a continuous gray scale
to be produced on a single scan. As seen in Fig. 2,there
is a manual discrimination threshold setting. Ordinarily,
this setting determines the level at which black is distinguished from white. To scan the picture of Fig. 3 this
discrimination threshold was varied with a sawtooth
waveform at a frequency approximately one half that of
the strobing frequency. The result was to produce the
familiar "halftone" effect in which the density of uni-

(a)

(b)

Fig. 3-(a) Grey-scale photograph. (b) Computer
halftone reproduction of (a).

formly black spots is proportional to the blackness of the
original picture. The picture display routine produced the
output photograph.

III.

EXPERIMENTAL INVESTIGATIONS

With the equipment that has been described, a series
of experiments were initiated with the goal in mind to program SEAC to recognize patterns of the type recognizable
by human observers. As a first step toward this goal, it was

224

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

decided to produce a library of picture processing subroutines which an investigator could use in programming
pattern recognition. This section describes the library of
routines that were written.
In pattern recognition the aim is to reduce the amount
of information to the minimum necessary for recognizing
one pattern from a group of patterns. In these experiments
the approach was to develop a library of computational
processes for simplifying patterns in order to obtain their
most significant features. Preliminary experiments were
concerned with determining those manipulations which
would prove to be the most informative. The compilation
of discrete routines for the performance of these elementary manipulations would provide the basis of a flexible
system for simulating many widely diversified pattern
identification logics. After determining the intended course
of his pattern analysis, the programmer needs only to refer to this file and to select those routines which in combination will best serve his purpose.
Furthermore, it should be possible for the analyst to
sit at the computer console and to draw from this tape
library several routines which he will select pragmatically
after studying the results of any preceding operations, and
thus guide the computer step by step toward recognition
of the pattern being studied. We distinguish two forms
of output in these routines-numerical data and transformed pictures. Numerical data can be read directly from
the computer but pictorial information must be converted
before it can be displayed by the picture output scope. The
picture display routine is used in such cases.
One of the simplest routines in the library counts the
total black area in a pattern. This program examines each
bit in sequence and tallies when the bit represents a black
area. Advantage is taken of the fact that in many patterns
there will be numerous words that are all black or all
white. By comparing whole words against constants of all
zeros or all ones, much time is saved. The area counting
process requires approximately thirty seconds on SEAC.
Pictorial information is an extremely informal sort of
information to feed into a computer; it is not stylized in
the same sense as numerical or alpha-numerical information. This is one of the fundamental differences that must
be faced in operating upon pictorial information. For one
thing, pictorial information often contains a good deal of
redundancy. The pictures obtained from the SEAC picture
input equipment require about 30,000 bits of computer
storage. However, the total number of bits of information
in the pragmatic sense of the term is probably somewhat
less than 100. In view of this fact, a routine was written
to provide the data necessary for efficient encoding of pictures. The routine analyzes pictures and tabulates the number of runs of each length of continuous black or white
points as they appear in the picture. On the basis of these
run lengths it is hoped that we will be able to determine an
efficient method of statistical encoding for these pictures.

Fig. 4-"Blobs" counted by SEAC.
Object

No.

Area

(0.25 rom squaresl

l
~

:«

•

163

142

181

206

404

Fig. S-Areas for each of the objects in Fig. 4.

One possibility is the use of the so-called Shannon-Fano
codes [5] - [8] which assign symbols of varying lengths to
the different run lengths to minimize the total code length
required for any given picture. Preliminary investigation
shows that code compressions of the order of at least from
6 to 10 times are possible.
Another routine counts the number of separate noncontiguous black objects (blobs) and measures their separate
areas [9]. Due to the restricted memory capacity of
SEAC, the routine cannot analyze a full-size input image.
Therefore, the image is compressed in the horizontal direction by a factor of four. The routine is devised to scan
sequentially through the image until a black point is found,
then to move systematically through the blob, cQunting its
points and erasing them until the blob is completely removed f rom the image. Because the compression in one
dimension is not linear, the computed areas are only approximately proportional to the original areas. The blobs
are traced so that objects with re-entrant profiles and nonsimply connected objects will be recognized as single objects. The time required for SEAC to count blobs in an

Kirsch et al.: Experiments in Processing Pictorial Information with a Digital Computer

225

(b)

(a)

Fig. 6-(a) Letter L with center of gravity.
(b) Letter L translated.

Fig. 7-Boundary of letter L.

image depends upon the number of blobs and their total
areas. An example of the use of the blob counter on the picture of Fig. 4 is shown in Fig. 5, on the preceding page. The
area of each blob has not been corrected for the factor of
4 compression.
A second blob counter, which operates with full precision but does not give the areas of the individual blobs, has
been planned, but not yet written. This routine examines
each bit and its immediate neighbors, generating a tally of
each new occurrence of a blob and adjusting when parts
of blobs merge or diverge. N onsimply connected blobs require special checking. There is no limit to the size of the
image that can be handled by this routine. The time required on SEAC to process any single pattern from the
scanner would be one minute.
Another simple code computes the center of gravity of
a pattern and translates the pattern rectilinearly so that the
center of gravity is at the midpoint of the image. Fig.
6(a) shows a pattern before translation and Fig. 6(b)
shows it after translation. The boundary of the square and
the center of gravity are shown as bright spots in the pictures. The translating routine can use any given set of
coordinates to determine the shift.
Fig. 7 illustrates the result of a routine which computes
what might be called a "first derivative" of the pattern.
Each 3-bit square is examined; if all nine bits are black

(a)

(b)

(c)

(d)

Fig. 8-(a) Letter L custered and complemented two times. (b)
Letter L custered and complemented four times. (c) Letter L
custered and complemented eight times. (d) Letter L custered
and complemented sixteen times.

the center bit is replaced by a white bit. The computer operates on three rows at a time, forming logical products of
all combinations of each bit with its adjacent bits. These
products are, in turn, logically multiplied to produce the
final result in which black bits remain only when one or
more neighboring bits are white. The effect of this socalled "custer" operation is to preserve the boundaries of
a pattern and to erase all the internal areas [10]. Fig. 7
shows the result of computing the boundaries of the letter
in Fig. 6. Notice that the boundaries contain most of the
significant information of the original picture but require
fewer bits. The time required for one "custering" operation is seven seconds.
I t was suggested that a thin line representation of certain patterns could be obtained by computing the boundary, reversing the image (i.e., forming the binary complement) , recomputing the boundary, etc. Upon consideration it became obvious that this treatment would not produce the proposed result, but the idea aroused considerable
curiosity as to what it would produce. As a result of performing this process, it was found, in fact, that the image
became more complex and difficult to analyze. In Fig. 8
there are four pictures of the "custered" letter L of Fig. 7
after it was "custered" and complemented a varying number of times. It is interesting to note that the small dust

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

226

(b)

Fig. 9-(a) Letter P with random noise.
(b) Expanded section of (a).

(a)

(b)

Fig. 10-(a) Fig. 9(a) custered and complemented.
(b) Expanded section of Fig. lO(a).

speck in the lower left-hand corner of Fig. 7 grows much
larger in successive pictures, and merges with the L in
Fig. 8 ( d). The next set of figures shows a very surprising result of the custer and complement operation. A program was first written which superimposes controlled random noise on a photograph of a letter P to produce Fig. 9.
This noisy letter was then custered 50 times to produce
Fig. 10. The "noisy" letter in Fig. 9 is still clearly recognizable in Fig. 10 even after being operated on 50 times.
Thus, the "custer and complement" is a process which
seemingly preserves the information in a "noisy" picture
but degrades the information in a "clean" picture.
In addition to these routines, some other simple routines
have been written. These include routines to superimpose
pictures, to smear pictures by translating and superimposing them, to magnify pictures so as to make their fine detail visible, to record pictures in permanent form on tapes
and wire, and a routine to analyze a set of pictures to determine the number of spots that are always black or always white in the set of pictures.
These routines perform elementary operations, although
in some cases, such as "custer and complement," the results of the operation have far from trivial explanations.
Some applied investigations have been undertaken to
utilize the potential of this new equipment. Such problems as character recognition, aerial mapping [11], and

automatic encoding of the chemical information in chemical-structure diagrams [12] are included.
Much interest has been developed in simulating character recognition logics by some of the preliminary experiments such as those of Greanias, et al., of IBM [2]. The
present picture scanning equipment adds to the practicality
of simulating recognition logics with computers by providing a convenient source of data input. We plan to simulate
some of the logics that have been proposed such as the
"peek-a-boo" system, in which characters are recognized
by locating the key points in a picture that characterize
letters uniquely.
An attempt is being made to program the computer to
generate elevation contour lines from aerial photographs.
The relative elevation of ground points can be determined
from two aerial photographs of the same area taken from
different points in space by measurements of the apparent
displacement of the elevated points. Some simplifying restrictions were made to reduce the problem to a convenient
size for a first attempt. The photographs to be used were
assumed to have been restituted to correct for tilt of the
focal plane and oriented so that the overlap or shift of
ground points occurs only in the x direction. It is much
less involved to deal with one-dimensional shifts than the
two-dimensional ones.
The contours are determined in three steps. First, similar areas on the two photographs are identified by checking
units of 44 bits for the correspondence of at least 39 of the
pairs of bits. Groups of bits that are mostly ones or zeros
are not compared. Next, the pictures are shifted with respect to each other to bring them into alignment. Finally,
all points with the desired parallax are computed and
stored in the contour diagram picture. The first two steps
of this program have been completed. The parallax determination and the preparation of output photographs remains to be done.
Another intriguing problem is to find a way that a machine could "look at" a diagram, such as a chemical structure diagram, and characterize it uniquely. The work to
date has not been concerned with the more symbolic information that appears on structure diagrams, such as element symbols, double bonds, etc. We have attempted to
treat only simple nets composed of vertices and bonds
drawn between them. The connection pattern has been
treated as a topological net and we are not concerned with
such things as size of angles, length of lines, width of
lines, and line breaks. The program we have been working
on will first locate most of the vertices by counting the
number and extent of clumps of "black" spots in each
line of a picture in both vertical and horizontal tracings.
Where these numbers change between successive lines a
vertex is indicated. Then, starting at a vertex the bonding
pattern could be traced from vertex to vertex. Thus far,
the programming is in a preliminary state. The actual coding and handling of the "housekeeping" procedures remains to be done.

Kirsch et al.: Experiments in Processing Pictorial Information with a Digital Computer

IV.

SOME UNSOLVED PROBLEMS

Thus far the discussion has been concerned with a report of experimental investigations. To those familiar
with the application of data processing techniques to new
fields, it should be apparent that such experimental investigations generally lead to the formulation of new problems in the two areas of higher performance equipment
design and proper utilization of such equipment. The
problems in automatic processing of pictorial information
that occur in these two areas will be formulated here. To
the knowledge of the authors, no solutions to these problems are available.

A. The Development of a High Performance Picture
Scanner for Computer Input
In using a digital computer to process pictorial information, it is unthinkable that any large quantity of pictorial information should be scanned and stored on conventional computer storage media like magnetic tapes unless a tremendous amount of reduction of information has
first taken place. The maxim that "one picture is worth
ten thousand words" is probably overly optimistic for
fairly common sources by about three decimal orders of
magnitude, if a "word" means 25 to 50 bits.
This estimate is based on a comparison of the number
of binary digits needed to describe highly stylized information in pictorial form and in such a form as to describe
only the "meaning" conveyed by the picture. This means
that in applications where it is either not possible or not
practical to encode (and thereby reduce) information in
a photograph, the best way to store a photograph is in its
original form.
However, to make such information available to a computer, equipment is needed that will mechanically handle
photographic information and that will be able to sense
the information for input to a computer. The mechanical
handling can probably be solved in many ways. Such devices as a microfilm rapid selector might be appropriate.
For the optical scanning of the photographic information, however, it appears that a device with performance
somewhat better than conventional cathode-ray-tube flying spot scanners is required. If we assume an average
document size of 8 X 10 inches then the scanner must be
able to resolve a field of the order of 103 X 103 = 106
spots. Although it is seldom necessary to scan a whole
pictorial source with this resolution, any section of a picture must be capable of being resolved with this precision.
The data rate for a computer like SEAC should be 1
m~gacyc1e. Thus, if it is necessary to scan a whole field,
it should be possible to scan the 106 bits in one second.
It is anticipated that with such a scanner, the computer
would first direct the scanner to locate information. The
most straightforward way to do this would be to have a
de focused scan or perhaps several different levels of defocusing. Thus a field of 210 X 210 bits might first be
scanned with a raster of 32 X 32 spots. Since each spot

227

would cover in turn an area of another 32 X 32 elementary spots, it would be necessary to be able to get something of the order of a 10-binary-digit reading of the light
value from any spot. It would be unreasonable to expect
an accuracy in such a reading, but such precision would be
required. In other words, upon successive scans of the
same large square array of 32 X 32 elementary spots, the
same reading should be obtained within 1 part in 2 10, even
though the reading itself is not accurate. The inaccuracy
can be compensated in the computer programs.
In addition to the performance of the actual scanner, it
would be desirable that the computer be able to specify
certain types of front-end processing which would be done
by suitable fast analog equipment, e.g., time differentiation of the scanning signal and insertion of logarithmic
response functions. Certain simple analog operations performed on the scanner signal can save a great deal of complex processing by the digitai computer.
With a scanner such as this and with suitable mechanisms for motion of the photographs, it would be possible
to get the photographic information into a computer at a
rate comparable to the present processing rate of the computer. The next problem would be how to use this information.

B. The Effective Use of a Picture Processing Computer
It was stated at the outset of this paper that an aim of
the present investigations is to automate some of the visual
processing of information done by human beings. The beginning of the processing operation is, logically, the statement of requirements. Therein lies the rub ! We can state
our requirements to a human being and expect some intelligent performance but we do not know how to do this
with a computer program. There seems to be fairly universal agreement among people as to what constitutes a
picture of an automobile, a letter "e," or the President of
the United States. This ability to recognize patterns is not
learned the way the ability to multiply numbers is learned.
Most of the pattern recognition ability of people exists on
a nonarticulate level. In order to program a computer to
duplicate this pattern recognition ability, it is necessary to
make explicit the techniques that people use and then instrument these techniques in computer programs [13].
To accomplish this we require an automatic program:...
ming technique in which macroscopic patterns can be defined in terms of more simple ones [14]. This type of technique which would assume the nature of an automatic
compiler would eventually enable a programmer to describe familiar objects in terms of other more simple but
nevertheless familiar objects. Thus a linguistic formalism
would be constructed which would continue to approach
closer at successive levels of approximation to the formalism used by people in describing pictorial information.
Experiments are being initiated at NBS on the construc:"
tion of such a compiler, however, there are no results
available yet.

228
V.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE
AREAS

ApPLICATION

PICTURE

PROCESSING

TECHNIQUES

Much of the work described in this report was motivated by the promise of application in the solution of important problems. Other areas of application have suggested themselves as the work progressed. We discuss below those classes of applications.

A. Analysis of Pictures
In this class of applications it is desired to subject a
picture to an analysis, the result of. which is to produce
some alphabetical or numerical data. The picture itself has,
in principle, no value for retention after the analysis. It is
desired to abstract some information from the picture and
store this.
The first application in this area occurs in information
retrieval of the type practiced in the U.S. Patent Office.
Documents containing drawings and schematics are to be
stored for purposes of subsequent reference by a computer. If a drawing can be coded so that an equivalent one
can be reconstructed from the code, this is sufficient for
storage purposes. Obviously, such pictorial considerations
as the quality of the lines in a circuit diagram need not be
preserved in the code. Thus we are led to the attempt to
recognize by machine such configurations as chemical
structure diagrams, electronic circuit schematics, and
drawings of mechanical configurations. These problems
show promise of yielding to the type of investigations described here.
Within the same class of analysis applications fall the
problems that involve counting objects in a picture. Here
we have such cases as the counting of particles in microphotographs of metallic structures, classification of particles in biological preparations, and analysis of tracks of
nuclear particles. In the area of astronomy, knowledge
comes mainly from photographs taken through telescopes
or other instruments and the analysis of these pictures now
requires considerable time in order to generate a rather
large body of data. Picture processing techniques might
be used for such problems as computing star positions and
proper motions, evaluating star brightness or magnitudes,
and automatically setting up star catalogs.

B. Transformation of Pictorial Information
In this class of applications, information is to be prepared for visual consumption by human beings. Generally,
the information is originally in such a form that it cannot
be used by human. beings. The question will be left unanswered whether human beings may be replaced by automatic processes as visual consumers of the information
produced in the applications in this class.
The first such application occurs in photogrammetry.
A stereo pair of aerial photographs is to be processed to
produce an elevation contour map. By techniques based
on principles described, it is believed possible to use a digital
computer to generate elevation contour maps. If investiga-

tions in pattern recognition proves successful, it may even
be possible to superimpose cultural information upon maps,
the whole process occurring automatically.
Another application of these techniques to the transformation of pictures was suggested to the authors by M. L.
Minsky of M.LT. Picture processing techniques could be
used to develop a good reliable set of photographs for the
planet Mars. We know that one of the main reasons that
a good photograph of Mars doesn't exist is that there very
rarely are conditions of perfect seeing where the entire
disk of the planet is clearly visible and all of the details on
it are plainly visible.
There are, however, several million frames of motion
picture film that were taken of the planet Mars during its
opposition. By an analysis of these photographs, abstracting those bits that represent good clear seeing in anyone
frame and putting them together to form a composite photograph of the disk of the planet, we may be able to get a
good reliable map of the true features of the planet.

C. Simulation of Picture Processing Systems
In this class of applications, the digital data processing
system in conjunction with its picture input and output is
used to simulate the behavior of a system or the model of
a system that processes pictorial information.
The most obvious use of such simulation techniques
occurs in applications to character recognition studies.
Fairly complex character recognition devices can have
their behavior simulated by a data processing system. In
this way devices can be "flown on the ground" without the
necessity of costly construction of apparatus.
A more unusual application of such simulation techniques occurs in the field of experimental psychology. In
attempting to explain human· vision, theories have been
propounded which are subject to analysis by simulation
techniques. Many operations that have neurophysiological
counterparts can be programmed on the type of research
facility described here [15], [16]. It is to be hoped that
the eventual use of general purpose picture processing simulation techniques will aid in encouraging the formulation of more ambitious theories of the functioning of the
human visual process.

VI.

CONCLUSIONS

In this paper a new type of research facility has been
described which allows complex investigations to be made
into the nature of· pictorial information and into ways in
which computers may be programmed to process such information with the same comparative ease that characterizes human processing of visual information.
The apparatus described here as well as the experiments performed are strictly of a research nature. Consequently no conclusion should be drawn as to the practicality of such processing as is described here. Before such
applications become practical it will be necessary at least
to solve the type of problems described here, namely, those

Kirsch et al.: Experiments in Processing Pictorial Information with a Digital Computer
of the design of high performance scanning equipment
and the development of automatic techniques for the recognition of visual patterns.
The applications of automatic pictorial information
processing techniques described here are not meant to represent the most important applications that can beanticipated. They are, rather, meant to illustrate typical classes
of processing techniques that can be automated if experiments of the type outlined here lead to successful conclusions.

VII.

ACKNOWLEDGMENT

The work described in this paper has largely resulted
from a project sponsored by the Rome Air Development
Center, USAF, and has also been supported in part by a
cooperative U.S. Patent Office-National Bureau of
Standards program to explore the potential mechanization
of patent search operations. The results reported have
been achieved through the contributions of a number of
individuals. In particular, J. L. Pike and M. A. Fischler,
helped to design, build, and debug the scanner. This work
was asisted by G. Crowther and O. Hall. Some of the experimental SEAC picture processing programs were written by R. B. Thomas. The interest and suggestions of
several other members of the Data Processing Systems
Division were also of assistance in the formulation of
some of the ideas expressed here.
BIBLIOGRAPHY

[1] Glauberman, M. H. "Character Recognition for Business
Machines," Electronics) Vol. 29 (February, 1956), pp. 132-136.

Discussion
V. M. Wolantis (Bell Telephone
Labs.): Do you find it satisfactory to distinguish only between black and white
rather than a number of shades of gray?
Mr. Kirsch: No, that is often not adequate. For purposes of looking at schematic diagrams or printed characters, two
values of light intensity are adequate.
However, for looking at the most interesting types of pictorial information,
people's faces and so on, we should like to
distinguish several shades of gray. What I
showed you was a poor compromise between having full gray scale rendition and
h~vin,\ o'lly the black and white rendition.
In the photograph of the person's face,
what we were doing was sacrificing some
of our resolution in order to get some gray
scale information, but certainly ideally we
would like to have considerably more gray
scale information.
Mr. Wolantis: To what extent would a
TV camera help solve the problem you
mentioned near the end of your talk?
Mr. Kirsch: A TV camera would help.
We would like to scan a million-bit pic-

229

[2] Greanias, E. c., Hoppel, C. J., Kloomok, M., and Osborn, J. S.
"Design of Logic for Recognition of Printed Characters by
Stimulation," IBM Journal of Research and Development)
Vol. 1 (January, 1957), pp. 8-18.
[3] Rabinow, ]. "Report on DOFL First Reader,'" Diamond
Ordnance Fuze Laboratories, Washington, D.C., Report TR128, November 26, 1954.
[4] Stone, W. P. "The Alphanumeric Character Reader-RADC,
ARDC, USAF," RADC-TN-56-194, AD 103-237.
[5] Fano, R. M. "The Transmission of Information," Research
Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, Mass., Technical Report No. 65, 1949.
[6] Huffman, D. A. "A Method for the Construction of MinimumRedundancy Codes," PROCEEDINGS OF THE IRE, Vol. 40 (September, 1952), pp. 1098-1101.
[7] Michel, W. S. "Statistical Encoding for Text and Picture
Communication," AlEE Conference Paper CP 57-723, May 1,
1957.
[8] Shannon, C. E. "A Mathematical Theory of Communication,"
Bell System Technical Journal) Vol. 27 (July, 1948), pp. 379473 (October, 1948), pp. 623-656.
[9] Dell, H. A. "Stages in the Development of an Arrested Scan
Type Microscope Particle Counter," British Journal of APplied Physics) Supplement, No. 3 (1954), p. 156.
[10] Kovasznay, L. S. G., and Joseph, H. M. "Image Processing,"
PROCEEDINGS OF THE IRE, Vol. 43 (May, 1955), pp. 560-670.
[11] Strand, O. N. "Mathematical Methods Used to Determine the
Position and Attitude of an Aerial Camera," U.S. Ordnance
Test Station, 1585, NAVORD Report 5333, March, 1956.
[12] Ray, L. C. and Kirsch, R. A. "Finding Chemical Records by
Digital Computers," Science) Vol. 126 (October 25, 1957),
pp.814-819.
[13] Selfridge, O. G. "Pattern Recognition and Modern Computers," Proceedings of the Western Joint Computor Conference (March, 1955), pp. 91-93.
[14] Dinneen, G. P. "Programming Pattern Recognition," Proceedings of the Western Joint Computer Conference (March,
1955), pp. 94-100.
[15] Clark, W. A., and Farley, B. G. "Generalization of Pattern
Recognition in a Self-Organizing System," Proceedings of
the Western Joint Computer Conference (March, 1955),
pp. 86-91.
[16] Farley, G., and Clark, W. A. "Simulation of Self-Organizing
Systems by Digital Computer," IRE TRANSACTIONS ON INFORMATION THEORY, No. P61T-4 (September, 1954), pp. 76-85.

ture in a second, but we also want to be
able to deflect to any spot or zone on the
picture with an access time of a few microseconds.
For the purposes of using the picture as
a computer memory, which is after all
what we were proposing that the high
performance scanner do, we would want
considerable reproducibility of scan. We
would want the scanning device to go back
and look at the same picture any number
of times over the course of a machine
computation, and be able to copy the same
information; I don't kno'w whether a
standard TV camera type of scan is capable of that type of performance.
Mr. Rellis: Can you state time required
to execute some of the programs cited?
Mr. Kirsch: Yes. The custer and complement for one cycle takes seven seconds
and thirty seconds to take the area of a
whole picture. To generate the coordinates
of the output point from the information
in the memory for a whole picture of
30,000 bits takes about one minute.
The most time consuming of the routines is the blob counter, which is a func-

tion of the complexity of the image being
blob counted, and here that routine will
take anywhere from about a minute to perhaps three minutes, or even more. However, these numbers are a reflection not of
the intrinsic complexity of the processes,
but rather of SEAC computation time, and
an indication of that speed is that the
SEAC add time is about a quarter of a milIi-second. So you can see that a faster machine certainly would be able to do the
processing more rapidly. It turns out, however, that by way of impedance matching
the operator's thinking time with the computer, this is not really too slow. The machine can test out an idea at about the rate
a person can generate it.
Mr. Rellis: Has any consideration been
given to the similarity between multiplycustered black obj ects on white, and fingerprints?
Mr. Kirsch: Yes, many people have
suggested this, and although we haven't
done any serious investigation, some people
have also suggested that perhaps by doing
inverse custering on a fingerprint, one
might get a picture of the criminal's face.

230

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Optical Display for Data-Handling System Output
JAMES OGLEt

INTRODUCTION

UE TO long experience in the field of business
machines, Burroughs has always been alert to the
fact that human operators are important links in
most computing and data-handling systems. Because of its
versatility, this link fills in where unattended techniques
are still insufficient, where decisions must be made in problems of extremely variable or unpredictable nature, and
where monitoring and manual intervention may be needed
due to functional failure.
One type of display device has been developed at the
Burroughs Research Center for conveying to operating
personnel the output information of data-handling systems. These devices utilize an optical medium known as a
lenticular screen. Such a medium consists of a large number of very small lens surfaces, either cylindrical or spherical,placed substantially in one plane of the lenticular
screen and having a common focal plane in which are recorded in a space-sharing fashion the image elements of
prerecorded messages.
Electronic engineers may be more familiar with the lenticular structure as used in the Lawrence television tube
than with optical lenticular screens. The functional similarities are there, but the methods of design treatment, the
various degrees of freedom, and the techniques of creating
physical embodiments are sufficiently different to preclude
considering the electronic and opticallenticulars as equivalent.
There have been numerous commercial applications of
the lenticular medium. Its first major appearance was in
decorative, dispersing, window glass not uncommon about
50 years ago. Here the cylindrical ribs, 0.1 to 0.5 inch
wide, served the purpose of presenting a repeated series of
prism angles effecting a transverse dispersion. No precision was required. The same principle is used in some
rear-projection screen designs and in single azimuth dispersion of light sources such as sealed-beam lamps.
These applications utilize lenticular screens without
need for any precise focal plane and no image information is carried by the screen itself. The minimal precision
allows some of these to be made of molded glass. In other
uses of the medium, the focal plane of the lenticules has
·an important function and carries prerecorded image information. Here the additional precision usually prescribes
that the lenticular elements be molded or cast in plastic.
We will mention a few; their geometry will become clear
when we describe the principles of lenticular output-data
displays.
t Burroughs Corp., Paoli, Pa.

In one type of three-dimensional photography, the lenticular screen is used for displaying consecutively different aspects of an object as might be seen by walking
around the object, each aspect being visible only to an observer's eye viewing the display at the angle from which
the aspect was photographed. Within certain audience
limitations, the two eyes of the observer can see two different aspects, thus giving true binocular vision of the object. Likewise, a simpler use is that where unrelated images are made to appear consecutively as an observer's
angle changes. These unrelated images may be messages
which appear to flash on and off, eyes that wink, and other
more or less sophisticated material.
BURROUGHS LENTICULAR DISPLAYS

An optical design is usually interesting in that it can be
operated in either direction, in other words, the ray trace
is usually reversible. This is the case in the lenticular
screen. We note immediately that if an illuminator is substituted for the observer in the last examples given, only
that image which the observer was able to see will be illuminated (see Fig. 1). Now if a device is constructed
containing at its rear end a number of lamps or a plurality of filaments, there will be associated with each illuminator discrete areas, one for each lenticule, in the focal
plane of the lenticules. In the case of cylindrical lent icules, these will be narrow lines; in the case of spherical
lenticules, these will be small areas the shape of which will
depend upon the configuration of the filaments. Now if
the focal plane of these lenticules carries a photographic
emulsion and if a stencil or negative is interposed between
a light source and a lenticular screen and if this light
source is made to occupy consecutively the positions of the
filaments in the final device, it will be sufficient to make as
many exposures as there are messages, substituting a new
negative each time the light source is moved to a new position. An appropriate diffusing screen placed over the developed lenticular will allow the illuminated image to be
visible to a sufficient audience.
TECHNIQUE AND DESIGN CONSIDERATIONS

A number of difficulties is encountered in the use of
this technique. First is the art of engraving master-die
surfaces in order to generate with adequate precision and
optical quality the minute surface elements involved.
Another is that of designing a system in which there is,
in general, only one surface to work with. Ordinarily an
optical device, such as a photographic lens, will have a
number of air-to-medium surfaces, thus enabling the optical engineer to introduce the necessary corrections to

Ogle: Optical Display for Data-Handling System Output

231

I-~~I----+I'W
DIFFUSER SCREEN
.-----PHOTOGRAPHIC MASK
rLENTICULAR SURFACE

THIN PRE'iSEO
PL.~STIC 0'" GLA.SS \,",~'Ln..:.

5PHERICUL.A.R
LENS

Fig. I-Geometrical schematic of one-azimuth lenticular digit display. Two of the ten light sources are shown with their associated ray trace and image records. Lenticular screen carries
cylindrical surfaces with horizontal axes.

assure that a substantial number of discrete image points
can be resolved through the desired pupil. We have found
that cylindrically-ribbed (i.e., single azimuth) lenticular
media can be made to resolve up to 10 images with adequate lack of crosstalk for a display purpose.
A practical technique for constructing two azimuth lenticular screens consists of placing two cylindrically-ribb~d
screens face to face with the axes of the cylinder of one
screen at right angles to those of the other. The adjoining
boundaries of the cylindrical elements now define square
celIu(ar pupils within each of which the refractive performance is similar to that of a spherical element (see
Fig. 2).
Two azimuth systems of crossed lenticular structure
can resolve in excess of 20 separate channels.
We have found that the quality of a screen image as
seen by an observer does not yield readily to analysis into
quantitative information of brightness, contrast and crosstalk, and that partly empirical designing and direct observation of models are still the most reliable ways of
obtaining a conclusive ana:lysis of a new application.
The light-handling capacity of these displays is of some
interest because its analysis will largely determine the
class of light source which is necessary today. If we consider the image-record plane, it is immediately clear that
the luminance of any image area can never exceed that of
the source divided by the number of messages. Considering further that the relative aperture of each lenticule is
limited by optical-design considerations, we find that a
further luminance reduction will occur as the outcoming
flux is dispersed in order to satisfy a suitable audience,
(except in certain special cases where a restricted audience can be tolerated). Further, certain clearance margins
in the optical geometry must be allowed. These must take
into consideration the mechanical tolerances of the light
sources themselves, the mechanical tolerances of the assembly, the geometrical optical aberrations of the rudi-

Eq\l"'~ENTS

PHO'TO
EMULSION ---~"'H!-.-J

Fig. 2-Two-azimuth lenticular display. Ray trace through screen
shown for "equivalent sphericles."

mentary lenticules, the phenomena of physical optics
(since these lenticules are usually very small), and the remaining imperfections in the manufacture of the lenticules. When all these factors are taken into consideration,
as well as the absorption factors which are inevitable in a
system involving several surfaces, we find that the only
practical illuminators for applications where the ambient
light is of an office level (30 to 60-foot candles) are tungsten filaments and that further, for practical reasons, these
tungsten filaments should be operated at low enough temperature so as to obtain consistently acceptable life. Typical embodiments specify filament currents from 80 to 200
rna and wattage from 0.5 to 1.5.
LOGIC AND ECONOMY

The information-handling capacity of such a device is
of some interest. If there are N channels or resolved images in each of L "sphericules" in the crossed lenticular
system, then the total information recorded in this system
should be NL bits. Clearly, the same information could
be handled in a system having N obj ectives provided that
each of such objectives was able to resolve L bits of image information. We are familiar with the 35-bit (7 X 5)
minimum matrix for the presentation of an individual alphanumeric character. One also finds immediately that
reading reliability and operator acceptance increase rapidly
if the number of bits per alphanumeric image is increased,

232

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

say by a factor of at least four, and continues to improve
appreciably until good graphic freedom obtains with a
matrix of 20 X 28.
The economical question then becomes one of determining whether for a particular application it is more desirable to utilize N obj ectives (in N small proj ectors) each
resolving L bits, or L equivalent sphericules each resolving N bits. We find that a numerical display will be
slightly more economical when built with a lenticular than
it would be built with a projector battery and that this
advantage increases as the amount of information per
channel increases.
An example of a large information display developed
for a link in the SAGE system is illustrated in (Fig. 2).
There are 20 channels in this display. The used screen
area is six square inches and each channel carries 3600
bits per square inch. This particular display was developed to replace a battery of signal lights and represents,
we believe, a tool which the human engineer can use to
good advantage to satisfy what might be called the system
requirements of a human operator. Interestingly, the human link in a system is the one least susceptible to redesign. With substantial training, an operator can learn to
utilize unfamiliar information codes, but in the final analysis we will always find that, as an information-receiving
center, the operator is a noisy device. It behooves us,
therefore, to address such a device with signals which will
activate the various functional aspects of the individual
and also supply a sufficiently large amount of information
so that these recognition and reaction functions can operate reliability. Thus, one might argue that the signaling of
one out of 20 pre-established messages constitutes relaying
information of no more than five bits for each message,

whereas the fully flexible tool in this instance utilizes over
20,000 bits and that the device is thus extremely redundant. However, when these highly redundant patterns can
be generated basically by one molding operation and one
photographic operation for each message, the potential
economy is self-evident.
Control complexity of the lenticular displays is naturally greater than that of coded displays. In the previous
example some means must be provided to deliver more
than five control lines to the display. One solution is to
apply power to one out of N (in this case 20) leads.
N + I leads must then be brought out of the device. For
maximum application freedom we prefer to bring out
both ends of each filament. This may allow the user to
simplify his decoder. Due consideration must be given to
sneak paths. In general, the glow of a filament in a sneak
path will be negligible if the voltage across the filament is
no greater than one third of the operating voltage. The tolerable sneak path will vary depending upon the type of filament, the type of display and the application requirements.
The individual application should be examined. Relay decoders have been used successfully to drive 10-channel digit
displays where sneak paths included three filaments in
series.
CONCLUSION

The lenticular medium lends itself to space-sharing display purposes in many applications. If the lenticular art is
mastered, good manufacturability can be obtained. Power
requirements are not negligible, the devices being best
suited to low-voltage, high-current applications. Where
electrical requirements can be met, excellent graphic freedom is obtained and this can materially assist the human
engmeer.

Devices for Reading Handwritten Characters
T. L. DIMONDt

N THE LAST five years, much thought and effort
have gone into the development of printed characterrecognition devices. Varying degrees of success have
been achieved. In some. cases, ingeniously distorted type
faces have been required. One might wonder why all this
interest exists. The answer is simple. Character-recognition devices help reduce the substantial cost of getting information into forms that computers can understand.
However, in creating devices that read printed or even
typed characters, we are not reaching back far enough

t Bell Telephone Labs., Inc., Murray Hill, N.].

toward the ongm of the information in the majority of
the cases. Only a little reflection will show that nearly all
of the information used by business data processing computers originates in the minds of humans. What is needed
are methods and devices which will allow these people to
produce, by simple and inexpensive means, the initial expression of their information in a form suitable for machine reading. Without these, there will be many situations, especially where the volume of input information is
large in comparison to the amount of processing, where
computers cannot be proven even if they cost little or nothmg.

Dimond: Devices for Reading Handwritten Characters
We have an example of this input problem in the Bell
System, where toll switchboard operators are producing
two billion toll tickets per year. These are the. records of
long-distance calls handled by operators. They are 20 X 5inch pieces of paper, each containing 20 to 30 characters
of information needed for processing. While there are plans
for ultimately eliminating these tickets by improvements in
switchboards, they will be with us for a long time. Some
idea of the magnitude of this input can be given by stating
that these two billion paper tickets produced each year
would make a pile 200 miles high or, if laid end to end, a
strip 150,000 miles long. What is more significant, it is
estimated that it would cost about $32,000,000 per year to
transcribe this information to cards by means of keypunchmg.
A broad look at possible methods by which humans can
communicate with machines, including computers, reveals
the following situation.
First, the human can communicate by physical actions
(generally involving the fingers) on keys, levers, dialSy
etc. Of these, the telephone dial is undoubtedly the one
used in the greatest number. On the other hand, the key
is used on a greater variety of machines, including typewriters, teletypewriters, calculating machines, keypunches,
and switchboard operator key sets. Second, the human can
communicate through physical action which produces a
document without the intervention of a machine (disregarding the pencil). This document, in turn, is used to
control a machine. Mark-sense cards exemplify this
method. Third, it seems probable that it will be possible
some day to produce machines which can interpret the
human voice reciting numerals and letters. Possibly we
may half facetiously suggest that ultimately the human
mind can directly control machines.
When one examines the methods just mentioned in comparison with handwriting, one must conclude, however
reluctantly, that it is pretty hard to beat handwriting as a
ready, economical, fast, and accurate means of expression.
Consequently, this discussion deals with two different
methods by which handwritten characters may be read. The
first falls in the category, mentioned above, of control by
physical action and involves a new device which permits
real-time communication with machines as characters are
written by a stylus. The second falls in the category of
communication through documents, and consists of simple
methods and devices by which handwritten characters can
be automatically read.
Let us consider the problems encountered in automatic
recognition of handwritten characters. To simplify the
discussion, it will be confined first to numerals. Of course,
the problem can be greatly simplified if it is permissible
to adopt an entirely new set of characters created specifically for easy machine reading. For example, characters in
the set of Fig. 1 could easily be recognized by a machine
scanning vertically and horizontally. The patent literature
discloses many. .such sets of symbols. They have the obvious and common disadvantages that writers must learn

233

them and become proficient and accurate in their use, and
that they cannot be understood by the uninitiated who
occasionally come in contact with them. Personal experience indicates that it would be very difficult to persuade
people to adopt them and that promoters of such systems
are viewed as enthusiastic but misguided.
Mark-sense marks cannot be considered as special symbols because it is the mark position rather than the shape
that carries the information. Mark sensing has the disadvantages of occupying considerably more space than
ordinary numerals and of being slow for humans to read.
If the idea of special sets of symbols is rejected, nothing remains but regular Arabic numerals. The problems
which are encountered in reading these will now be surveyed.

ARABIC
I

SPECIAL

2
3
4

II
III

5
6
7
8

::j::

Fig. I-Special handwritten symbols for machine reading.

Fig. 2-N umerals written without constraint.

Fig. 2 shows a collection of Arabic numerals chosen
from those produced by a random group of people. An
examination of these forces the conclusion that some degree of control must be placed on their writing not only to
enable machine reading but to reduce "sloppiness" which
makes even human reading difficult. The variable factors
which must be dealt with by an automatic reading device
are location, size, orientation, and shape.
Location is important, if for no other reason, because
it generally defines the meaning. For example, a given
number appearing in one place on a form may indicate
revenue and in another, expense. Also, of course, the machine's problem is greatly eased if it knows exactly where
the numeral is located. It may be remarked that while the
information content is in the shape rather than size or
orientation, nevertheless the machine must be able to recognize shape in the presence of variation in size and orientation.

234

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

"
~

Fig. 3-Cartesian-coordinate grid for character recognition.

/:l.J':/5
~7~<:!O

Fig. 4--Numerals with dot constraint.
A

Fig. 6-Range of variation permissible.
A

F't-B
E

ALLOWED
CONFI.GURATIONS
I: I:

CRITERIAL AREA
A

0/1
0/1

...,.

1ft

~ ~

011

'Y ;r "l1

Fig. 5-Set of bipolar coordinates for character recognition.

Any plan to control writing must take into account the
methods by which the symbol is to be recognized. The
method of recognition, generally proposed, is to examine
the character in a field of Cartesian coordinates as in
Fig. 3. The presence or absence of a mark in each rectangular cell is indicated by a television-scanning technique to a computer which can operate on the information
with all its considerable resources. The cells mayor may
not be contiguous. Study of this plan indicates that unless
very rigid writing controls are employed, this is a hard
and expensive way to recognize characters.
At this point we may conclude that these things are
needed to read handwritten numerals automatically and at
reasonable cost: 1) a means of constraining writing which
does not seriously affect writing habits, 2) a mode of machine examination of the symbols, under which the symbols appear invariant with reasonable changes in location,
size, orientation, and shape, and 3) compatibility between
1) and 2).
A simple solution is now described which encompasses
these three needs. First, the constraint is provided by
means of two dots around which the numerals are written
as shown in Fig. 4. The naturalness and ease of this
method are obvious from the figure. Second, in order for

tBlill

011

0/1

011

0/1

011

0/1

011
011

011

011
0

0/1

"ta CJ1?
(J (:J

o
o

0/1

011

011
0/1

0/1

Olt

011

Fig. 7-Truth table for numerals.

the numerals to appear invariant, the machine examines
them in two polar coordinate fields with the two dots as
origins. The machine is able to recognize the numeral by
sensing which of the radius vectors in the particular set
of Fig. 5 are traversed by the lines making up the numerals. (The two left-hand vectors are moved out of their
horizontal positions to avoid the ends of 3's and 5's.) The
use of the same dots both for the origins of the polar coordinate sets and for controlling the writing makes it
possible for simple machines to recognize numerals even
though they vary quite widely in location, size, orientation and shape. In Fig. 6, the numeral 3 is shown to be
invariant with a wide range of these four variables.
It remains to be shown that the set of radius vectors
crossed by each numeral is unique. This is done in Fig. 7
in which a binary 1 indicates a necessary cross, a binary 0,
a necessary noncross, and 0/1, indifference to a cross. A
point to be noted in the left-hand column of this figure is
the considerable tolerance of this method for the vagaries
of humans. For example, 1 may be written either to the
right or left of the dots since the associated transversals
are each unique and can both be interpreted as 1. The nu-

Dimond: Devices for Reading Handwritten Characters

COORDINATE SYSTEM

~••

Jr• • c;••

~•••
•~rP
•

·-(;J
-z.
•
.:

c:-• •
~.

••

235

[••
I·• ·• j.
••
:ts.j :C)
••

••
~

6)
••

l[
••

F.f
••

t::I• •

••

••
~

••

1.\ (

tJ.

,..

I;:
•• H
.$ :J:
:Q .JR
••
• •
• •
:x: •.'%•
'V: ~
••

I?••

·r tJ ti X
••

:$
=st

"l.

•

Fig. 8-Four-dot letter restraint-method l.

Fig. 9-Four-dot letter constraint-method 2.

meral 7 may be written with the vertical leg either between
the dots or to the right. However, a closed top 4 cannot be
permitted because it will appear the same as a 9. There are
many different truth tables that could be devised following
the philosophy indicated in Fig. 7 but trading among allowable variations in the formation of the different numerals.
Still others could be designed by adding a "goof" detection
feature which detects combinations not corresponding to
any numeral. The logic for the latter would be quite extensive, however, because there are 27 or 128 possible combinations of which a minimum of 10 are legitimate.
So far only Arabic numeral recognition methods have
been disclosed. The question naturally arises as to whether
the basic methods of controlling writing by dots and invariance in polar coordinate fields can be extended to include letters.
There are several ways of accomplishing this result.
Two will be discussed briefly. The first, involving four
dots, is shown in Fig. 8. The basic idea here is that the
first half of the alphabet is written about the left two dots,
and the rest, with a few exceptions, about the right two
dots. It will be noted that with a few exceptions which are
necessary to attain uniqueness of each character, all the
letters are regular upper-case block or drafting type. H is
lower case to avoid confusion with K. G and Q are somewhat specially formed. In passing, it should be noted that
block letters seem preferable to script because script writing is more likely to be undecipherable even by humans.
Evidence of this is the statement, "please print," onthe job
applications which some of you have filled out.
Another set of characters is shown in Fig. 9. Here ad-

vantage is taken of the fact that 13 of the 26 letters begin
with a vertical line, while 13 do not. In this embodiment
the letters G, K, Q and Ware slightly peculiar. 1
To mechanize any of the logical methods described
above, it is only necessary to devise a machin~ which can
detect marks in the long, narrow areas (her~after called
criterial areas) corresponding in position to the radius
vectors. An obvious way of doing this is with electro-optical scanning. Alternatively, a very simple reader can be
made by providing a sensing head made with printed wiring as shown in Fig. 10. In this figure, each criterial area
is made up of one long, narrow conductor connected to a
source of potential, and another, parallel to it, used as a
sensing element and connected to a translator. When the
head is properly placed on a piece of paper on which a
numeral has been written around dots with a conductive
lead pencil, the mark on the paper closes circuits between
the two parts of each of the criterial areas 2 which it crosses.
Hence, certain of the seven leads from the seven criterial
areas will be energized causing the translator to energize
a different output lead for each different character.
A translator using transistor logic and based on the
truth table of Fig. 7 is shown in Fig. 11. The input leads
A to G connect to the seven criterial area conductors similarly designated at the top of Figure 7. The RS lead connects to a contact which is closed to reset the translator
1

Proposed by Dr. L. A. Kamentsky of Bell Telephone Labs.,

Inc.
2 Subsequent to the presentation of this paper,
U. S. Patent
2,741,312 issued to R. B. Johnson was called to the author's attention. It discloses the use of the two dots and of the radial areas for
sensing conduction through the mark constituting the numeral. Relays are operated to control a card punch.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

236

~'.TENT"L
TR.ANSLATOR

TRANSLATOR

0125456789
__-C:;;""'O='"""="~--.o POTENTIAL

Fig. 12--Stylator.

Fig. IO-Reader.

F
D
A

c
G

RSo------.

AND GATE

~ OR GATE

Fig. ll-Translator circuit.

by restoring flip-flops F/F-A to FJF-G to a normal condition in which they energize the leads designated with a
zero sUbscript. The recognition of the numeral 2 will now
be followed. Conductors A, B, D, E, and G will have been
energized by conduction through the conducting pencil mark
although the energization of E is immaterial to correct
recognition of the numeral 2. Flip-flops of corresponding
designation energize leads of subscript 1. The end result
is that leads A 1 , B 1 , D 1 , E1 and G1 are energized as well as
leads Co and F o. Leads D1 and F 0 open gate 4. Gate 12 is
opened by the output of gate 4 and lead A 1 ; gate 18 by
gate 12 and lead B 1 ; gate 24 by gate 18 and lead Co; gate
30 by gate 24 and lead G1 • Gate 30 energizes lead 2 to indicate recognition of that numeral. The flip-flops are not
strictly necessary, but they aid by furnishing ample power
to drive the logic circuitry.

As mentioned before, a method will be described which
permits automatic recognition as the characters are written.
A novel device for performing this function for numerals is
shown in Fig. 12. A writing surface is provided on which
there are two guide dots surrounded by a set of criterial
areas consisting of seven conductors embedded in a plastic
plate. As a numeral is written with a stylus connected to a
source of potential, the stylus energizes, one at a time, the
conductors in the criterial areas involved in the numeral.
The combination of areas energized causes certain flip-flops
in a translator such as that in Fig. 11 to operate and drive
the rest of the translator to indicate the correct numeral.
The flip-flops are necessary because the criterial areas are
not all energized simultaneously. Alternatively, seven relays
may be used to replace the electronic translator of Fig. II.
This device has been tentatively christened a Stylator, meaning stylus translator or interpreter.
The problem arises in connection with the Styl(J)tor of
informing it when the writing of a character has been
completed. This is necessary because in some cases the
character changes from one to another during the writing
process and because the translator must be returned to
normal before a new character is written. A simple way of
incorporating this feature is to provide another conductor
in the writing plate which, when touched by the stylus,
causes the memory and translator circuits to return to normal as soon as the character already written has been recorded. This conductor may extend around the whole
perimeter of the plate so that it can be touched by a continued stroke of the stylus.
In the devices just described, no advantage is taken of
the information residing in the sequence in which the criterial areas are crossed. Of course, this information cannot be recovered from a character already written but it
is readily available in the case of the Stylator. This added
information is so meaningful that the two-dot system can
be used for letters as well as numerals.
Several uses have been suggested for the Stylator. It is
a competitor for key sets in many applications. It has been
successfully used to control a teletypewriter. It is attractive in this application because it is inexpensive and does
not require a long period for learning to use a keyboard.

Dimond: Devices for Reading Handwritten Characters

237

Stylator.
1) Clear by touching the stylus to the small area at the
lower right.
2) Make open. top 4's.
3) Keep ends of 3's and 5's out of area between segments E and F in Fig. 5.
CONCLUSION

Fig.

13~N umber

reader and Stylalor.

If the criterial areas are used to control the frequency of
an oscil1ator, an inexpensive· sending device is obtained
which may be connected to a telephone set to send information to remote machines.
Fig. 13 shows a combined number reader and Stylator
which can successively read four separate numerals from
a sheet of paper as well as recognize numerals as they are
written. The following set of rules will help in using the

Discussion

E. A. Etling (RCA Service Co):
What thought, if any, have you given to
the development of a system which places
no constraint on the handwriting of characters, other than broadly defined statistical bounds?
Mr. Dimond: The polar scanning technique can be extended to systems involving
different methods and degrees of constraint. I think that some degree of restraint is desirable because it requires
people to form their characters more carefully than they would if they followed
their normal habits, which may be so
sloppy that the characters are sometimes
unrecognizable even by humans.
P. Hersh (General Ceramics): How
does the Stylator distinguish between the
"early" versions of a character and the
final (correct) one?
Mr. Dimond: In these demonstration
machines, a separate segment is touched
by the tip of the stylus after writing is
complete to cause the information stored in
the translator circuitry and indicated by
lamps to be wiped out. In a commercial
machine, the touching of the segment
would first cause the information stored in
the translator to be transferred to some
sort of memory and would then restore
the translator to normal. To minimize the
effort required, the segment could consist
of a border surrounding the platen. It could
then be touched with the stylus by continuing the last stroke in writing a character.

Methods of constraining the writing of characters for
machine reading and machines for reading such characters
has been described. A new device called a Stylator has been
disclosed which permits real-time recognition of characters
as they are written on a platen.
ACKNOWLEDGMENT

The author wishes to mention that Dr. L. A. Kamentsky of the Bell Laboratories designed the first logic
circuits, constructed the first model and furnished many
valuable ideas as well, and that W. W. Gulden of the Cincinnati and Suburban Bell Telephone Company designed
and constructed the small model shown in Fig. 13.

M. J. Stoughton (Sears Roebuck):
\Vhat controls are you contemplating for
reducing operator errors?
Mr. Dimond: Errors may result either
because the operator writes a wrong number or because she forms it incorrectly.
Errors of both sorts may be reduced by
training. Nothing can be done in the design of the machine to prevent entirely the
former. In this respect, the problem is the
same as with a keyboard. In some cases
the error can be detected by a system of
control totals.
There are some things that can be done
to minimize errors due to incorrect formation. More criterial areas and a more
able translator would help. If the degree
of accuracy required justifies it, mentally
computed check numbers can be used. Suppose, for example, the number 13 is to be
recorded. The operator also records 24
which is obtained by adding 1 to each digit
of the number 13. At some later stage the
two numbers are automatically subtracted
to check that a difference of 11 is obtained.
L. C. Oesterich (US. Navy): You are
apparently adapting this reading device to
your toll ticket problem. Please outline the
system.
Mr. Dimond: The proposed reading device is one of the contenders for solving
the toll ticket problem. It would be used
in the following manner. The tickets furnished to the operators would have dots
preprinted on them around which the operators would write the characters, most
of which are numerals. These tickets would
be gathered up periodically and sent to a
processing center where they would be fed

automatically into a device which would
read the information and record it on cards
or magnetic tape for further processing.
E. N assell (Electronic Associates):
Wouldn't all cases where you consider
transmission be much faster handled by a
ten-key keyboard of some sort? And as
cheaply? As I see it, the best use for reading characters handwritten is when the
original recording must be made remote
from accessibility to transmitting or recording equipment. Since a form is necessary in any case, doesn't it appear the
conventional existing methods of electrographic sensing would remain superior?
Mr. Dimond: If we consider only the
originating device, there is probably not
much difference in speed or cost between a
ten-key keyboard and a Stylator. Two
other factors appear to tip the balance in
favor of the Stylator. First, preliminary
tests indicate that better accuracy is obtained in writing than in keying. Second, a
written document can be made at the same
time the Stylator is used as a sending device, if the Stylator is designed to use
capacitive coupling through the paper
rather than direct coupling between the
stylus and the criterial areas. If it is necessary to transmit letters as well as numerals, the keyboard would undoubtedly cost
more than a Stylator and would require
more skill.
I assume that in using the term electrographic sensing, Mr. N assell refers to what
is more commonly known as mark sensing.
The main difficulty with mark sensing is
that 10 to 20 times mOre space per character is required than for written characters.

238

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Automatic Registration in High-Speed Character
Sensing Equipment
ABRAHAM 1. TERSOFFt

INTRODUCTION

REA T strides have been made in the field of data
processing machinery. A wide variety of equipments has been developed which will quickly and
accurately perform assorted operations on data fed into
them. However, far less progress has been made on the
task of efficiently providing these equipments with input
data. The existence· of a severe language barrier has generally necessitated the use of human operators to translate
all data into machine language when it is first fed into a
data processing system. In an attempt to efficiently overcome this language barrier, a number of different character sensing systems have been developed to serve as automatic man-machine links.
One of the problems faced by most character sensing
systems is the relatively inexact fashion in which human
beings tend to position information on a document, due
both to economic considerations and to the flexibility of
the human beings who normally operate on this information. Of great help under such circumstances is the ability
of the character sensing equipment to register automatically and accurately on the specific information selected for
processing. In fact, it is just such general flexibility which
tends to distinguish an equipment suitable for field use
from one capable of operating satisfactorily only under
laboratory conditions.

SCANNING OF THE DOCUMENT

The automatic registration system to be described has
been successfully employed in a number of different Analyzing Readers.1 These high-speed character sensing
equipments all utilize a high resolution scanner (usually
mechanical) and photomultipliers to convert the optical
image received from the document into electrical signals.
As shown in Fig. 1, the document to be read is moved past
a reading station, and an image of the information on the
document is focused onto a scanning disk containing perhaps twenty to forty radial slits, each 0.010 inch wide.
This disk is normally caused to rotate at a speed of ten to
fifteen thousand rpm. Immediately in back of the scanning
disk, and swept by light projected through the radial slits,
is a fixed plate containing a slit 0.010 inch wide. As the
document is moved horizontally past the reading station,
an image of the information on it moves across the system
of· intersecting slits. A two-dimensional scan of the infort Intelligent Machines Res. Corp., Alexandria, Va.

Trademark, Intelligent Machines Res. Corp., registered U.S.
Patent Office.
1

Fig. I-Simplifi.e? view.of scanne,r shows rapid scanning
along one vel heal aXIS of movmg character image.

mation is thus obtained by means of a beam of Ii ht
passed by what is effectively a "flying aperture."
g
Considerations such as the horizontal speed of the document, the height of the field to be scanned on the document
and the size of characters to be read will combine to determine the magnification ratio of the optical system employed, the length of the slit in the fixed plate the number of slits in the scanning disk and the speed ~f rotation
of the dis~. Under normal circumstances, more than
twenty vertIcal scans are made across each character with
a~jacent scans ov~rlapping slightly. In designing the'scar:
mng system, care IS taken to insure that the effective cross
section of the s~anning beam will be substantially less than
the narrowest hne element normally occurring in the characters to be read.
. The moving spot of light passed by the intersecting slits
IS focused onto a photomultiplier, where it is converted
into. an electrical. sign~l whose amplitude is always proportIOnal to the mtenslty of the spot of light. Since the
length of the fixed slit is made slightly less than the chord
between adjacent radial slits at that point, each scan across
the ~x~d slit will produce a "black pulse" in the photomultIplIer output at the point where light is completely
blocked from the fixed slit. (See Fig. 2.) During the re-

Tersoff: Automatic Registration

'In

High-Speed Character Sensing Equipment

239

GATED
SCAN
SIGNAL

REFERENCE CAPACITOR
AND ·TIMING SAWTOOTIi
'14 PLATE

LOCATOR OUTPUT

It'EnDING

PHOTOt1Vl..TIPLIE.R
OUTPUT

CLIPPED ?J'--_ _ _
G/lT£D .5cn/i
SIGNIIL

I 2

GATE.D SCAN SIGNAL

I I CAPACITOR DRIVER

Fig. 3-Locator circuit marks time-in-scan
of the earliest signal it receives.

llt1IN6· PHI.Jlt.JNUL TIPL/£P.

OUTPUT

CLIPPED
TlHIN6 PULSE

Fig. 2-Development of a clipped scan signal and timing pulse.

mainder of each scan, pulses will occur only when the
scanning of a line element of a character produces a
marked diminution in the intensity of the light passed by
the intersecting slits. This photomultiplier output is then
fed to a video channel, where it is amplified and clipped at
the voltage levels (+ 15 and- 25 volts) used in. subsequent logical units. Also developed in the video channel
is a feedback voltage which holds the amplitude of the
"black pulse" constant, compensating for variation in document reflectivity and photomultiplier sensitivity.
A second photomultiplier and an exciter lamp, placed on
opposite sides of the scanning disk, are used in a similar
fashion to produce a timing pulse at the end of each vertical scan (see Fig. 2.)
BASIC OPERATION OF LOCATOR CIRCUIT

The operation of the aforementioned locator circuit,
which has been used for automatic registration in machines with the above scanning system, can be analyzed
. with the aid of Fig. 3. In essence, the voltage established
across the reference capacitor is compared with the timing
sawtooth. This establishes the point within the scan cycle
to be "marked for registration" by the switching of the
locator circuit's output. Switching occurs at the earliest
point in the scan cycle at which the information being
tracked is detected on the document. A feedback loop
causes the circuit to keep switching at this point until it is
either moved to a point even earlier in the scan cycle by
new information or reset to the end of the scan cycle at
the end of reading.
Referring to Fig. 3, the signals at points B, E, F, and G
are always at either + 15 or - 25 volts. The units II, 12,
and 13 are simple inverters whose outputs are clipped at

+ 15 and - 25 volts. They are designed to provide an
output which is always the exact inverse of their input.
The key potential in this circuit is the one established at
point A by the reference capacitor. When the locator is
not in operation, + 15 volts is applied to reset input point
B, causing VI to conduct until the capacitor is charged to
the same potential. With point A at + 15 volts, V3 conducts sufficiently heavily to bring its cathode, point C, to
+ 15 volts. Under these circumstances, V4 will normally
be cut off, conducting only during that brief period at the
end of each scan when the timing sawtooth on its grid
rises above the tube's cut off potential. Only during this
period of conduction will the potential on the plate of V 4,
point D, be sufficiently low to bring the output of 12 up to
+ 15 volts. At all other times point E will be held at - 25
volts. Since the signal at point F is the exact inverse of
that at point E, the AND gate input to II is enabled during almost the entire scan. Only the absence of a gated
scan. signal keeps point G at - 25 volts and point H at
+ 15 volts, preventing V2 from conducting.
To put the locator into operation, the potential of point
B is reduced to - 25 volts, effectively removing VI from
the circuit. The potential of point A still cannot rise above
+ 15 due to the clamping action of V2, whose cathode is
normally at + 15. However, whenever a gated scan signal
is detected during that portion of scan when point E is at
15 volts and II will try
- 25 volts, point G will go to
to drive point H to - 25 volts. It can seldom do this, since
the reference capacitor normally holds the plate of V2
above - 25 volts. Inverter II, however, will conduct to
the limit of its capacity, discharging the reference capacitor through V2 and driving the potential at point A down.
Tube V3 will then conduct less heavily, the potential at C
will drop to approximately that at A, V 4 will conduct for
a longer interval at the end of the frame, and input F will
disable the AND gate in each frame earlier than before.
This process will continue, with II driving the potential
at A lower and lower, and the output of 12 getting up to
+ 15 earlier and earlier in the scan, until the output of 13
holds off the AND gate from the earliest point in the scan

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

240

that a gated signal is detected until the end of the scan.
The potential at point A will then remain relatively constant, causing the leading edge of the locator output, point
E, to continue marking the earliest point in the scan that a
gated scan signal was detected. This situation will continue until the potential at A is either driven sti11lower by
the detection of a gated scan signal even earlier in the
scan, or reset to + 15 volts by the signal applied at point

1..ONI5

VE:R.7ICRL

LEFT

LONG

VERTICRt.

RIGHT

HORIZONTRL

TOP

HORIZONifiL

MIDDLE

HORIZONT'!L

BOTTOM

5;f~£XERJ~~Zl:,.
S:f:ZE:£1r!f~L

.',' :::""u1FR::::;f:'
SHORT

It was stated earlier that once set, the potential at point
A will remain relatively constant until either driven still
lower or reset to
15 volts. The primary reason this potential does not remain constant is that current drawn by
the grid of V3 tends to accumulate on the reference capacitor and slowly drive point A negative. To counteract
this effect, a large resistance is connected between this
point and a much more positive potential in order to draw
electrons away from point A at approximately the same
rate that grid current is supplying them. In general, it is
safer to choose a resistance value which causes the potential at point A to drift very slowly positive, since it will
simply be driven back down as soon as the output of 13
fails to hold off the AND gate while a gated scan signal
is present. If, however, the potential at point A is permitted to drift in a negative direction, causing the locator
output to mark erroneously a point earlier and earlier in
the frame, nothing can drive the potential at A and the locator output back until the unit is completely reset.
It should be noted that the speed with which the potential at point A can be set or reset to the proper values is
determined both by the capacitance of the reference capacitor, and by the rate at which electrons can be supplied by II and drawn off by the reset pulse.

ApPLICATION OF LOCATOR CIRCUIT

The basic method of operation of the locator circuit has
been described above. In practice, the circuit is sometimes
made to operate in slightly different fashions for different
applications. For example, in one application the scan signals used to set the locator could simply be those used in
the basic character analysis program. Other applications
may require that the locator be completely positioned and
performing its registration function before the information to be analyzed reaches the primary scanning station.
In such case, the locator can be positioned by scan signals
obtained from this same information as it passes one or
more pres canning stations. Using the mechanical scanning
system previously described, it is relatively simple to scan
simultaneously a number of lines displaced horizontally
or vertically from each other on the document. In all applications, however, the basic goal is to register in a known
manner on the characters to be read, and to do so with
sufficient precision to permit an analysis of the various
strokes comprising the characters. An example of the general manner in which characters are normally distinguished from one another is provided by the table in
Fig. 4. Here, eleven different stroke criteria are employed

VCR nell/.

LONG VERTICIIL
SII'IVLTilNGOl/S4Y
MIDOLE

IJPPERL EFT
I.IPPER R/WltT
UFT , . RlfiHT

(JPPER I.£FT --I--f.-I--

LEPT

~ R[GHT

PRo~E.CrING IlIGrrT
+ CONDITION 1'11.15T
-

- -t- +
+ + - + + - - +
+
+-. . + + - + - + + +
- - + + -t- + + - + +
++++-++-++ + -t- - + +
+ - +. - i - -t- + - - - i
- + - + +

CONOIT/ON

--1--

I--=-J + i
' ,'--'
--l--.L.

1+"1

.8£.

"""57 NOTSE

.j..'

----j--f-,-

--l----l--i---l

:JETE<::TEO
DETec.TED

Fig. 4--.Character is identified by correct combination
of detected and not detected conditions.

to provide a minimum of two differences between any
pair of numbers in a particular type face. This program
would not be possible if the lower, middle and upper portions of the characters could not be identified fairly precisely.
The character identification program shown in Fig. 4 is
the one actually employed in the Scandex 2 character sensing system, which automatically processes imprinted gasoline credit card invoices. This equipment reads the customer's account number, as imprinted on the reverse side
of an invoice card, and punches it into the same card.
Since carbon paper is used to make this impression, we
are sometimes confronted with smudges adjacent to the
characters to be read. In addition, the vertical registration
of the account number on the invoice card varies appreciably, being affected by the registration of the credit card
during the embossing operation, the position of the credit
card relative to the invoice in the imprinter, and the precision of the Scandex card feeding mechanism. To achieve
accurate registration on the account number despite the
presence of smudges, the locator is programmed to track a
point two-thirds below the tops of the characters. This is
accomplished, as shown in Fig. 5, by employing two simultaneous scans spaced a character width apart, summing
their scan signals, and operating on this sum. Thus, small
spurious information will not affect the location system.
With the point-in-scan marked by the locator circuit serving as a reference, the characters are then divided into top,
middle, and bottom thirds by appropriate time measuring
circuits.
Another Analyzing Reader, in which three locator circuits are incorporated, is presently being employed to
process automatically utility company billing stubs prepared on tabulating machines. An example of such a stub
is shown in Fig. 6. In this application it is necessary that
the machine record the first two and last four digits on the
top line, the four digits on the second line, and the total
on the bottom line of the stub. Here the problem is one of
2 Trademark,
Farrington Manufacturing Co., registered U.S.
Patent Office.

Tersoff: Automatic Registration in High-Speed Character Sensing Equipl'nent

241

·---2-·=r·Kf)~--T--.:--o--=rl~~

-----·-=r-o-o-------~-

.------- ----l •

b,Sl

SEP
I
.5A1VPGG

C#"i/?Acrcll? .:

$MVOL.
I

~rEt7 '8"
sc-?,u S/oA/A'L.

Fig. 7-Locator circuits track adjacent lines on billing stubs.

5VA? d,.c- # " 0 '
S/~,uA'~<;'

. r - - LOCATOR 'B"

ABent: F·GR=--r:r=--=:.~ L;GAfoA'"'"A~ ,
"----,--·,--KL-·-----·-·-F~-', ,iOc.Aril')--C::-

MNO, PQRSTUVW
XY.Z
Fig. 5-Scandex locator tacks point two thirds
below top of combined scan signals.
f)1f/£CTlvN OF /JOCI/MC.\r

PlEA'l€. 9IU.OS£ nwiSrue

, CIR/i,riO,v

WAs}l.mg~~'l:l±.~-:.::.-.:.: ::-..:..:::.-..:..: ::.;.-:.::: --::. =--::

P£f; RtNCI c/1PI1CITOR.
IlND 7iN/No 5nwrOOT/1

Gt?TEO

SCHilt

SIGNI?£

LOCIlTOk Ou/pur Ei!

p/4.7 ;Ecr/oA./

,,""

pcKVNlI!:,ur

Fig. 9-Locator circuits with slow and fast positive
drift of the reference capacitor voltage.

A fourth Analyzing Reader employing locator circuits
is an automatic mail sorter being developed for the Post
Office Department. In this machine, programming considerations make it desirable that we be able to register on
the bottoms of the various characters in the bottom line,
as with locator P in Fig. 9. This is achieved by connecting
a relatively low resistance between the reference capacitor
and + 100 volts. In this case, 18 megohms is used, causing the reference voltage to drift in a positive direction
relatively quickly when not being set down by input scan
signals. The danger in this approach is that the locator
might occasionally drift up to the previous line, and confuse it with the bottom one. To prevent this happening, a
second locator, with a much larger "pull-up" resistance (in
this case, 500 megohms) is made to track the same characters. This locator will behave more like locator Q in
Fig. 9, drifting upward at a much slower rate. Now, by
using locator Q as a reference, an alternate input for 10'cator P can be developed to prevent its drifting to a point
more than three fourths of a character height later in the
scan than locator Q. As can be seen, this permits locator P
to follow the bottoms of those characters in the bottom
line very closely without any danger of its drifting into
the previous line.
An interesting problem encountered in this same mail
sorter application was the need for determining immedi'ately the fact that we have begun reading a new line in a
staggered address. A normal locator circuit tracking the
bottom of the bottom line on a document will, in such a
case, simply drop down to this new line. How, then, do
we determine that this has occurred? The technique presently employed is to have an additional locator track a
point roughly three fourths of a character height below
the line being read. The presence of a scan signal earlier
in the scan than this point indicates that we have begun
scanning a new line in a staggered address.
A locator circuit can be made to track a. specified distance ahead of the input scan signal by simply delaying
the leading edge of its output pulse by a fixed amount before inverting it and feeding it back to the AND gate.
Unit M1 in Fig. 10 is used to accomplish this delay. As
can be seen from Fig. 10, the leading edge of the M1 out-

OllTPt/T

Fig. 100Locator circuit marks a specified interval before
the time-in-scan of the earliest signal it receives.

put pulse, and hence the trailing edge of the 13 output
pulse, follow the leading edge of the locator output pulse
by a fixed interval. The presence of M1 does not, however, alter the basic operation of the locator circuit itself.
Whenever the AND gate's output is positive, V2 will conduct and cause the locator to mark a point earlier and
earlier in the scan. It is only I3's disabling of the AND
gate that prevents scan signals from improperly advancing the locator output all the way to the beginning of the
scan. Scan signals can advance the locator only if they
occur while the output of 13 is positive. We see, then, that
scan pulses will continue to advance the locator until the
trailing edge of I3's output is brought to the same point
in the scan as the leading edge of the earliest scan pulse.
Of necessity, the leading edge of the locator output will
then be marking a point-in-scan which is the specified
distance ahead of the earliest scan signal detected.
CONCLUSIONS

In most high-speed character sensing systems, automatic registration is extremely helpful in: 1) overcoming
poor registration and extraneous marks on the document,
2) compensating for tilt of the information to be read, and
3) accurately dividing the characters to be read into vertical zones for purposes. of stroke analysis. A special locator
circuit is employed in a number of IMR Analyzing Readers to provide such automatic registration. Minor modifications of the circuit permit the control of such characteristics as the speed with which it can' be set, the speed
with which it can be reset, the timing stability of its
switching point, once set, and the degree to which it can
register on misaligned characters. Registration on a variety of types of information, either in restricted zones or
anywhere on the document, is possible. The usefulness of
this locator circuit has been thoroughly established through
its application to various Analyzing Readers presently
under development or in use in the field.
ACKNOWLEDGMENT

The basic locator circuit was developed by David H.
Shepard of Intelligent Machines Research' Corporation.
Techniques for applying it to different models of Analyzing Readers reflect the contributions of numerous members of the IMR staff.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

243

The National Cash Register High-Speed
Magnetic Printer
J.

SEEHOFt,.- M. ARMSTRONGt, G. FARLEYt, M. LEINBERGERt,
M. MARKAKISt, AND S. SMITHBERGt

1. INTRODUCTION
HE magnetic printer is designed to operate in conjunction with an electronic data processing system.
It operates on the principle of recording a latent
magnetic image on a special paper, which is essentially
magnetic tape with a white topcoat. The image is in the
form of an alphanumeric (or other) character which is
subsequently made visible by exposure to a ferromagnetic powder attracted to the magnetized portions of the
paper. The powder is coated with a thermoplastic resin
and requires a heating operation to fix it to the paper. An
example of the print is illustrated in Fig. 1.

0123-'56789
01.23-'56789
0123-'56788
0123.q.56788
0123-'56788

'r' -

'f -

'f- .
'f - .
'f -

may then be made visible by exposing the area to a black
permeable powder.
Dynamic operation of the printer is illustrated in Fig. 3.
The magnetic field is established by a coil wound around
a permeable bar. It is energized with pulses of a length
directly proportional to the length of vertical scan to be
recorded. Illustrated is a three-by (three vertical needle
sweeps per letter) format where the letter T;cis formed by
pUlsing the bar when needle 1 passes over the top area,
while needle 2 is traversing the entire bar, and when
needle 3 passes over the top area. The needles are arranged in the form of a helix on a drum rotatillg over the

ABCOEFGHIJKLM N9 PQRSTUVWXY2
ABCDEFGHIJKLMNOPQRSTUVWXY2
ABCDEFGHIJKLMNOPQRSTUVWXYZ
ABCDEF GH I .. IKLMNO PQRSTUVWXVZ
ABCDEFGHIJKLMNOPQRSTUVWXYZ

012:3"456789
0123"456789
,0123"456789
0123"456789
0123"456789

"7- .

7-.
-? - .
?-

7- .

Fig. I-Sample of print.

No Needle

Fig. 3-Three-by system.

Paper

Permeable
Core

Fig. 2-Effect of permeable needle on field strength.

II.

PRINCIPLES OF OPERATION

Printing is accomplished by utilizing the fact that a
weak diffuse field will not magnetize the paper while a
strong concentrated one will. Field concentration is obtained, as illustrated in Fig. 2, by placing a permeable
needle directly above the area which is to be printed. Print
t Electronics Div., National Cash Register Co., Hawthorne,
Calif.

bar and are spaced so that only one needle is within the
field of the bar at anyone time.
A completed breadboard of the magnetic printer (Figs.
4-6) consists of a spiral track of needles arranged around
the surface of a drum. Directly over the drum supported
by a fixture, is a permeable bar. The magnetic field which
places the latent image on the paper is established between
the needle points and the bar. During the time that a particular needle point is passing under the bar, a magnetic
flux path is established between the needle and bar only
when the coil is energized. Seven vertical scans are allotted
for each character (seven.:..by system). The scan (bar) is
energized for the full height of a given character (width
of bar) if a full vertical line is to be recorded for that portion of the character, or the scan is energized for only a
porti~n of the time as needed.
As an example, consider the formation of the letter L.

244

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

Cod.

Fig. 4-Block diagram of magnetic printer system.

Fig. 6-Breadboard model of magnetic printer.

••

••••••
1234567

Fig. 7-The letter L.

Fig. 5-Close-up view of the drum, bar, and paper.

The second scan is a full height scan, whereas the other
six are energized for only portions of each pass over the
bar. The resulting image appears in Fig. 7.
The encoder which has been utilized in conjunction
with the printer is a cathode-ray-tube device. Over the face
of the CRT is placed a photonegative mask which has on it
all of the characters which are to be used in the printer
(60 or more). The mask is easily replaceable with other
sets of characters or symbols. In back of the mask is a
photocell which converts the light energy from the phosphor on the face of the CRT to electrical pulses. When a
character is to be printed, 1 the binary coded representation
of the character enters the electronic input circuitry of the
printer. A decoding matrix then directs the horizontal,and
vertical deflection plates of the CRT to a certain portion of
the face of the tube, this point being at one corner of the
character (on the mask) which is to be scanned. The CRT
1 Along the edge of the drum is placed a magnetic layer. On
this layer are placed pulses in line with the first needle for a
revolution, each needle beginning a character, and each needle on
the drum. Three separate reading heads are utilized. Thus the
printer can "signal" the encoder or paper feed when a line begins,
when a character begins, or. when the nC'edle begins scanning the
bar.
'

beam then scans the character in seven sequential vertical
sweeps in synchronism with the passage of the seven
needles past the bar in the printer. The photocell senses
the beam when it illuminates the transparent portions of
the mask. In like manner, when the light energy is interrupted by the opaque portions of the mask, the photocell
lies dormant. The electrical signals from the photocell are
amplified and converted into current pulses for the coil
and magnetic field pulses from the bar. This CRT encoder
has been utilized for some time and has proven very reliable. Various other types of encoders, such as a magnetic
core matrix are feasible and would probably be less expensive.
As the drum rotates over the bar, a full line of characters can be printed. Since the width of the helix depends
upon the diameter of the drum for a given letter size, one
can design either a large drum to scan 120 characters per
revolution or a smaller drum to scan 40 characters per
revolution with three spirals (each with an independently
pulsed bar) wound on its surface. To achieve equal speeds,
the smaller drum would have to rotate three times as fast.
The present system scans 40 characters per revolution and
80 characters per line at a speed of 1200 characters per
second or 15 lines per second. Higher speeds than this
have not been attained with the present breadboard due to
mechanical vibrations set up upon rotating the drum at
high speeds. There is no reason to assume that increases in
speed by at least a factor of ten cannot be attained with
better drum design.
.
The paper that is used is formed by a series of layers:
1) Base kraft paper.

Seehof et al.: The National Cash Register High-SPeed Magnetic Printer

245

1 0 · . . _ _ - - - L - - - - - - - : :.......-

I;"

10 L - - - - - - I I - - - - - - - t 1000

100

Fig. 9-Effect of length to diameter ratio of
needle on effective permeability.

I
Fig.

l~N eedle

retentivity and trailing.

1.0
0.8
0.6
0.4
0.3

0.2
~o

::L

"CO
,..!t

Fig. 8-Experimental working model of a liquid inking chamber.

~ETS_

" '" ,,~

0.1
0.08
0.06

f - SHEETS, fl= 4~2
0.04 I . 0.03 I . WIRES, «=~
0.02

III.

DESIGN PARAMETERS AND EXPERIMENTAL RESULTS

WIRES

0.0 1

0.1

2) Magnetic oxide layer (can be permanently magnetized) .
3) White topcoat.
a) Forms a contrasting surface for the black ink.
b) Forms a multilith surface.
The printer produces only one copy at the time of printing. However, since the paper contains a multilith topcoat, it can be utilized to make any number of copies at a
later time by a simple multilith operation.
Inking is accomplished by passing the paper through a
liquid (Freon-113) suspension of iron particles (mapico
black coated with resin) which is in a state of rapid agitation. This operation can be accomplished with the apparatus illustrated in Fig. 8 at speeds greater than 15 lines per
second. For off-line operation a continuous feedthrough
would be utilized. For on-line operation, since it is not desirable that the paper stop for any length of time in the
inking solution (overinking would occur), a take-up reel
would be utilized so that the paper would be passed
through the inker in long passes (10 feet or more). As
soon as a pass through the inker was completed the paper
would stop, the agitation would stop, and therefore further
inking would cease. When another ten feet of paper built
up, the printer would signal the agitator to begin again
and another ten feet of paper would be inked.

0.2

0.3

0.4

0.5

0.6

0.7

at or ~t

Fig. ll-Build-up of flux in wires and sheets
subjected to sudden constant field.

obtained since it is directly proportional to the field magnetizing the paper. A certain minimum En is necessary as
a print threshold, and as high a value as possible should be
obtained consistent with the other variables. Since En ~ ILn
(needle permeability) H B (field due to bar), it ~an be se~n
that keeping all other parameters constant, the hIgher ILn IS,
the higher En will be. The effective permeability IL', is not
only a function of the material utilized but also of the
length to diameter ratio of the needle. The graph in Fig. 9
illustrates how important this ratio is. 2
If the needle utilized has "hard" magnetic characteristics (permanent magnet properties), and retains part of
its field after being pulsed, it will continue to print in a
weaker fashion if this field is above the threshold value of
Bn. To eliminate this "trailing" effect, "soft" magnetic
materials should be used. For example, if trailing occurs,
the letter T would look as shown in Fig. 10.
2. Needle Response Time: In te megacycle region, the
needle response time will be primarily dependent upon eddy
current effects. These effects are in turn functions of permeability, resistivity, and geometry. The graphS in Fig. 11
illustrates the effects of these variables for wires (needIes) .

A ..Needles
1. Needle Permeability: The flux density, En, concentrated
in the needle is a measure of the print quality that can be

2 R. M. Bozarth, "Ferromagnetism," D. Van Nostrand Co., Inc.,
New York, N.Y., p. 848; 1951.
3 Ibid.} p. 784.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

246

Fig. 12-Field distribution for magnetized paper area.

The larger the resistivity, and the smaller the diameter
and permeability, the quicker will be the rise time of the
needle. The same general argument will apply to the decay of the fields. An example follows:

H= 0,
H

= Hn ,

= 0

B t = Threshold flux density at which print just becomes visible on paper.
Bp
Flux density needed for dense printing.
flBe '= Transient flux density remaining after each pulse due to
eddy currents.

Fig. 13-Effect of ned Ie length

011

hysterisis characteristics.

Needle -mu-metal (I-' ~4 X 104).
' "IS attame d)
Let 1 - -Bn
- = 0.1 C'~.e., 90 per cent 0 fnse
J.LHB
at = 0.35.

Case I:

r = 0.005 inch
t

= 0.44 J.Lsec.

Case II:
r = 0.060 inch
t =

66 J.Lsec.

If the needle response time is too slow, a transient
"trailing" effect analogous to permanent retentivity will
be present.

3. Needle Shape: Fig. 12 is a lateral view of a piece of
magnetic paper containing a pulsed portion. The area
which will most attract the permeable powder is that portion exhibiting maximum field strength. It is apparent that
this will occur not at the center of a pulse, but at the edges,
where a short return path for the magnetic flux is available. Print density from a large needle will suffer by having central unprinted areas, while a very small needle will
have lesser area coverage on the paper. The effect of
powder diameter should be appreciable as the particle size
approaches printed area dimensions.
4. Effect of Needle Length: Curves 1, 2, and 3 of Fig. 13
illustrate the general behavior to be expected of the hysteresis curves as needle length is shortened while diameter
is kept constant. The heavy lines depict the static magnetization-demagnetization curve while the dotted line indicates the mode of demagnetization in a dynamic (high
frequency) case. After a short period of time, B will drop
to the normal static case.
a) The needle of Curve 1 will exhibit strong "transient retentivity and also permanent retentivity which

Fig. 14--Paper field definition and concentration
as a function of paper-needle distance.

will cause light print to appear on the paper even if
the bar is not being pulsed.
b) Needle 2 will exhibit weak "transient" trailing and
no permanent retentivity.
c) Needle 3 should show no "transient" trailing or permanent retentivity.
These effects were all demonstrated by experiment, and
strong printing with no trailing was obtained with soft
iron, piano wire, and mu-metal of 0.10 and 0.05 inch in
length.

5. Needle Distance from Paper: As the needle is moved
away from the paper surface, a marked diminishing of
field strength and definition is apparent (see Fig. 14).
Thus, optimum printing conditions will be present when
the needle is as close as possible to the paper surface.
6. Overlap of ,Needle Fields: In a three or multiple-by
system depicted in Fig. 3, it is possible for the fields of
successive needles to cancel each other if they overlap.
Fig. 15 shows what proper and improper needle spacing
can do in a two-by system.
7. Needle Pulsing: Six mu-metal needles were wound with
twenty-five turns of wire each and connected in series
through a commutator so that the pulses received were the
same that were ordinarily fed to the bar (see Fig. 16).
Experiments were conducted to determine the effect of:
a) Needle pulsing alone.
b) Needle and bar pulsing with same and opposite polarities.

Seehof et al.: The National Cash Register High-Speed Magnetic Printer

Weak

Medium

247

Strong

Fig. I7-Effect of bar field on paper field
concentration and definition.

Fig. IS-Paper field patterns as affected by needle overlap.
(a) Improper. (b) Proper.

field strengths), low permanent retentivity (no permanent trailing), and fair eddy current lowering (slight
"transient" trailing found at high field strengths). Transient trailing (eddy current effects) is apparent with all
needles at high field strengths, but decreasing the length
to diameter ratio of the needle almost completely eliminates this effect.
As a result of these experiments and analysis of design
parameters, the needles utilized in the printer were fabricated of short (50-mil length, 7-mil diameter) pieces of
mu-metal wire. The wire was then mounted in nonmagnetic stainless steel cylinders (60-mil outside diameters
with a 7-mil hole drilled down the center) to give rigidity
and to keep the wires away from the surface of the aluminum drum. The cylinders had no noticeable effect upon
the print quality.

B. Bar
Fig. I6-CIose-up of needle pulsing equipment.

c) Needle pulsing in the presence of a nonmagnetic
(lucite) bar.
All results were positive but it was felt that the slight
improvement in print quality did not warrant the extra
trouble inherent in mechanizing the commutator system.
The purpose of this experiment was to determine whether the directed field emanating from the needle point itself, instead of the bar, would improve definition of print.
Very little improvement resulted.
8. Needle Materials Tested: The following materials
(available in needle form) were tested to determine optimum results with respect to permeability, retentivity, and
response time:
a) Welding rod (commercial grade)
b) Piano wire
c) Phonograph needles
d) Ferrite
e) Mu-metal
f) Laminated mu-metal
g) Pure iron wire (reagent grade)
h) Pressed powered carbonyl iron (GQ-4).
The best over-all characteristics were obtained with mumetal which exhibits high permeability (response at low

1. Field Strength of Bar: Weak bar fields will cause no
print at all until the threshold of Bn is exceeded. Very
strong fields will cause printing over a wide area with loss
of definition, even though field strength is increased. An
optimum field strength exists (see Fig. 17).
The effect of varying the current through the bar from
0.5 amp to 3.0 amp is illustrated by Fig. 18, next page.
2. Effect of Stray Field from the Bar: Because of stray
field from the bar, weak print may appear both above and
below the desired line unless special precautions are taken.
This effect was substantially decreased by winding the bar
up to its very tip, thus effectively directing the field in the
desired direction. Fig. 19 illustrates the two cases.
3. Bar Distance from Paper: Since Bn :::::: tJ.HBJ and since
HB does not vary much with small distances from the bar,
one would expect approximately the same definition of
field in the two cases. Gradual diminishing of field strength
as the bar is moved from the paper should also occur (see
Fig. 20).

4. Electronic Response Time and Bar Permeability: The
rise time of the electronic circuitry pUlsing the bar should
be kept to a minimum. Besides being a function of various
circuit parameters, the electronic rise time is markedly
affected by the LjR ratio of the pulsing coil. The number

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

248

,t··t

•

II'

• o'

•

II",' ,'" 'II 11.. ··1

tl'l

III"

.,:

•

II",'

..,'

II" I.'

t,'" .•
to,"...

I 1',(

I II''''

-pnll'·',n~("I'It· ...·r;i ..1I"IKI ..., ....ult·'(:)fii:S t U\I

Increasing

t-;,rent

Fig. 20-Paper field concentration as a function
of paper-bar distance.
,-PRlC::A'HCIlf- t:-1';;iH I ..JKI M"'Of'(~RS'l' UH

50 mil ,,-metal.needle

::\~;.~

'. -PRI'::AHCllf- .... r;H I. JKI."''''(~P(:lRS I' UH

. -pmCAA(:Df'F'(;H '.,JKI MNPPQRSHJ\I
. -pmc AHCDf F'(;H I •. JKI MNOPQRS'T U\I

(

;..pmCAHCDf.fo(;H '.,JKI "'NOPQR~H"J\'

3.0

amp

-pmc AH(:Df f (;H I .,JKI MNOF'QRS'j

,j"

Fig. IS-Effect of bar current ..

(a)

(b)

Fig. I9-Bar field. (a) Properly wound. (b) Improperly wound.

of turns about the bar should be made an optimum with
respect to print quality, since not only rise time but H B is
a direct function of the inductance of the coil. To obtain a
given H B, (or a given number of ampere-turns) the current should be maximized consistent with other variables
and the number of turns held to a minimum. Bar permeability should be as large as possible to minimize number
of ampere turns necessary.

~.I.I.ss

Steel Cylinder

(a)

(b)

Fig. 2I-Effect of drum mass. (a) Poor print. (b) Good print.

d) Pressed carbonyl iron (all types)
e) Cold rolled steel.
Good print was attained with all bars tested, but ferrite
gave consistently better results. At present speeds, ferrite
gives acceptable results (as do the others). At higher
speeds, on the order of 10,000 characters per second, further experimentation may be necessary to provide a high
enough response time.
The final bar utilized in the printer was 0.1 inch wide,
0.875 inch high, and 8 inches long. 100 turns of wire were
wound in four layers of 25 turns each, concentrated on the
upper Ys inch of the bar (to keep the field well defined).
Continuous runs were not made, but it is anticipated that
heat effects with concomitant insulation breakdown of the
wire will occur under these conditions if precautions are
not taken. Magnet wire with high temperature characteristics will be necessary.

C. Drum
5. Bar Response Time: The magnetic response time of the
bar lags behind the pure electronic response, and, if too
slow, may cause transient "trailing" of the same type as in
Section III-A, 2 (needle eddy-currents).
6. Bar Retentivity: As long as the retentivity (H B) ret is
below the threshold value for printing, this may aid somewhat in acting as a "bias." Instead of pulsing the bar from
HB = 0 to HB > (HB) threshold, one may decrease the eddy
currents by pulsing from (HB)ret to HB > (H B) threshold'
The same type of analysis would apply to needle retentivity.

7. Bar Material: Bars constructed of the following materials were tested:
a) Ferrite
b) Mu-metal
c) Laminated mu-metal

The large mass and high electrical conductivity of the
aluminum drum used in the prototype magnetic printer are
the cause of large eddy current effects, so that it has been
found necessary to mount the mu-metal needles as far
from the drum surface as possible with present equipment.
No print at all was obtained when the needles were at the
surface of the drum, while marked improvement was obtained by moving the needles away from the drum surface
as indicated in Fig. 21.
Other approaches which would eliminate this effect are:
1) Fabrication of a nonconductive drum.
2) Coating the aluminum drum with a non conductive
layer at least two inches thick.

D. Aluminum Paper Guides
Fig. 22 illustrates the geometrical configuration of paper
guides, bar, needles, and paper; Fig. 23, the print obtained
from this system.

Seehof et al.: The National Cash Register High-Speed Magnetic Printer

249

.~paper

Paper

-2/

""---Gu-id-e

Fig. 22-Effect of paper guides.

~ Gu;de

Fig. 24-N ew guide system.

~~~~~~~t.~~;~:§.t,'1 ~~t:~~b,P.~~~t.q~
(

~e~~~~~~t~§.~~~~t~~b~~~~t.q~

Fig. 25-Print with new guide system.

Fig. 23-Print with paper guides.

While needle 2 of Fig. 22 is pulsing, needles 1 and 3 are
so close to the paper surface that the stray field from the
bar causes weak print to appear both above and below the
desired line. Letters below are displaced one needle to the
left and those above, one needle to the right.
For example, while needle 2 is tracing out the center
scan of the letter T, needle 1, which is displaced to the
left and which is beneath the line, is tracing the same scan.
Needle 3 is doing the same thing above the line and displaced one to the right.
A slight amount of retentivity is also apparent. This is
attributed to the presence of the· high conductivity aluminum guides with consequent eddy currents. These two
effects were eliminated by utilizing the following system
illustrated in Figs. 24 and 25.
By causing the paper to be bent sharply over the end of
the bar, it is possible to eliminate the effect of the adjacent
needles, since print threshold is very sensitive to distance
of the needles from the paper. The lines seen above and
below the print are due to an "edge effect" of the bar.
Extra strong lines of force emanate from this region and
the effect was easily eliminated by placing a 5-mil mylar
shim over the bar surface. This sufficiently lowered the
field of the edge to a value below the print threshold but
did not substantially lower the over-all field strength.

E. Paper
1. Paper Response Time: Since the magnetic layer in the
paper is composed of finely powdered ferrite particles,
with consequent minor eddy-current problems, it is probable that the paper response time is appreciably faster than
that of the bar and needles.
2. Paper Magnetic Fields: In order to most strongly attract the printing ink (fine magnetic powders), the magnetic particles in the paper should have characteristics of
maximum retentivity.

Premagnetize~

Fig. 26-Paper field patterns as affected by premagnetization.

3. Paper Orientation and Magnetization: Since the ferrite
particles in the magnetic layer are asymmetric, they should
be oriented during manufacture so that the axis of maximum permeability lies in the direction of the field which is
to be applied (i. e. perpendicular to. the plane of the paper) .
Premagnetizing the paper in a direction opposite to that
of the applied field should enhance printing density by
shortening the flux return path (see Fig. 26).
J

F. Inking
1. Liquid Inking: Investigations determined that inking in
a liquid medium did not have to be conducted under static
conditions. Agitation by a Fisher Vibrastirrer yields
greatly improved results. Movement of paper through the
liquid medium does not cause smearing or other deleterious effects on print quality. An experimental inking chamber is illustrated in Fig. 8.

PROCEEDINGS OF THE EASTERN COJtfPUTER CONFERENCE

250

2. Gas Phase (Dust) Inking: All experimentation in this
area yielded poor results with a large amount of background inking. There is also a serious danger of explosio.n
when utilizing small iron particles in a system where stahc
charges can be built up.
3. Inks and Field Attraction: If an unmagnetized particle
is brought into the field of a permanent magnet, the force
of attraction is as follows. 4 (See Fig. 27.)
M

O~E=-------

F (attraction)

-----....
;.0

6MM1 where
M induced magnetic
X 4 Ml permanent magnet

=- -

=
=

Fig. 27-Attraction of permeable particle by a magnetic field.

The larger the permeability of the particle, the larger
are M and the attractive force. At the same time it is desired to minimize the X 4 term by using fine particles to
allow a minimum distance of approach. Thus, since fine
particles have lowered permeabilities,5 an optimum ink
particle size must be experimentally determined in order
to maximize (F /mass) of the particle.
Particle shapes and interactions with surrounding fluids
will influence the rate and density of deposition upon a
magnetized paper surface. Maximum covering power per
particle is also desirable for best printing contrast.

4. Solvents: Freon-113 is an excellent dispersing agent because it has the following properties:
a) Volatile
b) Nontoxic
4

Ibid., p. 729.
Ibid., p. 45.

c) Noninflammable
d) N onviscous.
IV.

PRESENT CHARACTERISTICS AND FUTURE
CAP ABILITIES

The present characteristics and future capabilities are
listed below.
1) Speed-I-2000 characters per second at present:
20-50,000 characters per second ultimately.
2) Type face-External and replaceable fonts of 64 or
more characters each; unlimited number and kind of
characters are available with the restriction that they
must fall within the space usually reserved for capital letters.
3) Legibility-Comparable to typewriter print.
4) Noiseless-Except for paper feed.
5) Low maintenance-No mechanical moving parts except rotating drum. No breakdown due to solenoids
and moving hammers, no wear of type face, no ribbon debris.
6) Serial operation-No line buffer necessarily needed.
Continuous (nonintermittent) off-line paper feed
possible.
7) Copies-Unlimited number available from a multilith master which is printed directly. Only one copy
immediately available unless printing is conducted
in tandem.
8) Immediate visibility-Not available: lag of ten feet
of paper if used for on-line operation with take-up
reel; about two feet for off-line operation with steady
travel through inker.
9) Character size-Present machine has ten characters
per inch, each 0.1 inch high, width of printing line
80 characters. All of these parameters can be easily
varied.

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

On-Line Sales Recording System
]. s.

BAERt, A. S. RETTIGt,

AND

I. COHENt

INTRODUCTION

HIS PAPER presents an equipment description of
a pilot on-line Sales Recording System currently in
operation for the Associated Merchandising Corporation as part of a research project. This pilot system comprises point-of-sale units connected to a central computer
by means of an Input-Output Buffer Unit. The operating
characteristics of this system are such that they may be extended to include, for example, inventory and production
control, and the handling of transportation reservations.
Operation of the pilot unit which started in April, 1957,
has continued with highly satisfactory results. The Sales
Recording System has maintained an average "up-time"
record of over 97 per cent for the past nine months.
SALES RECORDER

The point-of-sale unit or Sales Recorder (Fig. 1) consists of a keyboard for manual input, in combination with
a character display, and a procedure indicator. There is a
punched tag reader for automatic input of the sales person
number, customer number, and merchandise stock number.
An output printer provides for a three-part sales check at
the point of sale. The entire unit is packaged over a cash
drawer and is contained within a vented aluminum housing, which opens completely for servicing.
The keyboard (Fig. 2), is of the type commonly referred to as a ten-key keyboard. Actually, it consists of
fourteen keys-ten numeric and four control keys. The
four control keys are: Enter, Nonmerchandise, Clear, and
Total; they allow the operator to control the procedural input to the machine,
The keyboard is designed to provide both manual and
electrical interlocks. When one key is depressed, another
key cannot be depressed, and two keys cannot be depressed
at the same time. The actual depression of the key is only
for the first position of the key stroke. From that point on,
the key is mechanically pulled down, and is held down
until the Input-Output Buffer Unit has recognized the
character being input as a legitimate one. When this routing is complete, the key is allowed to return to its normal
position. Each time a numeric key is depressed, that number is shown in a lighted character display window of the
point-of-sale unit. In this manner, a sales person can verify any item prior to striking the Enter key, which clears
the display.
t Electronic Data Processing Div., Industrial Electronic Products, RCA, Camden, N.].

Fig. I-Sales Recorder.

Fig. 2-SalesRecorder controls and indicators.

251

252

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

The output of the keyboard is four bits with parity.
These five bits, plus an additional control bit, are the information levels sent to the Input-Output Buffer. Operation of the keyboard can be at the sustained rate of six
characters per second.
The procedure indicator provides a visual indication to
the operator of where in the procedure the operator should
be at any given time during input of information to the
Input-Output Buffer. The indicator consists of a plastic
engraved drum driven by a stepping switch. The drum advances one step each time the Enter key is depressed. Upon
depression of the N onmerchandise key, the drum advances
to the nonmerchandise field. Upon depression of the Total
key, the drum advances to its original start position.
The punched-tag reader was designed to accept the
twenty-five column, Dennison tag, or abbreviation of it.
When a sales person or . customer number tag is read, the
tag is returned after read-in, so that it may be retrieved

board operation when a tag has been inserted in the reader.
The actual reading of each character on the tag is accomplished by "sensing" the tag with five· pins so connected
through linkages as to actuate switches in the presence of
holes.
When the Total key is depressed, the Sales Recorder is
placed in the output mode causing the cash drawer to open
and allowing the print-out of the sales check to begin.
However, in a credit type sale, if the purchaser did not
have a good credit rating, a Hold button will light, providing a bad credit indication to the operator, and preventing
print-out. Print-out will proceed if the operator presses the
Hold button.
The same mechanism that provides the character display
for the keyboard is also the principle portion of the output
printer. The output printer consists of ten numerical print
wheels fabricated from nylon. These wheels are set up sequentially directly from the information sent by the Input-

Fig. 3-Partially assembled unit with the Procedure Drum and keyboard in place and the tag reader and printer shown separately.

Fig. 4-Complete assembly with the cover removed.

by the operator. In the case of the merchandise tags, however, the tag is read and retained in the machine, which
prevents littering the counter with merchandise tags. Since
all of the tag information has been recorded on magnetic
tape, the merchandise tags may be removed from a bin and
discarded at the end of the day. The first character of any
tag indicates which of the three types it is and if the wrong
tag is inserted, the logic of the machine is such that the tag
reader stops, and the Clear key must be depressed in order
to retrieve the tag. The tags are read one step or character
at a time, 7-1/2 characters per second. Again, as in the
keyboard, the tag reader stops in a reading position and the
output is verified as a valid binary number, prior to advancing to the next reading position. An electrical interlock
between the tag reader and the keyboard prevents key-

Output Buffer Unit. When a line of print has been set up,
a print platen is released, the line is printed, and the paper
advances to the next print position. The speed of this operation averages about six characters per second. An automatic overprint provides a visual indication on the check
for credits or C.O.D. transactions.
The paper used consists of a three-part sprocket-fed
preprinted form. The first copy is obtained through an ink
ribbon impression and the back copies are carbonless paper,
and require no ribbon for printing. Since the form is preprinted, the paper feed machanism is programmed in conjunction with the printer so as to print only in the correct
blocks or spaces on the check and not in the preprinted portions.
The printer, keyboard, and tag reader are all separate

Baer, Rettig, and Cohen: On-Line Sales Recording System
units that plug into the base assembly. Fig. 3 gives an exploded view of the Sales Recorder while the complete assembled unit is shown in Fig. 4.
BUFFER

The Buffer Unit is a multiplexing device which allows
the transfer of data to and from as many as ten Sales Reco.r-ders. Working on a time-sharing basis, the buffer permits independent operation of each Sales Recorder. A
block diagram of the Buffer Unit is shown in Fig. 5.
The communication between the Sales Recorder and the
Buffer Unit consists of information flow within a closed
loop.
At both receiving ends of the loop the character is
checked for parity errors. If wrong parity is sensed, the
character is rejected and an error displayed at the Sales
Recorder. The keyboard of the Sales Recorder remains
locked until the error is cleared. As mentioned earlier, the
return of the correct character from the Buffer Unit to the
Sales Recorder unlocks the keyboard so that another character may be entered. This double check, together with the
relatively high power used in the transmission and the low
impedance at both ends, makes the system very reliable
from the point of view of the data exchanged, and insensitive to noise and crosstalk.
---------T-----SALES RECORDER

I
I

BUFFER UNIT

KEYBOARD

TAG

READER

--LEGEND~-M~MH~WNTR~~-

Fig. 5-Block diagram of the Buffer Unit.

The two information trunks are actually contained in a
single cable that connects each Sales Recorder, via a junction box, to its designated input terminal at the Buffer
Unit. To best describe the functioning of the Buffer Unit,
its operation will be divided into two modes:
1) Input: receiving information from the Sales Recorder-or entering transaction data, and
2) Output: sending the processed information to the
Sales Recorder-or the printing of the sales slip.
Let us consider the input mode. The 6-bit binary character in the form of voltage levels created by either depressing a key or reading a character from a tag in the tag
reader is transmitted by cable directly to the input magnetic core bank. (See Fig. 5.) There is a bank of input
cores associated with each Sales Recorder. The cores serve
the dual function of noise suppression and gating.

253

Also associated with each Sales Recorder is a portion of
a magnetic drum called a sector. That part of the drum containing all the sectors is designated as the Sector Channel.
Each sector has a storage capacity of 360 5-bit characters.
The sector is divided into preassigned fields corresponding to 1) the items listed on the Procedure Drum for input information, and 2) those items required for output
information. These variable fields, once established for the
desired application, are fixed in length. Thus, for instance,
there is a three-character field for the sales person number.
If insertion of a fourth digit is attempted, the Clear light
at the Sales Recorder wi11light, indicating that the capacity
of the field has been exceeded.
A special indexing track on the drum, called the Marking Pulse Track, provides the indexing mark within each
sector to indicate the location of the character last operated
on and the field starting position.
A magnetic core shift register in synchronism with the
magnetic drum provides a read-out pulse that transfers the
information from the input cores to the Input-Output Register at the time that the sector associated with the respective Sales Recorder is accessible. Since the arrival of the
information from the Sales Recorder is asynchronous with
respect to the read-out pulse, core logic is provided to assure that a complete character is actually placed in the
Input-Output Register. Once in this Register, the information is checked for parity, and if an error is sensed an error control flip-flop is set.
1£ a numeric character is present, it is written on the
sector at the location specified by the marking pulse. However, if the character in the register is one of the six operational commands, the matrix in the control unit is activated, so that the specified command level is generated.
These command levels enable their related logic to perform
the required function. In either event, a parity error prevents the processing of the character in the register.
The information transferred at the start of the sector
cycle remains in the Input-Output Register for a time interval equivalent to a sector period, approximately 4.1 msec.
All operations pertaining to the character received are executed within this period. Therefore, it is the Input-Output
Register that is being time shared, allowing the sequential
sampling of information in each input core bank.
Clock signals are provided so that synchronous timing
with the magnetic drum occurs.
Just prior to clearing the Input-Output Register, near
the end of the sector period, the contents are transferred
to the output core bank. If the error control flip-flop
was set, the Input-Output Register is cleared before the
contents are transferred, so that the output core bank contains all zeros. Shortly thereafter, the information is read
out of the cores to fire their associated thyratrons, thus
forming the return character which activates the decoding
relay matrix in the Sales Recorder. This in turn terminates
the transmission of the character to the Buffer Unit, and

254

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

also extinguishes the thyratrons by removing their plate
voltage. As was the case for the input cores, there is a bank
of output cores and thyratrons associated for each Sales
Recorder.
The information levels transmitted by the Sales Recorder remains present for several complete auxiliary memory cycles, about 130 msec, at which time an error indication
is made, unless the correct character echo is returned
earlier. Since all zeros are returned to the Sales Recorder
when parity error is detected, the information transfer is
not interrupted. Therefore, a second or third chance is
afforded to correctly process the transmitted character,
and thereby minimize the possibility of transient errors
stopping the operation.
Let us now consider the output mode where information is being printed on the sales slip. In the output mode,
the information to be transferred to the Sales Recorder
originates from the sector storage. However, a character
is only transmitted when a request for information is made
by the Sales Recorder, and then only one character is
transferred per request. The buffer, in processing this request, will allow a character in sequence, as indicated by
the marking pulse, to be transferred from the respective
sector to the input-output register. Once in this register,
it can be read into the output cores and the thyratrons fired
in much the same manner as is done during the input mode.
The central computer processes a transaction only if the
input of data from a Sales Recorder has terminated. This
is signified by a total symbol entered at the beginning of
the respective sector when the total key at the Sales Recorder is depressed. Before the buffer can respond to the
Sales Recorder's request for a character, the processing
of its associated sector by the computer must be completed.
After the central computer has signaled that this processing
is completed, it has no further control over the respective
sector.
The seventh line in the output trunk furnishes the overprint level to the Sales Recorder. Another thyratron per
Sales Recorder is provided and operated by a special control flip-flop. This control storage is activated by a program
controlled character coming from the sector storage.
Each sector may be in a different operational state at a
given time independent of one another, and thus there is
no interference or interruption of communication between
a Sales Recorder and its associated sector storage.
RECORDER CENTRAL

The central computer, referred to as Recorder Central, is
a general-purpose internally programmed digital device
with a fixed order code. As shown in Fig. 6, it comprises a
magnetic drum, an arithmetic unit, a control unit including
a clock and control pulse generator, a small high-speed
magnetic core memory, and an operator console to provide
program and operator control. The magnetic drum contains, in addition to the sector channel, a random access

stock and credit reference file, the program storage, the
work space for transaction processing, and the necessary
timing tracks. The small high-speed memory of twentycharacter capacity is used for all operations, except transfers within the drum.
The computer is a one-address, variable word, numeric
machine. An instruction word consists of an order code
of two characters, and an address area consisting of four
characters. The order code was specially designed to
facilitate file processing as well as rapid calculation. The
order code contains instructions for communication between a Sales Recorder sector and the computer, arithmetic computations, decision and control operations, file
processing, and console input and output via paper tape
and monitor printer.
COMPUTER
MAGNETIC DRUM

(

BUFFER
UNIT

LEGEND
~OATAPATH
.~

CONTROL PATH

Fig. 6-Block diagram of the Recorder Central.

Upon recognition of the Total symbol by the computer,
the entire contents of that sector are transferred to the
working storage. The sequence of words and their positions within the sector remain the same. Thus, transaction
processing requires a minimum of editing and rearrangement for output printing.
The input data in the working storage is analyzed to
determine how the transaction is to be processed. If the
transaction requires the verification of a customer's credit,
the customer's charge number is processed against a credit
exception file. All stock numbers are passed against the
stock reference file to determine price and city, state and
federal tax information. These data are obtained for all merchandise items sold in the transaction. Prices are extended,
subtotal and total calculated, and required information
listed for printing on the sales slip, after which the contents of the working storage are written out to the transaction record magnetic tape. After receiving a check signal
from the tape station, the information in the working storage is then transferred to the respective Sales Recorder
sector and the Input-Output Buffer notified that transaction processing is completed.
To determine the beginning and end of the variable sized
items on the drum processed by the Recorder Central, item
markers are used. The working storage can be changed in

Baer) Rettig) and Cohen: On-Line Sales Recording System
layout to represent any kind of sales check or business form
corresponding· to the Sales Recorder sector layout. The
ability to vary the reference storage message sizes to conform to variable word requirements allows great efficiency
to the use of the drum.
Access to the file storage is hastened by avoiding long
indexing searches. Messages within the file may be either
extracted, deleted, changed, or added, by separate orders.
Variable sized criteria may be used with these instructions
to extract desired information. Thus, one may be interested
in all swim suits, in all bikini swim suits, in all bikini swim
suits with blue polka dots, by adjusting the criterion accordingly.
To gain access to the desired data of a message, a
mathematical transformation on the criterion of the message is used. This allows minimum delay in locating the
messages and thus speeds up over-all transaction processing. More important) however, by avoiding the use of indexing routines,messages can be entered into the reference
storage or extracted from it without the requirement of
prior sorting and collating. Thus, the problems of file maintenance are considerably simplified and external processing
is appreciably reduced.

255

All transaction processing operations are carried out
automatically. If information concerning daily transactions
is required at any time by the Electronic Data Processing
System the current transaction record magnetic tape can
be remotely disconnected, and a new tape connected, by
the Sales Recording System.
The Recorder Central contains many built-in checking
features. Redundancy checking is used throughout the
equipment to determine errors in transmission of characters and to isolate their sources. Arithmetic operations
are repeated and results compared. Orders are checked
before they are carried out. These, among others, are designed to insure against incorrect processing. However, in
a system used for on-line processing the ability to maintain
continuous operation is of paramount importance. Thus,
the Recorder Central is designed to attempt to overcome
any error a fixed number of times before it will stop operation. This will discriminate between transient errors and
those due to catastrophic breakdown. In the latter case, a
complete set of machine status indicators is available at
the console, specifying exact portions of an order in which
failure had occurred for ease and rapidity of maintenance.
Plug-in type module construction is used to facilitate
troubleshooting, preventive maintenance, and replacement
in case of failure.
CONCLUSION

Fig.

7~Recorder

Central and Input-Output Buffer Racks
with console and power supply.

The control console may be used to monitor system operation or to provide means for manual control of the Recorder Central. A view of the Recorder Central is provided
in Fig. 7. All control flip-flop indicators in the machine are
displayed as an aid in maintenance and program debugging. All areas of the computer may be interrogated from
the console. Information can be introduced either manually
or by means of paper tape. A monitor printer is provided
for printing the contents of the high-speed memory or any
portion of the auxiliary memory when desired. Marginal
checking facilities are also controlled from the console.

The Sales Recording System represents a great step
forward in providing the means for data integration within
a department store. The Associated Merchandising Corporation's research installation has demonstrated beyond a
doubt that on-line Sales Recording systems are a reality.
Here, for the first time, a variety of transaction types, as
broad as the store· desires, may be processed directly from
a point-of-sale unit, with complete computation performed
by a fast, accurate, and versatile high-speed computer, including an automatically printed sales slip.
But more than this, the point-of-sale unit can be used
to either interrogate the reference file and thereby gain
immediate access to any desired information, or actually
enter new reference information, directly from its remote
location.
It is important to note that this information need not be
in any ordered form. Furthermore, a magnetic tape record,
made for each transaction entered, provides a direct and
reliable means of furnishing the information to an Electronic Data Processing System for data handling. The
Electronic Data Processing System can have access to the
magnetic tape at any suitable time without interrupting
transaction processing by Recorder Central.
ACKNOWLEDGMENT

The authors wish to acknowledge the contributions of
Albert Burstein, Edward Damerau, Robert Grapes, Andrew Ling, and Felipe Tanco in preparing this paper.

256

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE
Discussion

J. T. Wallace (Eastman Kodak Co.) :
Is the Dennison tag reader available separately and what is the general means of
reading the tag?
Answer : Yes, it is. The reading is accomplished by pins going through the
holes and consequently making switches.
N. J. Dean (Ramo-Wooldridge Corp) :
What is the capacity, speed, and formtape, drum, etc-of the random access
reference file?
Answer: The random access reference
file which is a portion of the drum has a
capacity of approximately 72,000 characters. The speed is approximately 1500
rpm with a character rate of slightly less
than 100 kc.
Question: How many characters in· a
complete transaction? If more than one,
is the entire message buffered before presentation to the computer?
Answer: There could be up to 360 characters. That is the full capacity of a sector, there being 10 sectors for the sector
storage.
The normal transaction will have a
smaller percentage and therefore the sector capacity was laid out so that you could
process a maximum of five merchandise
transactions and five nonmerchandise type
transactions per sale.
An entire message is retained in the
sector until the "Total" indication has been
given. At this time, the complete sector is
transferred ino the "Working Storage."
Mr. Dunham: Do the marking pulses
and clock signals originate from the magnetic drum?
Answer: Yes.
Question: Will this point-of-sale unit
include customer billing?
Answer: Ultimately it could, by providing all the necessary information to an
Electronic Data Processing System.
Question: How do you plan to handle
the sales slip; namely, is a copy to be re-

turned to the customer-a generally accepted billing practice?
Answer: One copy of the sales slip is
given to the customer at the time of purchase. The other two are retained by the
store; one to be included with the monthly
statement for charge sales, and the other
for record purposes.
Question: I would like an idea of the
general size and cost of a complete system
for a medium size department store.
Will one computer handle any number
of sales recorders? How many pieces of
intermediate equipment are required?
What provisions would be made for
leaving an alternate means of recording
sales in the event of computer failure for
even a short time during the day?
Answer: For cost figures our sales department should be contacted.
The number of Sales Recorders that
may be handled by a single computer is
a function of the computer handling rate
as well as the random access rate to the
reference file.
A paper tape punch could be used as a
stand-by unit to capture the required information for future data processing,
while the sale may be handled by normal
sales book method.
Question: What is the expected cost on
each point-of-sales recorder?
Answer: For cost figures our sales department should be contacted.
Question: Do you intend to use telephone lines to connect a remote point-ofsales recorder to files and computer?
Answer: Long-distance hookup, that is,
beyond the confines of a store, has not
been considered as yet.
Question: How many point-of-sales recorders can be attached to your system?
Answer: The system is a pilot system,
and therefore provision for only ten Sales
Recorders was made.
Question: What is the maximum number of input-output units you can couple

to the described system?
Answer: For this pilot system, thirty
Sales Recorders were set as the maximum,
and still maintain an average transaction rate of one a minute with an average
queueing rate of one second for the transaction processing.
Question: What provisions are made
for multiplexing or sequencing between the
many point-of-sale units and the central
computer?
Answer: Each Sales Recorder has a
sector allocated to it. When the sector
comes under the read heads, the associated Sales Recorder will receive information if it's in the output mode or furnish
information if it's in the input mode.
Question: Can a department store economically justify the use of these Sales
Recorders?
Answer: This is a question for the individual department store to answer.
Question: How much time is required
to process a transaction at the point of
sale? Did you find that this time requirement was in any way objectionable to the
customer?
Answer: The transaction time varies
with the data to be entered, of course.
However for average length transactions
a rate of less than one minute per transaction may be sustained. This was found to
be in no way objectionable to the customer.
Question: If the point-of-sales unit can
enter new reference information, as on
credit, can a customer operate one pointof-sales unit in the absence of the salesman and thus improve his credit rating
prior to making a large purchase?
Answer: No. Although special coding is
provided to allow entry of data from the
Sales Recorder, it requires a knowledge of
the code (nonmerchandise code) and a
special sales person number. Adequate
programming checks are provided to accomplish this ..

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

257

Organization of Simulation Councils, Inc.

HIS YEAR Simulation Councils, Inc., accepted
with great pleasure S. N. Alexander's (1957 EJCC
Chairman) invitation to hold its annual meeting in
conjunction with the Eastern Joint Computer Conference.
This allowed members of the regional groups to meet at a
national conference where all members had a definite interest. The Simulation Councils participated in two of the
formally scheduled sessions, and in addition, four informal
sessions were held at which papers were presented by its
members.
The following background information is provided to
familiarize members of the AlEE, ACM, and IRE with
the objective of the Simulation Councils.
Simulation Councils, Inc., is an organization formed
to improve communications at the working level concerning methods of using simulation techniques and equipment to facilitate the study, design, test, and analysis of
physical systems. These objectives are quite broad and
naturally are very much related to the use and application
of electronic computation.
The beginning of this organization occurred in N ovember, 1952, when a group of persons associated with the
operation of general purpose analog computers decided
to hold informal meetings for discussing the details of
problems and methods for their solution, equipment, and
new ideas. It was decided that the minutes of such meetings would be published in an informal newsletter which
would be made available to all those interested, including
those who were unable to attend.
The group holding this first meeting evolved into the
Western Simulation Council. Following the same concept
and procedures, five other Councils were organized: the
Midwestern Simulation Council, Eastern Simulation Council, Central Simulation Council, Southeastern Simulation
Council, and the Canadian Simulation Council. All these
Councils have joined into an international group which is
called Simulation Councils, Inc. The purpose of this
group is to provide a medium for publishing the minutes
of the meetings of all regional groups, and to arrange
annual meetings to which all the regional groups are
invited.
The entire concept of the Simulation Councils is based
on informality. Membership in anyone of the regional
groups may be obtained by attending a meeting and signing the attendance register, or by requesting the Chairman
of that Council to place your name on the mailing list.
Meetings are held approximately every second month
with from 20 to 100 members in attendance. The meetings
usually are held at the computing facilities of organizations
interested in simulation. An attempt is made to choose
I

different hosts for each meeting so members may visit a
wide variety of simulation and computational facilities.
Meetings usually begin with brief talks on a subj ect selected to promote discussion, and an attempt is made
to encourage all present to join in the discussions. This
informal discussion is the essence of the Simulation Coun-,
cils effectiveness; everyone gets a chance to compare his
technique and equipment with others having similar interests and problems.
The Simulation Council ,Newsletter, as informal as the
Council discussions, has appeared every month since N 0vember, 1952. At first it was privately published by
mimeograph, but increased demand made this impractical
and since April, 1955, it has appeared as a separate and
editorially autonomous section of Instruments and Automation. Individual issues of the Newsletter are as different as the wide variety of subjects and the personalities
of people discussing them. The result is that while reporting month after month the offhand remarks as well
as the serious thoughts of those developing the techniques
and designing the equipment for simulation, the N ewsletter reflects progress in the allied fields of analog and
digital computation and data processing as well.
Anyone interested in the Simulation Councils is invited
to contact the Steering Committee Chairmen of the various
regional groups listed below.

Western Simulation Council
Dov Abramis, Convair, Pomona, Calif.
Midwestern Simulation Council
Warren Jackson, Standard Oil Co., Midland Building,
Cleveland, Ohio
Eastern Simulation Council
Hideo Mori, Hydel, Cambridge, Mass.
Southeastern Simulation Council
Robert Johnson, Georgia Institute of Technology, Research Area 4, Atlanta, Ga.
Central Simulation Council
James Pierce, Beech Aircraft, Wichita, Kan.
Canadian Simulation Council
F. W. Pruden, Analog Computation and Simulation
Group, Mechanical Engineering Division, National Research Council, Ottawa, Ont., Canada
ABSTRACTS OF PAPERS

Papers presented by members of t~e Simulation Councils, Inc., at the four informal sessions during EJCC are
abstracted on the following page.

258

PROCEEDINGS OF THE EASTERN COMPUTER CONFERENCE

1) Physical Simulation in Airplane Control System Problems,
P. G. Hurford, McDonnell Aircraft Corp. This paper describes
the use of physical simulation including the pilot as an aid in the
solution of lateral control problems of modern jet fighters. To
accomplish this, an analog computer and control system are combined with a movable chair which imparts to the pilot the "feel"
of rolling motions. The effects of various chair, control systems,
and airframe parameters are determined with this system.

2) Design and Utilization of a Three-Axis Simulator, M. Paskman and R. Edwards, Aircraft Armaments, Inc. The time and cost
savings resulting from utilization of a three-axis simulator are
discussed. The design evolution of a three-axis, gimballed system
is described, including specifications, mechanization, and analysis
of the system. Test procedures are outlined, after which the use
of a general purpose analog computer to solve typical control
system problems is described.

3) Synthesis of Closed-Loop System by Means of Analog
Computers with Real Gyros or Accelerometers in the Loop,
B. W. McFadden, Micro Gee Products, Inc. Presented are the
advantages and techniques in the synthesis of a closed-loop system
(such as an autopilot or autonavigational system) by including the
real gyro or accelerometer in the loop with an analog computer.
Also given are the characteristics of a simulation table with a
threshold of less than one second of arc which makes it possible
to determine closed-loop performance about null under the influence
of real, small discontinuous nonlinearities such as friction, noise,
and deadband.

4) A Discussion of the Procedures and Practical Problems Relating to Real-Time Simulation Using Control System Hardware,
Eaton Adams, Jr., Convair, Fort Worth, Texas. The techniques
and problems involved in real-time flight simulation using hardware
are described. A three-degree-of-freedom missile trajectory problem, including nonlinear aerodynamics, is used as an example.
Problems of analysis and correlation with analytic results when
nonlinear hardware is included in the simulation are discussed,
along with other difficulties arising in such a simulation.

5) A New Dead Time Simulator for Electronic Analog Computers, Millard Brenner and Jerome D. Kennedy, Electronic Associates, Inc. Up to now, the accurate simulation of dead time,
performance of auto and cross correlations, and long-term arbitrary
function storage have been difficult to achieve on an electronic
analog computer. This paper deals with the applications and operational principles of the SIMULAG, a variable delay, multichannel magnetic tape unit developed by Electronic Associates, Incorporated.
6) A. C. Diode Fun,ction Generators, C. L. Cohen and
D. S. Peck, Nuclear Products-Erco Division of ACF Industries,
Inc. Circuits employing silicon diodes as used in aircraft flight
simulators are described. The circuits involved perform function
generation such as slope changing, limiting, jump functions, absolute
value generation, and magnitude selection.
7) Future of the Transistor Analog Computer, Robert Bruns
and George Milligan, Jet Propulsion Laboratory, California Institute of Technology. A portable analog computer employing transistors has been developed to study the possibility of analog computers
employing semiconductor devices only. The transistor operational
amplifier and servomultiplier circuits for this computer are described, along with the performance capabilities of these components. The direction of future analog computer design is discussed.

8) Open-Shop Operation of a Large Real-Time Simulator,
Stanley Rogers, Convair, San Diego, Calif. The problem of
maximizing the use of simulators under open-shop conditions is
reviewed. Topics include scheduling, operator training, equipment
reliability and maintenance, problem-checking methods, and suggested future trends for large simulators.

9) A Real-Time Simulation System for Use with an Analog
Simulator, Robert M. Beck and Max Palevsky, Packard-Bell Computer Corp. Reported are some newly developed digital devices
which can be employed with analog computers. These devices include an extremely high speed incremental computer for computing
nonlinear analytic functions, a function generator that employs
photographic techniques for storage, and, a 0.01 per cent analogto-digital and digital-to-analog converter. The converter also performs multiplication and division within the conversion process.
10) Simulation of Aircraft Landing Gear Dynamics on the
Geda Analog Computer, P. J. Hermann, Goodyear Aircraft Corp.
A simulation of the vibration dynamics of a landing gear has been
set up in order to study the effects of braking torques and other
influences in exciting the landing gear into undesirable vibration
amplitudes. Assumptions in formulating the physical model of the
landing gear are discussed, as well as the development of the
equations of motion. The difficulties in determining numerical values
for the various constants are related, and the results of the computer simulation are presented.
11) Direct Simulation on Analog Computers through Signal
Flow Graphs, Louis P. A. Robichaud, Canadian Armament Research and Development Establishment. For physical systems which
can be considered as made up of combinations of various elements,
it is not necessary to write out the equations before determining
the analog computer circuit. A procedure is presented for going
directly from a physical system to an analog computer representation of that system. Each physical unit is represented by a computer unit such that all terminal variables are in evidence in that
unit as they are in a matrix or flow graph representation of the
physical unit. The computer units then are interconnected in a
manner similar to the physical units.

12) The Use of Quaternions in Simulation of the Motion of
Rigid Bodies, A. C. Robinson, Wright Air Development Center.
In simulation involving real three-dimensional coordinate transformations, it is customary to represent orientation by Eular angles
or direct cosines. The use of the four-parameter quaternion notation which avoids both "gimbal lock" and redundancies is outlined. It is shown that from the standpoint of accuracy and
amount of equipment required, the quaternion method is superior
to either of the two usual techniques. Other advantages and
limitations are set forth.
13) On the Loop and Node-Analysis Approach to the Simulation of Electrical Networks, Joseph Otterman, University of
Michigan. Simulation of an electrical network on an analog computer by the loop or node-analysis approach is quite often unsatisfactory because of hidden regenerative loops. Such regenerative loops arise when the computer setup contains more integrators
than the order of the differential equation describing the network.
Instability may result because the actual computer components in
the loop containing an excess integrator depart from ideally exact,
prescribed values. Presented is a procedure for tracing the loop
currents in such a way that there is one-to-one correspondence
between the number of integrators in the simulation setup and
the count of independent energy-storing elements in the network,
i.e., the order of the differential equation describing the network.
The generality of the procedure proves that it is always possible
to trace the loop currents in such a way that excess integrators are
avoided. A parallel procedure for node analysis is discussed briefly.

14) Application of High Speed Compressed Time Scale Computer to Engineering Problems, Joseph Miasnick, General Electric
Co. Application is made of a high-speed compressed time scale
computer to certain engineering problems. The place of this type
of computer in the over-all problem-solving picture is described.
Specific application is made to the simulation of the performance
of a starter-engine combination. The simulation is used to determine the effect in performance of varying cutout speed, torquespeed characteristics, and engine application. The computer also
is used to determine the inertia of the test stand.

Organization of Simulation Councils, Inc.
15) Simulator for Use in Development of let Engine Controls,
Emile S. Sherrard, National Bureau of Standards. Recommendations and cost estimates for a simulator capable of representing
a twin-spool, after burning, turbojet engine and its controller are
given. The simulator is intended as a tool to be used by a group
of engineers engaged in developing engine control systems. The
simulator is designed to determine stability and performance of
the engine control system, to be useful in both early and late
stages of the development, and to operate in real time so control
hardware may be operated with a simulated engine.
16) Ducting Air-Flow Characteristics By Electrical Circuit
Measurements, William H. Sellers, General Electric Co. The weIlknown dynamical analogies are developed and reviewed for electrical, mechanical, and acoustical systems. The analogous electrical
circuit then is derived for a physical system consisting of a
cylindrical tube, containing a butterfly valve arbitrarily located
within the tube. An actual circuit consisting of inductances, capacitors, etc., is then constructed and the values of voltages and
currents are recorded by a brush recorder. These recorded values
are analogous to the pressure and flow of the fluid within the
duct. Photographs are presented depicting the actual wiring and
measurement arrangements. Also, a comparison solution, obtained
by a standard analog computer, relates the relative error and
reliability of the direct circuit measurement method.

259

17) The Real Time Simulation of Turbo-let and Turbo-Prop
Engines and Controls, D. L. Dresser, Allison Division, General
Motors Corp. Some problems faced by the designer of control
systems for turbo-jet and turbo-prop engines are outlined. The use
of the analog computer as a tool in designing controls for future
engines and testing hardware is discussed. Two specific examples
of full-scale simulation are presented. The first one is a turbo-jet
simulation based upon component turbine and compressor date, and
the second one is a turbo-prop simulation which uses test-stand
dynamometer data. Also, two linearized procedures are given. Computer techniques for representing a variety of nonlinearities are described.
18) Noise and Statistical Techniques in Analog Simulation,
Henry Low, The Ramo-Wooldridge Corp. The role of the analog
computer in obtaining the impulse response of a linear system is
discussed, including the adj oint method for obtaining the impulse
response of time-varying linear systems. Techniques for treating
nonlinear systems on an analog computer are discussed. The types
of statistical quantities corresponding to random inputs to the
analog computer are described, and an illustrative example is
worked out. The theory of operation of various types of analog
computer noise generators is explained. Techniques for measuring
the statistical characteristics of low-frequency noise generators
are presented.

Source Exif Data:
File Type : PDF
File Type Extension : pdf
MIME Type : application/pdf
PDF Version : 1.3
Linearized : No
XMP Toolkit : Adobe XMP Core 4.2.1-c041 52.342996, 2008/05/07-21:37:19
Producer : Adobe Acrobat 9.0 Paper Capture Plug-in
Modify Date : 2008:11:16 12:16:50-08:00
Create Date : 2008:11:15 16:09:09-08:00
Metadata Date : 2008:11:16 12:16:50-08:00
Format : application/pdf
Document ID : uuid:fe7c0349-ee25-44b9-8de5-4657f957b494
Instance ID : uuid:d1ba3e04-6555-4d8c-bdd0-063bd0a96c96
Page Layout : SinglePage
Page Mode : UseOutlines
Page Count : 260

EXIF Metadata provided by EXIF.tools

1957 12_#12 12 #12

1957-12_#12 1957-12_%2312

Navigation menu

Versions of this User Manual:

Views

Navigation