1961 05_#19 05 #19

1961-05_#19 1961-05_%2319

User Manual: 1961-05_#19

Open the PDF directly: View PDF .
Page Count: 668

Download
Open PDF In Browser	View PDF

Proceedings of the

WESTERN JOINT COMPUTER CONFERENCE
~

. . . . . . . . . . . " . . . . . u· . . . . . I.uuu'uu,"uu,uuu"u,""""""'"'''"uu"'"'''"",u""uuoo,,,,uuuuuu",uoooooouoouoooouoouoooo,,oooooooooouuoouuuuu

May 9- 11, 1961

EXTENDING MAN'S INTELLECT

Los Angeles, California

Spo·nsors:

THE INSTITUTE OF RADIO ENGINEERS
Professional Group on Electronic Computers

THE AMERICAN INSTITUTE OF ELECTRICAL ENGINEERS
Committee on Computing Devices

THE ASSOCIATION FOR COMPUTING MACHINERY

Vol. 19

Price $4.00

PROCEEDINGS OF THE
WESTERN JOINT COMPUTER CONFERENCE

PAPERS PRESENTED AT
THE JOINT IRE-AIEE-ACM COMPUTER CONFERENCE
LOS ANGELES, CALIF 0, MAY 9-11, 1961

Sponsors
THE INSTITUTE OF RADIO EN·GINEERS

Professional 'Group on Electronic Computers
THE AMERICAN INSTITUTE OF ELECTRICAL ENGINEERS

Committee on Computing Devices
THE ASSOCIATION FOR COMPUTING MACHINERY

Published by
WESTERN JOINT COMPUTER CONFERENCE

ADDITIONAL COPIES
Additional copies may be purchased from the following sponsoring
societies at $4.00 per copy. Checks should be made payable to any
of the following societies:
INSTITUTE OF RADIO ENGINEERS
1 East 79th Street, New York 21, N.Y.
AMERICAN INSTITUTE OF ELECTRICAL ENGINEERS
33 West 39th Street, New York 18, N.Y.
ASSOCIATION FOR COMPUTING MACHINERY
2 East 63rd Street, New York 21, N.Y

© 1961 by
National Joint Computer Committee

The ideas and opinions expressed herein are solely those of the authors,
and are not necessarily representative of, or endorsed by, the 'WJCC
Committee or the NJCC Committee.

Manufactured in the U.S.A. by
Griffin-Patterson Co., Inc., Glendale, Calif.

WESTERN JOINT COMPUTER CONFERENCE COMMITTEE
General Chairman

Walter F. Bauer, Ramo-Wooldridge, Canoga Park, California

Vice Chairman . .

Keith W. Uncapher, The RAND Corp., Santa Monica, California

Conference Administrator

Robert W. ·Rector, Space Technology Laboratories, Los Angeles, Calif.

Program Chairman . . . .

Cornelius Leondes, UCLA, Los Angeles, California

Associate Program Chairman

Paul Armer, The RAND Corp., Santa Monica, California

Associate Program Chairman

J. D. Madden, System Development Corp., Santa Monica, Calif.

Associate Program Chairman

John McLeod, Convair-Astronautics, San Diego, California

Publications . . . . .

Glenn Morgan, IBM Corp., Los Angeles, California

Hotel Arrangements

William Dobrusky, System Development Corp., Santa Monica, Calif.

Public Relations

Santo Lanzarotta, DATAMATION, Los Angeles, California

Finance ..

William S. Speer, United Aircraft Corp., Costa Mesa, California

Registration

Marvin Howard, Ramo-Wooldridge, Canoga Park, California

Exhibits ..

Richard H. Hill, Ramo-Wooldridge, Canoga Park, California

Printing and Mailing

L. C. Hobbs, Aeronutronic Div. of Ford Motor Co., Newport Beach, Calif.

Trips . . . . . . . .

Joel Herbst, Ampex Computer Products Co., Culver City, California

Women's Activities

Phyllis Huggins, Bendix Corp., Computer Div., Los Angeles, California

Public Relations Consultant

Lynn-Western, Los Angeles, California

Exhibits Manager . . . . . .

John Whitlock, Oakton, Virginia

iii

NATIONAL JOINT COMPUTER COMMITTEE
Chairman
Morris Rubinoff
Moore School of Engineering
University of Pennsylvania
Philadelphia, Pennsylvania

Vice Chairman
J. D. Madden
System Development Corp.
Santa Monica, California

Secretary-Treasurer
Margaret R. Fox
National Bureau of Standards
Department of Commerce
Washington, D.C.
AlEE Representatives
Morris Rubinoff
Moore School of Engineering
University of Pennsylvania
Philadelphia, Pennsylvania

IRE Representatives
Richard O. Endres
Rese Engineering Co.
Philadelphia, Pennsylvania
Frank E. Heart
Lincoln Laboratories
Lexington, Mass.

G. L. Hollander
Hughes Aircraft
Fullerton, California

Charles W. Rosenthal
Bell Telephone Laboratories
Murray Hill, New Jersey

R. R. Johnson
General Electric Co.
Phoenix, Arizona

Willis H. Ware
The RAND Corporation
Santa Monica, California

C. A. R. Kagan
Western Electric Co., Inc.
Princeton, New Jersey

A eM Representatives
Paul Armer
The RAND Corporation
Santa Monica, California

Ex-Officio Representatives
Harry D. Huskey
University of California
Department of Mathematics
Berkeley, California

Walter M. Carlson
E. I. du Pont de Nemours
Wilmington, Delaware

R. A. Imm
IBM Corporation
Rochester, Minn.

J. D. Madden
System Development Corp.
Santa l\ionica, California

Arnold A. Cohen
Remington Rand UNIVAC
St. Paul, Minnesota

R. W. Hamming
Bell Telephone Laboratories
Murray Hill, New Jersey

Bruce Gilchrist (ACM)
IBM Research Center
Yorktown Heights, New York

Headquarters 'Representatives
R. S. Gardner (AlEE)
American Institute of Electrical Engineers
New York, New York

iv

L. G. Cumming (IRE)
The Institute of Radio Engineers
New York, New York

FOREWORD
The technical papers included in these Proceedings represent all the
technical presentations made at the Western Joint Computer Confer~
ence. In addition, there is included a message from the Chairman of
the National Joint Computer Committee, Dr. Morris Rubinoff. We are
proud of the excellent contributions recorded here, so many of which
directly support the 1961 WJCC theme "Extending Man's Intellect."
Also, we are happy that we have been able to make these Proce:dings
available at the time of the Conference, thus enhancing the benefits
of the Conference to registrants and making available the information
in a timely manner.
It should be recognized. however, that the papers presented herein
have not been selected by the usual procedures wherein a paper is
refereed as to its appropriateness and edited for its content. Neither
the NJCC nor the WJCC Committee can take responsibility for the
accuracy of the facts or opinions expressed. We are confident, however, that the overwhelming majority of the papers presented here are
responsible in all ways. Many papers were called but few were chosen;
we are happy to record them here for the continuing advance and
lasting annals of information processing technology.
WALTER F. BAUER
General Chairman
1961 Western Joint Computer
Conference

v

MESSAGE FROM NICC CHAIRMAN
(3) Prepare, publish, and disseminate information of
a tutorial nature to laymen, high school teachers
and students, government offices and officials, etc.

This is an historic occasion. The close of this 1961
Western Joint Computer Conference will signal the
change-over in administration of Joint Computer Conferences from the National Joint Computer Committee to
the American Federation of Information Processing Societies (AFIPS), with broader scope and greater flexibility. As you know, AFIPS is a society of societies organized to represent through a single body the professional
societies of the American computer and data processing
world. The enthusiastic response to the formation of
AFIPS is highly gratifying and lends encouragement, confidence and a sense of mission to those whom you have
charged with conducting its activities.

( 4 ) Maintain relations among American and foreign
technical societies through conferences and symposia, cooperation with other societies in organizing sessions at their conferences, provide reference material to other societies on the computational sciences.
(5) Maintain membership in the International Federa-

tion of Information Processing Societies (IFIPS).
(6) Aid in certain actions of member societies involv-

There are times when the path to the future is best
appreciated through a re-examination of the past. I would
like to quote from a letter dated December 15, 1959,
written by the late Chairman of NJCC, Professor Harry
Goode, who contributed so much both to NJCC and to
the birth of AFIPS:

ing participation and cooperation by more than
one society.
(7) Sponsor the ICC's."

The Constitution of AFIPS reflects these views in their
entirety. With your frequently demonstrated cooperation
and support, the Board of Governors of AFIPS will
continue to conduct our successful Joint Computer Conferences and to represent the United States in our International Federation, IFIPS. As new societies join the
Federation, it will gradually provide the hoped-for broad
representation of the American information processing
profession. We will seek to establish AFIPS as the information center on data processing including not only
bibliographies of written material, but also a calendar of
events of computer activities in the United States and
throughout the world, a roster of individuals active in
information processing, and a current file of developments
in progress or recently consummated. We plan to establish
a speakers' bureau to carry information on the information
processing field to educational institutions and professional
societies. We plan to establish a public information committee which, through the media of personal contacts, press
releases and tutorial articles, will make available to laymen,
to government agencies, to affiliated and member societies
and to the profession as a whole, the present status and the
probable future of information processing in the United
States.

"I believe the major objective in the formation of the
society is to provide for information flow in all other instances than those provided for by the individual societies to their members.
"There are four types of such flow:
( 1) Information flow between members of information

processing societies nationally.
(2) Information flow between our national informa-

tion processing society and foreign information
processing societies.
(3) Information flow between societies in the informa-

tion processing profession and other professions.
(4) Information flow from the information processing

societies to the general and educational public.
"If we can recognize a firm set of objectives such as
these (which of course need to be rewritten into a proper
set of words), then what the society is to do is relatively
clear-cut.

"The functions follow immediately from the objectives:
(1) Act as the American representative body on mat-

I trust that with your continued cooperation and support
our efforts will meet with a long string of successes.

ters related to computing application and design,
in a broad area of computational and information
processing sciences.
(2)

Respectfully submitted,

Advance the field by stimulating research into new
aspects of computer sciences emphasizing the
cross-pollination of ideas among member societies.

Morris Rubinoff, Chairman
National Joint Computer Committee

vi

TABLE OF CONTENTS
DIGITAL SIMULATION

Page

1.1
1.2
1.3
1.4

"Simulation: A Survey" by H. H. Harman, System Development Corp ........................................................ .
"Management Games and Computers" by J. M. Kibbee, Remington Rand UNIVAC ..................................
"An On-Line Management System Using English Language" by A. Vazsonyi, Ramo-Wooldridge ................
"Application of Digital Simulation Techniques to Highway Design Problems" by A. Glickstein,
S. L. Levy, Midwest Research Institute ..............................................................................................................
1.5 "The Use of Manned Simulation in the Design of an Operational Control Systen( by M. A. Geisler,
W. A. Steger, The RAND Corp .........................................................................................................................

11
17
39
51

MICROSYSTEM ELECTRONICS
2.1 "A Survey of Microsystem Electronics" by Peter. B. Meyers, Semiconductor Products Division of
Motorola, Inc .......................................................................................................................................................
2.2 "Testing of Micrologic Elements" by R. Anderson, Fairchild Semiconductor Corp .........................................
2.3 "Interconnection Techniques for Semiconductor Networks" by J. S. Kilby, Texas Instruments, Inc .................
2.4 "Microsystem Computer Techniques" by E. Luedicke, A. Medwin, Radio Corporation of America ........~ .....

63
75
87
95

MODELING HUMAN MENTAL PROCESSES
3.1
3.2
3.3
3.4

"Modeling Human Mental Processes" by H. A. Simon, Carnegie Institute of Technology .............................. ll1
"The Simulation of Verbal Learning Behavior" by E. Feigenbaum, University of California .......................... 121
"Simulation of Behavior in the Binary Choice Experiment" by J. Feldman, Universify of California ............ 133
"Programming a Model of Human Concept Formulation" by C. 1. Hovland, E. B. Hunt, Yale University .. 145

RECENT ADVANCES IN COMPUTER CIRCUITS
4.1 "Parallelism in Computer Organization Random Number Generation in the Fixed Plus Variable Computer
System" by M. Aoki, G. Estrin, University of California, T. Tang, National Cash Register Co................. 157
4.2 "The CELLSCAN System~a Leucocyte Pattern Analyzer" by Kendall Preston, Jr., Perkin-Elmer Corp ..... 173
4.3 "Application of Computers to Circuit Design for UNIVAC LARC" by G. Kaskey, N. S. Prywes,
H. Lukoff, Remington Rand UNIV AC ................................................................................................................ 185
4.4 "Wide Temperature Range Coincident Current Core Memories" by R. S. Weisz, N. Rosenberg, Ampex
Computer Products Co ................................................................................... :................................................... 207
PROBLEM SOLVING AND LEARNING MACHINES
5.1 "Descriptive Languages and. Problem Solving" by Marvin Minsky, Massachuetts Institute of Technology ...... 215
5.2 "Baseball: An Automatic Question-Answerer" by Bert F. Green, Jr., Alice K. Wolf, Carol Chomsky,
Kenneth Laughery, Massachuetts Institute of Technology ............................................................................ 219
5.3 "A Basis for a Mathematical Theory of Computation" by John McCarthy, Massachusetts Institute
of Technology ...................................................................................................................................................... 225
INFORMATION RETRIEVAL
6.1 "Information Retrieval: State of the Art" by Don R. Swanson, Ramo-Wooldridge .... :.................................. .239
6.2 "Technical Information Flow Pattern" by M. M. Kessler, Massachusetts Institute of Technology ................. .247
6.3 "A Screening Method for Large Information Retrieval Systems" by Robert T. Moore, Princeton University .. 259
AUTOMATA THEORY AND NEURAL MODELS
7.1 "What Is an Intelligent MachineT by W. Ross Asby, University of Illinois .................................................... 275
7.2 "Analysis of Perceptrons" by H. D. Block, Cornell University .......... :............................................................... 281
7.3 "Physiology of Automata" by Murray L. Babcock, University of Illinois ........................................................ 291

vii

TABLE OF CONTENTS, continued
NEW HYBRID ANALOG·DIGITAL TECHNIQUES

Page

8.1 "Combined Analog-Digital Computing Elements" by Hermann Schmid, Link Division, General
Precision, Inc. ...................................................................................................................................................... 299
8.2 "Optimization of Analog Computer Linear System Dynamic Characteristics" by C. H. Single,
E. M. Billinghurst, Beckman Instruments, Inc .................................................................................................. .315
8.3 "Design and Development of a Sampled-Data Simulator" by J. E. Reich, J. J. Perez, Space Technology
Laboratories, Inc. . .................................................... _......................................................................................... 341
8.4 "Digital Clock Delay Generators and Run Counter tor a Repetitive Analog Computer" by T. Brubaker,
H. Eckes, University of Arizona ............................. _......................................................................................... 353
LARGE COMPUTER SYSTEMS
9.1 "Trends in Design of Large Computer Systems" by C. W. Adams, Charles W. Adams Associates.................. 361
AUTOMATIC PROGRAMMING
10.1
10.2
10.3
10.4

"Current Problems in Automatic Programming" by Ascher Opler, Computer Usage Company ..................... .365
"A First Version of UNCOL" by T. B. Steel, Jr., System Development Corporation ..................................... .371
"A Method of Combining ALGOL and COBOL" by J. E. Semmet, Sylvania Electric Products ..................... .379
"ALGY-An Algebraic Manipulation Program" by M. D. Bernick, E. D. Callender, J. R. Sanford,
Philco Corporation .............................................................................................................................................. 389
10.5 "A New Approach to the Functional Design of a Digital Computer by R. S. Barton, Computer
Consultant ............................................................................................................................................. _.............. 393
10.6 "The JOVIAL Checker" by M. Wilkerson, System Development Corporation .................................................. 397

MEMORY DEVICES AND COMPONENTS
11.1 "Factors Affecting Choice of Memory Elements" by Claude F. King, Logicon, Inc ........................................ .405
11.2 "A Nondestructive Readout Film Memory" by R. J. Petschauer, R. D. Turnquist, Remington Rand
UNIVAC ...............................................................................................................................................................411
11.3 "Tunnel Diode Storage Using Current Sensing" by E. R. Beck, D. A. Savitt, A. E. Whiteside, Bendix Corp .. ,427
11.4 "The Development of a New Nol1destructive Memory Element" by A. W. Vinal, Federal Systems-IBM ..... ,443
11.5 "High-Speed Optical Computers and Quantum Transition Memory Devices" by L. C. Clapp, Sylvania
Electronic Systems................................................................................................................................................475
APPLIED ANALOG TECHNIQUES
12.1 "Optimization of a Radar and Its Environment by GEESE, General Electric Electronic System Evaluator
Techniques" by L. Berger, R. M. Taylor, General Electric Defense Systems Department.. ......................... ,490
12.2 "The Spectral Evaluation of Iterative Differential Analyzer Integration Techniques" by M. Gilliland,
Beckman Instruments ............................................... _......................................................................................... 507
12.3 "An Iteration Procedure for Parametric Model Building and Boundary Value Problems"
by Walter Brunner, Electronic Associates, Princeton University ...................................................................... 519
12.4 "Analog Simulation of Underground Water Flow in the Los Angeles Coastal Plain" by D. A. Darms,
E.AJ. Computation Center, H. N. Tyson, IBM Corp .................................................................................... .535
PATTERN RECOGNITION
13.1 "A Self-Organizing Re~ognition System" by R. J. Singer, Aeronutronic Div., Ford Motor Co ..................... 545
13.2 "A Pattern Recognition Program that Generates, Evaluates and Adjusts its Own Operators" by L. Uhr,
University of Michigan, C. Vossler, System Development Corp .................................................................... .555
13.3 "An Experimental Program for the Selection of Disjunctive 'Hypotheses'" by M. Kochen, IBM Corp ........ .571
13.4 "Time-Analysis of Logical Processes in Man" by U. Neisser, Brandeis University .......................................... 579

viii

TABLE OF CONTENTS, continued
Page

COMPUTERS IN CONTROL

14.1 "Computer-Based Management Control" by A. J. Rowe, Hughes Aircraft Co.............................................. .587
14.2 "American Airlines 'SABRE' Electronic Reservations System" by M. N. Perry, W. R. Plugge,
American Airlines .................................................... _......................................................................................... 593
14.3 "Real-Time Management Control at Hughes Aircraft" by D. R. Pardee, Hughes Aircraft Co ........................... 603
14.4 "The 465L (SACCS) Computer Application" by P. D. Hildebrandt, System Development Corp ..................... 609
THE "HUMAN" SIDE OF ANALOG SYSTEMS
15.1 "The Computer Simulation of a Colonial, Socio-Economic Society" by W. D. Howard, General
Motors Corp .............................................................. _......................................................................................... 613
15.2 "X-15 Analog Flight Simulation-Systems Development and Pilot Training" by Norman Cooper, North
American Aviation, Inc. . ........................................ _......................................................................................... 623
15.3 "Analog-Digital Hybrid Computers in Simulation with Humans and Hardware" by O. F. Thomas,
Simulation and Computer Center, U.S. Naval Ordnance Test Station .......................................................... 639
15.4 "The Automatic Determination of Human and Other System Parameters" by T. F. Potts, G. N. Ornstein,
A. B. Clymer, North American Aviation, Inc ................................................................................................... 645
LIST OF EXHIBITORS........................................................ _......................................................................................... 661

ix

1

1.1

SIMULATION: A SURVEY
Harry H. Harman
System Development Corporation
Santa Monica, California

Introduction
Simulation may be traced back to the beginning of time -- be it the make-believe world of
the child at play,or the adult make-believe world
of the stage. The impetus for modern scientific
simulation came with the development of analog
computers in the 1930's; and progressed even
further when the electronic digital ,computers wee
created.
The very definition of an analog computer contains the notion of simulation, viz., a
device which simulates some mathematical process
and in which the results of this process can be
observed as physical quantities,such as voltages,
currents, or shaft positions.
While there is no
doubt that the analog computer represents one aspect of simulation, the truly new simulation advances came with the digital computer. In the
past two decades, since the development of Mark I
by Howard Aiken and since Eckert and Mauchly designed the ENIAC, tremendous strides have been
made in science and technology ascribable directly to the flourishing new computing discipline.
The revolutionary impact of the electronic
computer on our society may well be equal to
that of atomic energy -- and may actually surpass
it in the long run. A direct consequence of the
computer is the burgeoning activity which collectively goes under the name, "simulation".
The
growing awareness, and popularity of this field
of activity is evidenced by a recent article in
Business Week in which a parallel is drawn
between the group of simulation experts and the
group of painters known as the Futurists.
Just
as the art works might bear no direct resemblance
to the subjects for which they were named, so the
mathematical formulas, flow diagrams, and computer outputs bear no direct resemblance to the
physical world which they simulate.
Moreover,
this symbolic art ·'represents a massive assault
on tradition -- in this case, the traditional art
of managing large organizations!'I6rhis assault -involving scientific systems analysis and simulation techniques -- first occurred on military systems problems, but more recently has found its
way in~o business and industrial systems problems
as well.

subject is the raison d'etre of this session. In
my review of the work in this area, I came across
John Harling's paper, "Simulation Techniques in
Operations Research -- A Review".5 From the title
it would appear that my work had been done for me.
His opening remarks draw attention to the fact
that "simulationtt is a somewhat ill-defined subject and that considerable confusion exists in
the terminology employed, and he goes on to say:
"The term 'Monte Carlo' is presently
somewhat
fashionable; the term 'simulation' is to be preferred, because it does not suggest that the
technique is limited to what is familiar to statisticians as a sampling experiment." (p.30?) He
equates "simulation" with "Monte Carlo methods"
and thereby implies a much more restrictive usage
of simulation than is intended in thepresent Survey.
The term "simulation" has recently become
very popular, and probably somewhat overworke~
There are many and sundry definitions of simulation, and a review and study of some of these
should help us gain a better perspective of the
broad spectrum of simulation.
Webster only provides the fundamental notion that simulation is
an act of "assuming the appearance of, without
the reality'.
Thomas and Deemer 20 suggest the
following paraphrase of Webster: "to simulate is
to attain the essence of, without the reality."
Note that the substitution of "essence"
for
"appearance" makes the vital difference between
the scientific and the casual use of simulation
It not only is not necessary that the simulator
not "appear" as its real-life counterpart,
but
frequently attempts to imitate reality
closely
may be detrimental to the purposes of the simulation. For example, to expedite the training of
pilots a relatively accurate duplication of the
cockpit is n-ecessary for the trainer, but to duplicate the bulky whole of the airplane would defeat the purpose of the simulator. Thomas and
Deemer advise that ·'we should deplore the tendency to introduce trappings and ornaments in ~
lation to gain the 'appearance' of reality when
~the 'essence' which we need." (p.5)
In a technical dictionary? the term "simulator" is defined as follows:

Definitions
To appraise the current work in simulation
and to apprise you of the general status of this

A physical system which is analogous
to a model under study (as, for instance,
an electric network in which the elements

2
1.1
are in correspondence with those of an economic model). The variables of interest in
the model appear as physical variables (such
as voltages and currents) and may be studied
by an examination of the physical variables
in the simulator. (p.267)
This definition covers what we
normally would
consider simulation when accomplished by analog
or digital computers. Nonetheless, it is not the
universally accepted definition, alternatives being proposed by practically each separate field
of application.
Thus, in the area of Opemtions Research,Harling5 states:"By simulation is meant the technique
of setting up a stochastic model of a real situation and then performing sampling experiments upn
the model.The feature which distinguishes simulation from a mere sampling experiment in the classical sense is that of the stochastic mode!'!
(p.307) As noted above, this definition of simulation is equivalent to the Monte Carlo technique; and is, in fact,
almost identical with
the definition of the latter provided by A.S.
Householder: 6
The Monte Carlo method may briefly be
described as the device of studying an ar~
ificial stochastic model of a physical or
mathematical process •••• The novelty ;- of
the Monte Carlo method~ lies rather in the
suggestion that where an equation arising
in a nonprobabilistic context demands a numerical solution not easily obtainable
by
standard numerical methods, there may exist
a stochastic process with distributions
or
parameters which satisfy the equation, and
it may actualry be more efficient to construct such a process and compute the statistics than to attempt to use those standard methods (p.v).
While this represents a very powerful and useful
technique in simulation,
Monte Carlo does not
encompass all the legitimate scientific aspects
of simulation.
In their book, System Engineering,4 . Goode
and Machol give a ha~zen or more examplesof
simulation in which the Monte Carlo Method is
used in queueing problems. They do not,however,
take the foregoing definition.
Instead,
they
define simulation to be "the study of a system
by the ~ut-and-try examination of its mathematical representation by means of a
large-scale
computer". (p.403) While some people (not in this
audience) might object to the qualifier that
a
"large-scale comnuter" be the means of the study,
they would certainly grant its modus operandi.
This is an operational definition~d as such
it proposes more or less exact procedures to be
followed in executing a program of simulation.
Specifically, Goode and Machol propose a series
of steps (pp.404-7) including the choice of computer
(analog or digital,
in particular);con-

struction of the computational flow diagram (it
being assumed that the mathematical model of the
system has been formulated);
determination of
preliminary (analytical) solutions;
choice of
cases to be treated, with a view toward reducing
the number of runs; data reduction and analysis
(some to be done run by run); and consideration
of the simulation of human beings (by same simple
anal~tical
function or by actual inclusion in
the simu~ation).
A type of working definition is proposed in
the field of Management Science.
Here, simulation is conceived as "the science of employing
computational models as description for the purposes of (1) learning, (2) experimenting, (3)
predicting in management problems."19 A similar
definition, which more specifically delimits the
area of consideration, is the following:l
The systematic abstraction and partial duplication of a phenomenon for the purposes
of effecting 1) the transfer of training
from a synthetic environment to a real environment; 2) the analysis of a specific
phenomenon; or 3) the design of a specific
system in tenns of certain conditions, behavio~, and mechanisms.(p.6)
The behavioral scientist, accumstamed to laboratory experimentation,puts it even more directly:8
"By simulation, we mean a technique of substituting a synthetic environment for a real one -- so
that it is possible to work under laboratory conditions of control."
The foregoing definitions range in emphasis
from a sampling plan (which distorts distributions in order to obtain relatively
efficient
estimates of the parameters)
and the mere use of
a large-scale computer, to a simple delineation
of the area of inquiry.
What they have in common is an attempt to substitute other elements
for some or all of the real elements of a system.
Perhaps the simplest and most direct definition
of simulation is merely the act of representing
some aspects of the real world by numbers or
other symbols that can be easily manipulated in
order to facilltate lts study.
In thlS sense,
simulation is one of the oldest analytical tools.
Classifications
However simulation is defined, there nmains
the problem of selecting the appropriate elements
of a system to be simulated.
Which aspects are
represented, and how they are represented,
constitute the distinguishing characteristics of the
different types of simulation. Hopefully,
these
considerations should also provide for the meaningful classification of simulation types.
After an exhaustive search of the literature, and several months' cogitation, the writer
was reluctantly forced to conclude that there is
no completely adequate taxonomy of
simulation

3
1.1

types. Perhaps some day a reasonable basis will
evolve for classifying simulation types into major and subordinate categories, and the practitioner will be assisted thereby; but at the present time, we can do very little in that direction.
About the best that has been proposed (see
for example, I. J. Good3)
is a single continuum
on which the model is classified according to its
degree of abstraction from the real-life system,
operation, or procedure. Thus, the focus is on
the simulation model and its relationship to its
real-life counterpart. This conceptual basis for
ordering simulation types follows:
(1) In the most extreme instance (ultimate
or trivial, depending on your point of view), the
real system can be used as the "model" to gain
knowledge about itself. However direct and simple
it might sound, it is usually neither practical
nor feasible to detennine the inherent properties
of a system by observing its operations. Limited
time and resources often force the use of morler,
less expensive methods than the ·'identity simulation".
(2) Only one step removed from the reallife instance is the attempt to replicate it with
the highest degree of fidelity,
by means of an
operational model of the system in its normal environment.
A SAC mission flown to test the air
defenses of the United States is an example of
an essential replication of a war situation. Enemy bombers are replaced by SAC bombers; ADC fires
no weapons. Such "replication simulation" really
involves very little abstraction from reality,and
also provides very little gain;
except tv make
possible the limited study of selected dangerous
or future situations. A subcategory of this classification might involve essential replication of
operational gear while employing abstracted inputs.A case in point is the Air Defense Command's
System Training Program (discussed below).
(3) Next, along our continuum, the replkation might be attempted in the laboratory instead
of in the field. Here it is necessary to choose
the relevant features of the real system for representation in the laboratory, and also to decide on the means of such representation. A system
may be made up of such diverse elements as people,
hardware,operating procedures, mathematical functions, and probability distributions. A laboratory model might consist of the actual replication
of some elements and the abstraction and substitution by symbolic representation of others. It
should be noted that every kind of substitution
is possible: people are often simulated by hardware, but the reverse is also done. A wide range
of simulation types is encanpassed by ulaboratory
simulation", and perhaps is best exemplified by
operational gaming.

(4) More clear-cut abstraction from reality
is involved in the complete ttCXl1lPJ,ter simulation"

of a real system. In some circles this is the
only admissable type of simulation.
There is no
roam for human beings or real hardware components
in this model of the system. All aspects of the
system must be reduced to logical decision rules
and operations which can be programmed.
If the
model of the system consists only of mathematical
functions, the simulation is said to be deterministic.
If it also includes probability distributions then it is stochastic.
This type of simulation is quite cammon in operations research ,
with a popular example being a "computer simulation" of a (hypothetical) business finn.
(5) The highest degree of abstraction leads
to the complete "analytical simulation", wherein
the real system is represented completely Qv mEans
of a mathematical model and a solution ( at least
theoretically) can be obtained by ~ical mea n s.
Essentially, the problem here is that of solving
a set of equations. Even if a closed form is not
available, approximate methods (including Monte
Carlo) can be employed to get a solution.
The
least and the highest degrees of abstraction -t1identity simulat ionttand complete nanalytical simulation"- may not be of much experimental value,
but they do provide useful conceptual bounds for
the simulation continuum.

Need for further classification.-- While the
foregoing considerations provide a fundamental
(philosophical) continuum on which
simulation
types might be ordered,
it is not sufficiently
discriminating.
The bulk of the simulation studies reported in the literature would fall into
one or two categories only. Further, more detailed distinctions could lead to generalized principles and thus to the full development of a
discipline of simulation.
The additional dimensions of simulation cannot be adequately determined at the present rudimentary stage of development of this field.
Dichotomous classifications.-- What is frequently done as an alternative is to break the
total field of simulation into two classes. Commonly encountered examples of such dichotomy, or
pOlarity, is detenninistic-stochastic; deductiveinductive; analytical-physical; computerized-manual;
or one of the many variants of these. An important consideration is the absence or presence of
at least one human being in the simulated model.
While this seems to offer a real
distinguishing
characteristic, it does not help nearly as much
as anticipated. There can still be
stochastic
models which are simulated entirely in a computer,
or by means of a computer and people.
For this
reason, the writer discarded an earlier plan in
which the primary dichotomy was into "automatonsimulation" and tlbio-simulation". Differences in
simulations' that are fully computerized and those
that involve human beings may be useful ,but should
probably be subordinated to more fundamental classification concepts.

4
1.1

Even this crude classification scheme may
provide a useful guide in planning a simulation
experiment. As a general rule, increasing experimental control can be attained by moving in the
direction of a complete mathematical model, but
unfortunately this usually is associated with decreasing realism.
The more that is known about
the properties of an element of a system, the better it can be simulated. Imperfectly understood
system elements probably should be used tt as is"
in the model rather than approximated in a probabilist:i.c manner or by decision rules.
Adequate
simulation of a system in the laboratory requires
a detailed systems analysis with particular atte~
tion paid to the functional structure of the various tasks and the operations to be performed by
the human beings in the system.
Since the human
actions are certainly of a stochastic nature,realistic simulation of a man~achine system
can
best be accomnlished by having the human elements
in the model.
Classification by objectiveo-- An alternative breakdown of simulation activities can be
made according to the purpose or objective of the
simulation. The principal categories usually employed are evaluation, training, and demonstration. With the emergence of very large military
command and control systems, the old trial-anderror method had to give way to simulation as the
primary technique for the design and development
of such systems, as well as for the evaluation of
alternative solutions to system problems. Again,
in the implementation and operation of such systems, simulation has been found to be a very effective device for training. Not only have simulators been employed for individual flight instruction in place of expensive and dangerous
procedures, but similar efficiencies have been
realized in training groups in total system operations through simulat ion.
This is one of the
chief objectives of mana~ement games as well as
the specific training programs of military systems.
In the demonstration role, simulation serves as a
means of indoctrination -- to exhibit the feasibility of a complex system.
Simulation as a research tool
While this very brief account of the uses of
simulation for evaluation,demonstration,and training immediately points up its value, some IIDre definite indication of the advantages of simulation
as a research tool in the study of complex ~~s
seems to be in order. First of all, the real system in the field is not as amenable to control as
a simulation of it. At the same time there is no
interruption of the on-going activities in order
to conduct the research.Also, productive research
requires the taking of quantitative measurements,
which again can better be accomplished in a simulation study than by observati~n of the actual
system.
These primary advantages are really the advantages of the laboratory over the field,regard-

less of whether it is a chemistry laboratory or a
digital-computer laboratory. Simulation as a research technique ha~ more specific advantages:
(1) It can compress or expand real time. A
business operation of a year can be simulated in
minutes in order to study long term trends or to
study the operations under varying alternatives.
On the other hand, the process can be slowed down
to permit the more detailed study of critical situations.
(2) It provides the ability to experiment,
test,and evaluate new systems or proposed changes
to existing systems in advance of having to make
firm commitments.Aside from great economy of time,
simulation of this type makes it possible to consider hypothetical systems which may be dangerous
or impossible to try any other way.An interesting
example involves the procedure the Cornell Aeronautical Laboratory employed in designing and CO~
structing the Mark I perceptron for the automatic
identification of simple patterns. They first demonstrated by simulation on a computer (IBM 704)
that such an experimental machine could be built.
(3) It makes for more economical experimentation, both in time and money. A complete t1 computer simulationU of a system usually can be run in
very short time once the program has been developed. However, the cost of creating a large-scale
computer simulation program can be prohibitive.
Usually it is justified because of continued experimentation with the model, but on occasion the
payoff may be so great as to ,justify even a single
trial.
(4)It permits the replication of experiments
under different conditions. An important example
is the replication of economic time-series, which
just could not be accomnlished without simulatio~
Review of simulation activities
Extent of literature.-The acceptance of
simulatlon eVldently has been widespread -- as
witness the increasing n'~ber of simulation studies in the last decade. Prior to 1951 there was
nothing in the scientific literature on this subject. The most recently published bibliogranhyl5
contains 344 entries (including 6 other bibliographies)and except for one reference (IIA Simplified j'lar Game, 1897) the earliest article is· dated 1951. Two other bibliographies merit special
mention.
Malcolmll
presf.:nts what he terms "a
fair sampling of simulation literature to date'~
Concerned primarily with the application of simulation to manage~ent problems, he subdivides the
165 titles into industrial and military applications am separates simulation games from the rest.
The other,12 while not snecific:llly Addressed to
simulation,presents 477 references'to the closely
allied subject of systems research.
One of the
interesting aspects of the latter bibliography is
that it also contains a topical outline of the
field and each reference is assigned to one or

5
1.1

more of the classification categories. The extent
of the literature on simulation has grown to such
immense proportions, in so short a time, that the
truly scholarly exploration of this field loans as
a formidable effort for all but the most serious
student.
No attempt will be made here to review the
content of different simulation studies. The objective is only to indicate the
scope of such
studies.One such collection of 17 studies appears
in the "Report of Systems Simulation Symposium';
published in 1958. These include typical inventory-control, scheduling, cargo handljng, and waiting-line problems on the industrial side; related
problems on logistics systans peculiar to themilitary,as well as military"laboratory simulations",
incorporating systems of men and equipment;
and
even same methodological considerations directed
at increasing the speed of simulation and statistical problems associated with Monte Carlo sampling.
As regards the technical aspects of simulation, the results of current research activities
appear, principally, in the Operations Research
journal, 'specialized statisticat journals, and
publications of various research institutes. Of
special interest is the report of the first Symposium on the Monte Carlo Method6 and two subsequent symposia 17,18 on the same subject.
Operational gaming.-- The simulation studies that have attracted the most attention in
recent years may be described by the generic term
"games" -- intended to cover such activities as
war gaming, business management games, and operational gaming in general.
In their excellent
article, Thomas and Deemer 20
first distinguish
the basic concepts of simulation, Monte Carlo,and
operational gaming; present a brief reviewof some
of the theory of games of strategy; and then compare the approaches of gaming andncn-gaming techniques to competitive situations.
The role of
operational gaming is best expressed in their
words:
Although simulation and Monte
Carlo
methods are often used in gaming we feel that
the essence of operational gaming lies rather
in its emphasis on the p~ of a game.
There is playing to formulate a game, playing to solve a game,
and playing to impart
present knowledge of a game. Thus we define
operational ~aming as the serious use of
playing as a primary device to formulate a
game, to solve a ~ame,or to impart something
of the solution of a game. (p.6)
In practical applications, the technique of
gaming is aimed principally at providing practice
in working through alternative sequences in considerable detail. Within the framework of a particular game certain input parameters canbe altered to provide innumerable variations. When human
teams participate in such games, they not only

gain practice in comprehending the consequences
of particular moves and sequences of events, but
also gain some insight into the perspective of
the participants.
The development and present usage of management games is reviewed by Joel Kibbee in the following paper on this Program.
He stresses the
importance of computers in this area, an~ discusses the building of models and programming of management games.
It should be remembered that
non-computer or manual business games (e.g., as
developed by Stanley Vance at the University of
Oregon and by John L. Kennedy at Princeton University) have considerable merit as tools for
management training and_development as well.
Management control.-- Perhaps aneof the most
powerful tools for management control of largescale programs is the activity known as PERT
(Program Evaluation Review Technique). This ~tem
of charting the key milestones into a network for
the accomplishment of an objective, dependent on
many and diverse factors, was first developed in
conjunction with the Polaris program. lO As a result of such management control, the Polaris program became operational two years earlier than
originally anticipated.
A similar technique developed for the Air Force by Douglas Aircraft
Company in conjunction with the Skybolt program
is PEP (Program Evaluation Procedure). The PERT/
PEP program evaluation techniques now are being
extended to almost all Army, Navy, and Air Force
weapons systems. 9 Among other computer-based methods for monitoring schedules being developed
is SCANS (Scheduling and Control by Automated
Network Systems) at System Development Corporation. The aspect of these techniques which is
especially germane to this Session is the optimization of networks through simulation. By devising a "computer simulation" of the scheduling
technique, alternative management decisions can
be tried, and from the output an optimal solution can be determined. Closely related to these
types of programs is the Decision Gaming work on
which Dr. Vazsonyi reports later in this Session.
Social behavior.-- Turning to another area,
of studies on simulation of social processes being carried out in universities and research
laboratories from coast to coast. His survey is concerned with research in the behavioral sciences which
use computers in the sL~ulation of social behavior. The studies range from experiments in interactions and conformity of small groups to in=
tergroup relations in the community to the behavior of an entire society and international relations.
~llis Scott14 , calls attention to dozens

Vehicular traffic.--8till another area which
is receiving more and more attention is that of
vehicular traffic control.
While the' earliest
works, by H. H. Goode, G. F. Newell, and others,
only date back about six years, the activity has
been gaining considerable momentum since
then.

6
1.1
Research is going on in all parts of the count~
The extent of the national interest is evidenced
by the conference on transportation research convened by the National Academy of Sciences last
fall. About 150 participants from government, industr,r, universities, and research institutions
met to review and formulate a program of research
on transportation in the United States. A more
recent conference13 was devoted exclusively to
the utilization of simulation as a research tool
in the areas of highway and vehicle improvement,
traffic control and enforcement,
and driver and
safety education.
An example of a physical model for studying
driver performance, car construction, and road
des.ign is the "driving simulator" at the UCLA
Institute of Transportation and Traffic Engineering. The cab of this simulator consists of a
standard station wagon on a treadmill of steel
rollers, which faces a lO-ft high semicircular
screen and with a small screen on the car's rear
window. Movie projectors throw traffic scenes
on both screens and a battery of instruments record changes in steeringwheel movement, acceleration, braking, and in the driver's breathing rate
and in emotional stress.
Although the ultimate goal is to ~der the
total system, including the driver and the traffic, at this stage of development of methodology, it seems wise to distinguish "driving simulationt1 from "traffic simulation". Early work on
traffic simulation was restricted to one or two
lanes of very short stretches of highway, and required inordinate amounts of comouter time.Nonetheless, such work pointed to the feasibility of
running simulation studies of traffic flow. A
much more extensive model of expressway traffic
flow has been developed at the Midwest Research
Institute, and is reported by Glickstein and Levy
later in this Session.
Simulation in

man~achine

laboratory research

The foregoing review points to many exciting and challenging activities -- emerging as a
result of the development of the digital electronic computer, the use of simulation, and the increased awareness of the "systems approach".Thus,
the study of large, complex man~achine systems
has become possible.
Just as trial-and-error experimentation has
been a respected technique in the development of
the classical sciences,so in the study of complex
systems the new techniques of smulation may be
employed to explore and to define the problem itself. The direction and course of study of a manmachine system should be permitted (at least in
the early stages) to be altered and restructured
during the simulation and according to insights
gained from the simulation itself. This use of
sL~ulation as a new kind of research tool is perhaps the outstanding feature of such laboratories

as RAND's Logistics Systems Laborator,y and SDC's
Systems Simulation Research Laborator.y discussed
below.
NEWS.-- Entire laboratories have been built
to exploit simulation for teaching purposes and
evaluation of systems.
Perhaps the first such
facility to be conceived (in 1945), but which was
not funded until 1950 and then took eight years
to build) is the simulator at the U.S.Naval War
College at Newport, Rhode Island.
This facility
and the exercise conducted in it is called NEWS
(NaVal Electronic Warfare Simulator). At" the
heart of the system is a very large analog computer (known as the Damage Computer) which is designed primarily to assess damage and to provide
feedback to the several forces playing, to indicate their remaining effectiveness. the exercise
is primarily a training device -- used in war
gaming, in the final stages of tactical training
of naval officers from the fleet.
SRL.-- Another laboratory in which simulation was employed as the principal tool was the
Systems Research Laboratory (SRL) of The RAND
Corporation. From 1951 to 1954 this laboratory
employed simulation to generate stimuli for the
study of information processing centers.
The
esse~tial features
of a radar site were created
in the laboratory and by carefully controlling
the synthetic inputs to the system and recording
the behavior of the group it was possible to stuqy the effectiveness of various man~achine combinations and procedures.
STP.-- The research in SRL eventually gave
rise to the Air Defense Cormnand' s System Training Program (STP) -- probably the largest-scale
simulation effort ever attempted.
STP is now in
operation throughout the United States, as well
as in Alaska, Canada, and Europe. Training exercises are conducted in the normal' working environment at the radar sites, direction centers in
the SAGE system, Division Headquarters, and higher
cormnands.
Fundamental to this vast program is
the creation of problem materials by means of an
IBM 709 and special off-line and EAM equinment.
Through these means synchronized radar pictures
for large areas of the country are simulated Wong
with other innuts required by the operating system, e.g., fli~ht plan information, intelligence
and weather information, and commands fram higher
headquarters.
Also, various lists and maps are
prepared for the trainers to assist them in
observing and recording crew actions in order to
furnish feedback on system performance to the
crew immediately after each exercise. Through
simulation of this type it is possible to provide
exercise of air defense procedures and regulations, applicable either in peace or in war situations, at a fraction of what it would cost with
tfreolication simulationtt •
L5L.-- In 1956, the Logistics Systems Laboratory-{LSL) was established at RAND under Air

7
1.1

Force sponsorship. The first study in this labor~
atory involved the simulation of two large logistics systems for purposes of comparing their effectiveness under different governing policies
and resources. The system consisted of men and
machine resources together with policy rules on
the use of such resources in simulated stress situations such as war. The simulated environment
required a certain amount of aircraft in flying
and alert states while the systems' capability to
meet these objectives were limited by malfunctioning parts, procurement and transportation delays,
etc. The human participants represented management personnel while higher echelon policies in
the utilization of resources were simulated in
the computer. The ultimate criteria of the effectiveness of the systems were the number of
aircraft in commission and dollar costs.
While
the purpose of the first study in LSL was to
test the feasibility of introducing new procedu~
into an existing Air Force logistics system and
to compare the modified system with the original
one, the second laboratory problem has quite a
different objective. Its purpose is to ~prove
the design of the operational control system
through the use of simulation. The complete description of this study is presented by Dr. Steger
later in this program.
ASDEC.-- A somewhat different type of facility in which simulation is employed to test and
evaluate electronic systems is the Applied Systems Development ~valuation Center (ASDBC) of the
Naval Electronics Laboratory at San Diego.Recently the Navy Tactical Data System was being evaluated. The operational system was simulated by
means of actual hardware components such as the
Univac M460 computer and cardboard mockups of display and control equipment. The facility includes
an analog-to-digital computer which
generates
synthetic radar data used in the testing of operational systems.
NBS Study.-- Perhaps the largest single step
in the exploitation of simulation for research
purposes was the recent Feasibility Study~nduct­
ed by the National Bureau of Standards.
The
broad objectives of this study are best indicated
in its opening paragraph:
This report presents the results of a
study of the feasibility,
design, and cost
of a large-scale tool to be used in a research program on man-machine systems. This
tool facilitates the simulation of complex
weapon, systems for purposes of laboratory
experimentation with human subjects in the
system feedback loops. It is intended to aid
in the optimization of system performance
through studies of man-machine dynamics. It
incorporates capabilities which represent a
substantial advance over those of existing
facilities for research on man~achine systems.
Feasibility was demonstrated through the actual
design, implementation and operation of a scale

model of the desired facility. The work done at
the National Bureau of Standards provided
the
fundamental guidelines and philosophy for the more
ambitious laboratory facility being built by the
System Development Corporation in Santa Monica.
SSRL.-- Recognizing the importance of recent
work in simulation, as well as recognizing the
need for continued and expanded support for the
further development of this area, with particular
emphasis on its use in the study of complex manmachine systems, SDC decided to c reate a generalpurpose, computer-based, facility in which such
research could be conducted. Plans for the Systems Simulation Research Laboratory (SSRL) were
initiated about fifteen months ago and are about
to come to fruition.
My report on SSRL is in
greater. detail because of my involvement and familiarity with it.
The physical facility, covering about 20,000
square feet, has just been completed. The main
experimental operations space is a room approximately 45 x 50 feet with 20-foot clearance from
floor to ceiling. It is completely surrounded by
an elevated observation area. This large room may
be divided into appropriate smaller areas bymearn
of movable walls. Adjacent to the large, highceiling space are smaller, standard height experimental areas, which a Iso may be adjusted in size
-and shape to acco~modate the operations and observation requirements of specific projects.
A basic concept in planning a laboratory of
this kind is the distinction between universaltype and project-specific type equipment. Of the
former type,the most important E the general-purpose digital comnuter. A Philco 2000 system was
selected and is now installed. Another major
piece of equipment is a transducer that permits
human beings and other real-time elements of a
system to com~unicate with the computer. Such a
real-time switch and storage unit (RL-IOl) has
been designed and built at SDC and will be ready
for integration with the computer next month. An
internal telephone system (up to 120 stations), a
public address system, recording facilities for
any audio line, and a closed-circuit
television
system round out the general-purpose equipment of
the laboratory at this time. The specific hardware requirements for the first couple of projects
are now being determined.
Another basic concept is a general-rur~ose
programming system. Perhaps some day we will have
will
a general-purpose simulation program which
greatly facilitate the execution of research projects. For the present, however, we refer to the
basic utility program system for the Philco 2000
operating with the HL-I01. At SDC we are using a
problem oriented language, known as JOVIAL, 'ofhich
is patterned after Algol (the International Algebraic Language). The principal effort involves
the preparation of a JOVIAL Translator for the
Philco 2000; but which has been written in such
a manner that preliminary testing and actual compilation could be done on an IBM 709.
Also, an
executive control program hasbeen developed which

8
1.1

takes cognizance of the requirements introduced
by the RL-IOI and of the unusual nature of ~he
applications of the Philco 2000. The programmlng
for the initial research projects is proceeding
concurrently with the utility programming.
The new laborato~ is expected to enhance tie
present research efforts of SDC and to open entirely new avenues of research endeavor. In the
former category are a number of research projects
that have necessarily been limited in scope, but
which can now be broadened because of the new facilities.
One such area is that of automated
teaching. Successful research in this area has
been conducted at SDC in the last two years, but
the constraint of a single student to the teaching machine has been a severe limitation. ~his
made the gathering of statistical data very tlffieconsuming.
Also, any potential application of
automated teaching techniques in the academic or
the military or industrial organizations would
certainly require more efficient means than individual tutoring.
Thus, the next stage in this
research effort is to create a Computerized Laboratory School System (CLASS) which project will
be studied in SSRL very soon.
Another example of present research at SDC
which can be expanded through the medium of the
new laboratory is the study of Management Control
Systems. At the present time, the research consists of a "computer simulation" of the behavior
of a business system.
This model enables the
study of the reaction of the organization
to
specific changes under alternative sets of decision rules. As interesting as the computer sImulation might be, it will be found lacking in a
basic ingredient insofar as acceptance by realworld managers is concerned. That ingredient is
the true human variability in decision making. The
particular model certainly can be made more valid
-- albeit, more complex and less controllable -by introducing human decision makers at certain
critical points in place of decision rules. Such
a tllaboratory simulationll model, at a later phase
of the research, will be possible in the SSHL.
The first new research endeavor to exploit
the SSRL facilities is a study of a terminal air
traffic control system operating in a post-19 70
air environment. Projected increases in traffic
volume and aircraft speeds indicate that terminal
control zones will increase in size and will therefore inclUde many airports wi thin a single complex.
Coordination among many airports of the control
of high density traffic of widely differing performance characteristics poses significant problems of organization and planning. It is believed that in order to effect the safe, orderly and
expeditious flow of air traffic in a termina complex, there will be a need for a new planning agency in addition to the control agencies in intimate
contact with the details of the environment. The
general purpose of this Droject is to investigate
the functional interactions among the cmtrol agencies, and to evolve alternative hypotheses regardin~ superordinate planning agencies.

In the first phase of the project the configuration simulated is an air traffic control system for a two-airport terminal complex. The system consists of the operators and equipment representing the following agencies for each of the
two airports:Stack Control,Aoproach Gate Control,
ADProach Control, Departure Control, and Flight
Data Processing.
Some of these agencies include
human operators while others are represented by
completely automatic processes.
The objectives
of Phase I are to study inter-airport coordkation
problems and to identify significant variables fur
future systematic investigation. Additional p1anning and coordination f-,mctions will b~ a~ded
in subsequent configurations as they are lndlcated by Phase I results. This project -- involving
a "laboratory simulation" model --is an excellent
example of the utilization of the best aspects of
the broad range of simulation techniques in order
to experiment with a complex man-machine system.

In this wide range of simulation work which
we have reviewed two distinct
activities stand
out, neither one taking much cognizance of the
other. On the one hand, simulation work is being
done in the Operations Research field which may
be classified - largely as "computer simulationu •
On the other hand, there is the group of behavioral scientists, experimental psychologists in
particular, engaged in the simulation of environmental conditions which may be called "laboratory
simulation'l. Each of these groups could learn a
great deal from the other. Furthermore, there is
increasing evidence that ft pure" simulation will
have to be modified if it is to stand the test of
validation. What is necessary is the marriage of
the two approaches -- a realistic possibility in
the new man~achine system laboratory.
References
1.

Bogdanoff, E., et ale Simulation: An Introduction to a new Technology. TM-499.
Santa
Monica,Calif.: System Development Corp., 196~

2.

Ernst, Arthur A. Feasibility Study for a ManMachine Systems Research Facility. WADC Technical Report 59-51. AST1A Doc. No. AD 213589.
Good, 1. J. Discussion in "Symposium on Monte
Carlo :Methods," J.Royal Stat.Soc.,B 16(1954),
68-69.

4. Goode, Harry H., and Machol, Robert E. System
Engineering. New York, N.Y.: McGraw-Hill Book
Co., Inc., 1957. Pp. xii + 551.

5. Harling, John."Simulation Techniques in Operations Research - A Review,1t
search, 6 (1958), 307 -19.

Operations Re-

9
1.1

6.

Householder, A.H.,Forsythe,G.E., and Germond,
H.H. (eds.). Monte Carlo Method. Natl. Bur.
of Standards Applied Math.Series,No.12, 1951.

7.

Kendall, Maurice G., and Buckland, ~'lilliam R.
A Dictionary of Statistical Terms. New York,
N.Y.: Hafner Publishing Co.,l957.Pp.ix + 493.

8.

9.

Kennedy, J.L., Durking, J.E., and Kling, F.R.
"Growing Synthetic Organisms in Synthetic Environments. 1t
Unpublished Paper, Princeton
University, 1960.
Klass, Philip J. !TPERT/PEP Management Tool
Use Grows,"
Aviation {leek, Vol. 73,
No. 22
(Nov. 28, 1960), 85-91.

10. Malcolm, D.G., Rosebloom, J.H., Clark, C.E.,
and Fazar, W. "Application of a Technique for
Research and Development Program Evaluation,1l
Operations Research, 7 (1959), 646-69.
11. Malcolm, D. G. "Bibliography on the Use of
Simulation in Management Analysis, "Operations
Research, 8 (1960), 169-77.
12. McGrath, Joseph E.,
and Nordlie, Peter G.
Synthesis and Comparison of System Research
Hethods, ApD.B (1960). ASTIA No. AD 234463.
13. National Conference on Driving Simulation.
Sponsored by Automotive Safety roundation,
Public Health Service. Santa Monica, Calif.,
Feb. 27 to March 1, 1961.

14. Scott, Ellis. Simulation of Social Processes:
A Preliminary Report of a Survey of Current
Research in the Behavioral Sciences. TM-435.
Santa Nonica ,Cali f.': System lJevelopment Corp.,
1959. Pp. 15.
15. Shubik,Martin. "Bibliography on
Simulation,
Ga~ing,
Artificial Intelligence and Allied
Topics,11 J. Am. Stat.Assoc.,S5 (1960),736-51.
16. Silk, L. J. f'The Gentle Art of Simulation, II
Business Week, Nov. 29, 1958, 73-82.
17. "Symposium on Monte Carlo Hethods,1'
Stat. Soc., B 16 (1954), 23-75.

J. Roy.

18. Symposium on Monte Carlo Methods (Herbert A.
Meyer, ed.). New York, N.Y.: vlileyand Sons,
1956. Pp. 382.
19. The Institute of Management Science Bulletin,
Vol. 5, No. 2 (1958), 'p.3.
20. Thomas Clayton J., and Deemer, Walter L., Jr.
t'The Role of Operational Gaming 'in Operations
Research," Operations nesearch,5 (1957),1-27.

11

1.2

MANAGEMENT GAMES AND COMPUTERS
By
Joel M. Kibbee
Director of Education
Remington Rand Univac
Division of Sperry Rand Corporation
Management Games
Management games, although a relatively new
educational technique, are being widely utilized,
and much discussed. They are primarily of concern
to the educator and to the research scientist, but
since many of these games are played with the aid
of an electronic computer, they should be of interest to computer people in general. In addition to the use of a computer for existing games,
new games are being developed and will require programming. Many papers have been published on the
educational aspects of management games; this
paper has been written primarily to arouse interest in them as a computer application.
The first management game to become widely
known was one developed by the American Management Association in 1956. It continues to be
used, along with four or five other games since
developed by A.M.A., as part of their management
education courses and seminars. Over one hundred
different management games now in use are listed
in a forthcoming book on the subject*, and more
than 30,000 executives have participated in at
least one of them.
It seems worth-while to give a brief description of a typical game play for those who have
never participated. The Remington Rand Univac
Marketing Management Simulation will be used as
an example.
The game session begins with a briefing. At
this time the instructor describes to the participants the type of company they are about to
manage, the economic environment, the general
nature of the product, and the competitive forces
they will face. He also discusses the scope of
their authority, the functions to be filled, the
decisions to be made, and the information they
will receive.
The participants are divided into management
teams, and after the briefing, the various teams
meet to develop an organization, set objectives,
and decide upon policies and procedures. In a
typical game; involving perhaps forty to fifty
_*Management Games, by Joel M. Kibbee, Clifford J.
Craft, and Burt Nanus, soon to be published by
the Reinhold Publishing Company.

executives, there might be six teams each with
seven or eight members.
Games are played in "periods," with a period,
depending on the particular game, being a simulated day, week, month, quarter or year. The
Univac Marketing Game takes place in months. The
participants are given operating statements for
December and begin by making decisions for January. In addition to the operating statements
they are also provided with a case history, sales
forecast, and data on material and operating
costs, production facilities, and shipping times.
The decisions for January are processed by
the computer and operating reports are produced,
and are returned to the participants. Decisions
are now made for February, and so forth, perhaps
for one simulated year. In a typical play the
companies have a half-hour in which to make decisions, and reports are returned about ten minutes
after the decisions have been submitted.
In the Univac Marketing Game, each company
manufactures one product and markets it in three
different regions, East, West, and South. All
companies are competing in the same c'onsumer
market. The managers set price, spend money on
advertising, hire Or fire salesmen, set the
salesmen~ compensation rate, set production lev~l,
engage in special market research projects, etc.
The total market is shared among the companies
according to their pricing and advertising policies, according to the number of salesmen and
their degree of training, and so forth. The operating reports show the sales obtained, the net
profit achieved, inventory on hand, etc. The report also s~ows the number of salesmen on hand;
companies can pirate experienced salesmen away
from one another.
Each management attempts to achieve the
largest possible accumulated net profit, and a
"winner" might be proclaimed. This is usually
discouraged, however, as good performance in a
management game as in real business, depends on
many factors such as return on investment, share
of market, personnel policy, and numerous others
that contribute to success. At the end of the
game play, a discussion session takes place.
This "critique" is held to focus attention on the

12

1.2
lessons which were to be taught, and it gives the
participants the opportunity to review their performance, discuss management principles with
other members of the group, and receive feedback
from the game administrator and other observers.
Existing management games vary widely in the
types of models used. The original A.M.A. Game,
and similar games developed by IBM, UCLA, Pillsbury and other organizations, are concerned with
general management. The Carnegie Tech Management
Game is based on the detergent industry. Other
games exist for banking, petroleum, telephone exchanges, insurance companies and super markets.
Some games concentrate on a particular management
function, such as marketing, materials management
or manufacturing. There is a game concerned with
the management of a gas station, and three different existing games are based on an auto dealer
model. The military have, of course, been playing
war games for many centuries, and they are now
utilizing computers for this purpose.
Most games are played by one or more management teams, where a team might be made up of anywhere from one to twenty executives. Since most
games contain some obvious measure of performance
there is always a "rivalry" between teams. There
mayor may not be a direct "interaction." In a
marketing game various teams may be in competition for a common market, and the action of one
team, say price of its product, will affect all
other teams. In an inventory control game, on
the other hand, each team is attempting to achieve
the best performance beginning from the same conditions. A game with interaction is like tennis,
a game without interaction is like golf. Both
such games engender rivalry.

computer. This has given rise to an obvious
classification into manual games and computer
games. Computers are used for management games
for the same reasons they are used in any business application, primarily to perform computations speedily, accurately, and automatically.
Computer games can also be more flexible, as will
be discussed more fully below.
Some rather odd advantages have been attributed to manual games. It has been stated, for
example, that manual games can be made very
simple and easy to play. Obviously a computer
game can be made just as simple, though there is
a temptation to use the full capacity of the
computer with a resultant complexity that is not
needed to satisfy the educational objectives.
Similarly, one published statement claims that an
advantage of manual games is less time pressure
on the participants. Just because the computations are made rapidly does not mean that the
participants must make their decisions quickly.
Manual games do have advantages. They usually
cost less initially, do not require special
facilities, and can be scheduled as desired.
Good design is required to keep down the number
of clerks and administrators needed. There are
so many advantages to computer games, however,
that, at least for this author, they see~ well
worth the development and operating costs.

Manual and Computer Games

Since games vary widely in complexity, it is
not possible to state how much computer time or
programming is generally involved. However,
several moderately complex games which the author
helped develop, and which were designed for a one
or two day management exercise, involve something
of the order of 3,000 to 5,000 instructions, and
took from three to nine man-months to program.
The total development cost of a game includes
more than programming, however. More or less
time can be spent on the creation of the model,
consulting fees can be required, as well as the
cost of materials and test plays. Most often
the programmer is involved in the development of
the model as well as the computer coding. These
moderately complex games are usually designed for
a one day management exercise and might involve
five to eight hours of computer time, although
the computer might actually be used only part of
the time and be free for other processing between
game periods. As an example, a play of the
Univac Marketing Management Simulation, lasting
for perhaps eight to ten simulated months, and
accommodating about fifty executives, might cost
$300 to $500 for computing time. The cost is
about the same if the game is played discontinuously using a Univac Service Center. However, a
group playing a game might also have non-computer
expenses for special facilities, materials or
staff.

Management games, like the business situation
they simulate, require that information be processed and calculations made. The extent of the
computations depend on the particular game, and
may require anything from a pencil to a large

It is not necessary to have a computer on
the premises to conduct a management game. Recently, five different cities in the Midwest
simultaneously played the Univac Marketing Game
through the use of leased telephone lines. Of

The word "competitive" has often been used in
the classification of games, usually as synonymous
with interaction. There is, however, an economic
meaning for "competition," namely, competing for a
share of a common market. A marketing game would
include competition, but such competition can be
either of an interactive form, with the competitors being other participating teams, or a noninteractive form, in which the economic competition is built into the model itself.
Games also differ as to the level of management for which they are intended. They differ
widely as to the complexity of the model. Some
are meant to be played quickly, others require
considerable analysis. In general, then, a
management game is a dynamic case history in which
the participants, faced with a simulated business
situation, make decisions, and are fed back reports based upon these decisions.

13

1.2
more interest, however, is what we call the "disc;ontinuous" mode of play. Decisions are made
perhaps once a week and mailed to a service center
and the reports are mailed back. This enables the
game to be played without infringing upon regular
production time since the game can be run at any
odd 15 minutes at the convenience of the center.
Many educatio~al, industrial and professional
organizations are at this moment engaged in discontinuous plays of management games.
Management Games And Computer Personnel
Because of the widespread use of management
games, and their ever increasing growth, it is
likely that most computer installations will at
one time or another find themselves involved in
running a game session. Several computer manufacturers have developed management games and
happily supply the programs to their users. Most
computer games developed by other business Or educational organizations are also available.
Directors of Training or of Management Development
are now generally interested in this new tool and
will probably be getting in touch with the manager
of the computer installation if they have not already done so.
The computer installation may also be called
upon to help develop and program a new management
game. Irrespective of their interest in education,
computer personnel will find that management games
can provide an excellent orientation to data processing for top management. It is sometimes difficult to get a company president to watch a payroll demonstration, or a matrix being inverted,
but, as personal experience has shown, he can become very much interested in the computer as a result of his involvement in a management game.
The data processing manager might consider
the use of management games for training personnel within his own department. A game called
SMART for systems and p=ocedures managers was developed a few years ago, and the System Development Cocporation developed a game called STEP for
use in training programming supervisors.
The Educational Advantages Of Management Games
Very little research has been done on the
validity of management games as an educational
tool, but a similar statement can probably be
made about most educational techniques now in us~.
The most striking thing about a management game is
the involvement on the part of the participants.
One is continually impressed with the way in
which executives will work "after hours" in planning the operations for their simulated companies,
and it is generally necessary to bring in sandwiches and coffee rather than to attempt to interrupt the play for a luncheon or dinner. Motivation is an important aspect of learning, and
management games are sometimes used at the start
of a course or seminar merely to stimulate
"students" towards a greater receptivity for
lectures or other types of training that will

follow.
Management games are superior to other educational techniques for demonstrating the importance of planned, critically timed decisions; the
necessity of flexible organized effort; and the
significance of reaching a dynamic balance between interacting managerial functions. They can
also demonstrate the need for decision-assisting
tools, such 'as forecasts, control charts and
budgets. They can demonstrate to management the
power of a scientific approach to decision making.
Management games are educational tools, and
to be effective should be used along with other
techniques as part of an overall course or seminar.
The briefing and the critique are as important as
the actual play. Furthermore, most game sessions
involve many "incidents" which are not actually
part of the computer model. The participants
might be asked to formulate a personnel policy, Or
to submit special reports. Job rotation, promotions, appraisals, and so forth can all be made
part of the exercise. Much of the activity that
takes place -- in organizing, planning, com~u­
nicating and controlling -- is in addition to the
numerical decisions which are submitted to the
computer. Management games continue to be demonstrated fOT a variety of reasons, but a demonstration is not a course, and a participant
should not form his opinion as to their educational merit from a few hours engaged in a demonstration play.
Building A Management Game
M3nagement games are constructed for educational purposes, and their construction is premised on a set of educational objectives. Working
from the objectives, a model which simUlates a
business situation is constructed. One of the
most important constraints on the model is a need
for simplicity.
A game must be simple to play. This does
not mean that it needs to be easy to make good
decisions, but the participants should not have
to devote considerable time and energy to learning
the rules. It requires skill and experience to
abstract from the real world those elements of
major importance so that a playable game will result. It is here that a programmer must work
closely with the team that is building a game,
and vice versa, the team should get the programmer
into the act as early as possible. Skillful programming can do much to simplify the mechanics of
play for the particip3nts. A good program facilitates the manner in which the players submit their
decisions, and attempts to set up procedures which
will keep clerical errors to a minimum, and even
possibly have the computer edit the decisions for
obvious errors.
In the Univac Manufacturing Game, for example, there are quite a few shipping decisions.
Suppose a company decides to ship 100 Clanks to
the Western Region, and 100 Clanks to the Eastern

14
1.2
Region, but only has 160 Clanks on hand. The
computer automatically interprets the decision as
a desire to ship equally to both regions and 80
are sent, with no interruption in the game session. Similarly in games which have a limited
cash on hand, the computer can issue emergency
loans at some interest rate to accommodate an
error on the part of the participants in spending too much money.
In some games, it is only necessary for the
participants to circle code numbers on a decision form, rather than write out quantities with
possible errors in the number of trailing zeros.
The object of ga~es is seldom to teach the
participants to be less careless in their clerical tasks, and a good computer program can do
much to eliminate clerical chores entirely so
that the teams can concentrate on their decision
problems.
While in university courses, and often in
military games, the time spent on the game session might be lengthy, in most business applications only a day or so of a course or seminar can
be devoted to the game exercise and simplicity
for the participant is extremely important. This
is one of the main constraints on the model, and
even more on the computer program.
Experience has shown that the mathematical
model used in management games can be extremely
simple. One may begin with a curve, perhaps relating a demand to advertising expenditure,
which is defined by a table of fifteen points.
Later it is found that if only five points are
used, the play from the standpoint of the
participants is identical, and the author has
actually used games in which all relationships
are linear -- though there are many interactin3
ones -- without any apparent difference in the
training experience. In the usually short time
a company has to make its decisions, the mere
fact that demand might depend on six or seven
marketing decisions presents adequate complexity
without the use of complex curves.
From the standpoint of the computer computations, it is not usually important whether a more
or less complex curve is used, though extreme
mathematical complexity could lengthen the computation time even in a large computer. However,
throughout the complete model a surprisingly
amount of simplification can be introduced without violating the educational objectives.
Simplicity is important for the game administrator as well as the participants, since it should
not be necessary to have a large staff to conduct
game sessions.
As one continues to emphasize simplicity,the
question of realism always arises. Some games
are used to teach a specific skill or technique,
perhaps production planning and control. In such
cases a realistic and adequately complex model
might be necessary. But most management games
are used to teach general principles, for instance

the importance of planning and control rather
than the specific relationship between inventory
carrying costs and stockout costs. For such
games verisimilitude and not realism is the most
important attribute.
Verisimilitude is the appearance of reality.
It is as important in management games as in the
theater. As long as the relationship between
price and demand, for example, seems similar to
what goes on in the real world, and sufficiently
engrosses the participant in the exercise, it is
not important that the actual curve used, assuming even that it was known, is identical with
that obtained from a detailed study of real data.
Usually one is attempting to train a manager in,
for example, marketing principles, and not the
way in which a particular product with which he
is concerned wili behave in the real market.
The word "Management Game" has generally
been used with a training implication, and the
term "Simulation" used when the object is to
solve a problem, or actually guide management
decision making. Under this terminology a
simulation model must have vatidity, in the
sense of an ability to predict the future, if it
is to be of 'value; the management game model,
used for training, must more often stand the
test of verisimilitude. In fact, it is possible
that a~ over concern with realism can produce a
game that is too complex, too difficult to play,
and can actually destroy verisimilitude, and the
involvement on the'part of the participants
which is so important.
The programmer working together with a
team that is constructing a management game can
do much in the cause of simplicity and verisimilitude. It is this very ability which makes
computer games, in general superior to manual
games. The reports returned to the participants
can resemble the actual reports which they
obtain in everyday life. Furthermore, there is
essentially no limit to the amount of special information that can be provided to simplify the
role of the participant. Thus, a report can give
total costs, unit costs, and percentages. It can
provide a variety of research type information
and statistics. Little of this is usually possible in a manual game. The American Management
Association's Top Management Decision Making
Simulation permits management to engage in
market research studies, at a suitable cost,
which will provide them with the answer to
questions concerning the number of orders for a
.product that might have been received had a different price been set. From the computer standpoint, it is only necessary to loop through the
same set of computations using the research
price rather than the actual price.
A good computer program can also simplify
the task of the game administrator. In the
American Management Association's General Management Simulation, it is possible during the play
to develop additional products through a p~ogra~

15

1.2
of research and-development. When a product has
been developed the staff sends a pre-printed
letter to the company informing them of this.
However, the computer program has been arranged
so that a special report to the administrator is
prepared each period, and the computer itself, on
this report, informs the administrator when to
send out a particular letter. This same report,
incidentally, informs the administrator about
various aspects of the company's performance so
that he is better able to control the session,
and to provide feedback at the critique.
There are many ways in which the relationship between quantities can be introduced into
the model. The most common approach is to use a
mathemati~al function; this itself may take the
form of a graph, a table, or an algebraic expression. Another approach is the use of judges.
There might be one or more "expert" judges who
make a subjective evaluation of the effect of
particular policies. Such a judge should not be
confused with a clerk, sometimes called a judge,
umpire or referee, who, in a manual game, merely
performs the arithmetic computations according to
pre-arranged formulas. This is the function of
the "computer," whether it be a man or a machine.
The "expert" judge is a person with specific
knowledge and experience, who arrives at a
subjective evaluation of the decisions, hopefully
without bias towards the particular participants.
The judges themselves may be a group of
participants, and can be considered to be one
of the playing teams. One might have five
competing companies in a general management game,
together with two teams of , competing bankers who
may invest money in the various companies. A
model may use various combinations of equations,
judges, etc., and in the example just cited only
the capitalization aspect is relegated to judges.
In some games only certain non-quantifiable
factors may be left to observers who can directly
influence company performance in the role of
judges.

on the basis of lowest price. Similarly, raw
material could be offered to the highest bidder,
and the cost would then not be based on a
specific value built into the model. Bidding
techniques have been highly successful in many
games. A particularly good example is the
"Management Business Game" produced by the
Avalon Hill Company, an excellent "parlor" game
that can have serious uses.
Parameters
The mathematical meaning of parameter
applies mainly to the constants used in the
algebraic relationships between quantities, but
for games it is convenient to extend this concept to all constants. For instance, the cost
of carrying inventory in a game might be a
function of the opening value of the inventory
A and the closing value of inventory B, namely,
.05 (rA + sB). The .05 is also a parameter,
but it has been given a specific value for this
illustration. If we set r = 1 and s = 0, we
have a cost dependent only on opening values.
We might prefer to set r = and s =
which is
slightly more realistic. Thus, we are changing
the nature of the model by our choice for rand
s. Similarly, the cost of raw material is a
parameter which may be changed for different
game plays. The inclusion or omission of certain
~actors can be controlled by parameters acting
as switches. Whether or not certain information is to be included on the report can be controlled by a parameter which can take on the
values of 0 or 1. In well-designed computer
games, extensive tables of such parameters are
used, and in this way seemingly different games
can arise from the same model by the choice of
a particular parameter set.

t

t ,

One normally attempts to compute in advance
the parameter values that will be used in a
particular game, but this is usually fol~owed by
actual parameter studies once the program has
been checked out, and, of greater importance,
by numerous test plays. There is always a range
of feasible parameter values, and the particular
set to be used in a particular game session will
depend on the game administrator. A game should
be packaged with many such sets (and they may be
labeled "highly stable som.ewhat price insensitive
industrial product" or "highly competitive very
price sensitive consumer product" and so forth.)
Thus a single game program can be used for different games, and ideally one can imagine one
large model with sufficient parameters to allow
it to be adopted to a variety of industries and
si tuations.

In one game based on a real product, the
judges consisted of several members of top
management, and their own deliberations and
arguments as to the evaluation to be placed on
the various marketing policies being exhibited by
the teams playing the game proved to be, in their
own opinion, an extremely valuable educational
experience. At the other extreme, one might use
judges who did not even know that a game was
being played. For example, marketing policies
together with advertising copy could be presented
to an external group of simulated "customers,"
for their own preference in the products being
offered.

Extreme Values And Limits

A very objective and unbiased manner for incorporating the necessary relationships in a
model is by the use of bidding techniques. For
example, each company might submit a closed bid
as to how many units they could supply at a
particular price, and the demand would be awarded

Like any good program the ga~e program must
be prepared to handle any decisions, no matter
how ridiculous or extreme. Many a program has
"hung up" during a game session because the
game designers were convinced that only a
particular range of decisions might rationally be

16
1.2
made. Not only must the program handle extreme
decisions, but a rational result must be obtained.
For example, a product might be priced somewhere
around $10.00, but the program should be ready to
accept prices from $0 to $999999999 or whatever
the upper limit is that is imposed by the input
design. At some suitably high price the demand
will go to zero, but it must remain there, and
a curve must not be used that departs from zero
again above $999 just because one does not expect
such a decision to be made. It must also be decided what the demand will be if a price of $0 is
set, even though the company that makes it will
be losing money on every item sold.
The American Management Association's
General Management Simulation is immune to participants who might take extreme decisions. It is
possible for a company to fire all of its workers,
close down plants, etc. The program not only
will handle this, but will change its overhead
cost distribution procedure accordingly! As in
a~y computer program, the programmer must be prepared to have any quantity, unless limited by input format (though that is just another method
under his control) take on any value from minus
infinity to plus infinity.
The Future Of Management Games
Management Games are only a few years old,
but one ca~ already look back with fondness on
their infancy, and look ahead with confidence to
their maturity. Like most babies, the first few
games were very much alike; they usually modeled
the marketing or manufacturing of a durable good
and were slanted at higher management. Today
new games, like new teen-age singers, are
arriving on the scene with increasing rapidity.
Management games have been used primarily as
an educational tool. Their use in training is increasing and will increase, and will also spread
well beyond the area qualified by the word
"management." In addition, they will undoubtedly
have considerable application in research,
problem solving, personnel testing, and as a
direct aid to management decision making.
There are still no management games for
mining companies, the fishing industry, or the
mink farm. The entertainment field -- TV,
publishing, motion picture studios -- need
management as well as "talent." Government
city, state or national -- provides a large area
of application of management games for training.
Labor unions, universities and professional
associations also have managers. It is fairly
easy to write down hundreds of training situations which could well use this new educational
tool. And perhaps somebody ought to build a
game to teach people how to build a game.
Games completely different from those nO'N in
use can be expected. A super-game could be constructed which included manufacturing companies,
financial institutions, service organizations,

suppliers of raw material, and even a couple of
management consulting firms. Management games
are having extensive use in management education,
but there is probably an even greater need for
new tools in supervisory training.
Psychologists and sociologists have long
used humans, as well as animals, to study human
behavior. Much work has been done with small
task performing groups. The computer opens the
possibility of new uses of simulation in the
life
sciences, and one can expect an increase
in the number of laboratories now doing such research. Management games can also be expected to
play an important role in economic research.
While the simple manual management game has
a purpose, and is extremely useful in many
training situations, one can safely predict an
increasing use of computers in the management
game area. This paper was presented because of
increasing importance of management games to
computer people. It is hoped that the interested
reader will read elsewhere for those more
important aspects of management games related to
their construction, their educational utilization,
etc.
And it is also hoped that he will find the
opportunity of playing one -- it is not only fun,
it is educational.

17
1.3

AN ON -LINE MANAGEMENT SYSTEM
USING ENGLISH LANGUAGE
Andrew Vazsonyi
Ramo - Wooldridge
Canoga Park, Calif.

Summary
The demonstration model of an on-line management system presented in this paper aims to
provide increased mapagement capability to executives charged with planning and control of
large scale research development and production
programs. The technique is formulated as an
exercise in Decision Gaming and special emphasis is laid to the problem of providing capability
for quick and optimum reprogramming of dollars,
manpower, facilities and other resources. The
task of planning and control is structured into
two components, the more routine tasks are assigned to the equipment, whereas problems
requiring executive judgment are delegated to
players of the Decision Game. Through the use
of mathematical models and computer routines
the consequences of proposed reprogramming
actions are presented to the players in terms of
financial and manpower requirements, facilities
loading. etc. Through a step by step manmachine process, optimum programs and the
best utilization of resources is reached. Management data is retrieved and manipulated on an
on-line basis and all operations of the equipment
are executed through every day English commands.
All data is displayed on cathode ray tubes and
projection screens, including instructions to the
players on how to operate the equipment and how
to play the Decision Game. Input to the equipment is provided through (1) a permanently
labeled keyboard, (2) a blind keyboard that can
be provided with appropriate labels through a set
of plastic overlays. The computer action resulting from depressing of keys must be programmed
and is not wired permanently. By providing a set
of independently operated input-output consoles,
connected on-line to the same computer system,
a significant advance in the art of the design of
management systems is provided.
Intr oduction
The management planning and control technique
described in this presentation has been developed
for certain military and civilian activities with
the purpose of assisting executives in evaluating
and re -programming complex activitie s. However, it is believed that the technique is equally
applicable to the planning and control of other
large scale research, development, production
and constructlon programs.
In order to apply the planning and control technique to an activity it is neces sary to divide the

activity into "elementary" programming blocks.
The technique requires that first alternate scheduling and financial data on each of these planning
blocks be developed. In addition, it is necessary
to formulate explicitly the inter -relationships
between the planning blocks. These relationships
specify the per mis sible time phasing of the elementary programming blocks, the associated
dollar and manpower requirements, facilitie s
requirements and other financial rE;,quirements.
The planning and control technique employs a network analysis of the various activities involved
and permits the exploration of a large number of
planning combinations.
The primary purpose of the management
planning and control system is to assist executives
in re-programrning. As an illustration, suppose
that plans are compared with progress, a deviation is observed and re-programming of the different activities is required. For instance, it
might be necessary (1) to cancel a program, (2)
to stretch another one out, (3) to accelerate one,
or (4) to decrease or increase production quantities. Another situation when the need for reprogramming arises, when there is a budgetary
change and financial trade-offs between various
programs must be considered. For instance,
executives may want to know that if a particular
deadline is postponed by six months, how many
dollars and what manpower can be made available
to another program, and by how much can this
other program be accelerated.
The actual program analysis and re -programming activity is carried out through the medium
of a Management Decision Game.
Brief Description of Decision Gaming Technique
The Decis ion Game is to be played in three
steps. As a first step the players of the Decision
Game gather in the Control Roqm where the Game
is to be played and where the various information
displays can be retrieved with the aid of the computer system. The displays present all the important planning factors relating to the activities
to be re -programmed. The time phasing of
various missions, deadlines and goals, and the
associated loading of various facilities can all be
displayed. The associated financial information
can also be shown with sufficient detail so that
financial consequences of re-programming decisions can be made. Provisions are made to
retrieve further back-up information when requested, from a file of status of progress and

18

1.3
alternatives.
As a first step of re-programming a comprehensive analysis of the status of the programs is
carried out. The computer system is provided
with the capability of furnishing status information on a real-time basis and in everyday English.
Information related to all matters pertaining to
the progress of various programs is displayed in
cathode -ray tubes or in projection screen. After
this status analysis is completed, the players
have adequate information to perform the second
step of the Game.
This second step of the Game consists of making a "move". Such a Ilmove ll may involve a titne
shift of some of the deadlines, mile stones or
goals and / or a change in the delivery quantities
involved in the program. As suggestions for reprogramming Ilmoves II are tnade, the proposed
changes are put into the computer system through
the use of an appropriate keyboard and Communication Display Tube. At the direction of the
operator the computer and associated equipment
takes over and the third step of the Game, that
is the re-programtning computations, are
carried out.
This third step of the Decision Game is executed by the cotnputer in accordance wit,h mathematical models and associated computer routines
stored in its memory. Within a time span of
seconds the computer prepares a new program,
including all the new deadlines the new phasing
of sub -programs, facilities loading, manpower
and financial implications. When the computer
finishes the cOtnputations, the data is presented
to the players through cathode -ray tubes and/or
slide projections. By examining the various
displays and by retrieving tnore detailed infortnation, the players can evaluate whether the suggested solution to the re -progratntning problem
is satisfactory.
In tnost situations the first suggested program
will result in conditions that are not acceptable
to the players of the Game. Therefore, after
considering the results of the Ilmovell and discussing further implications of the data, a new
proposal for re-programtning will emerge and a
new cycle of the Decision Game will be entered
upon. By a series of steps it is pos sible to
develop a final program that is acceptable to the
players.

the examination of a panorama of alternatives.
The analysis can be performed only if -in the memory of the computer, techniques for examining
many alternate possibilities are stored and programmed. It is recognized that it is itnpossible
to store all the possible alternatives and therefore, a method to study alternatives must be
provided. The analysis is made pos sible by the
application of tnathematical modeling techniques
and by the storage of certain basic system paratneters. The mathematical model uses the eletnentary programming activities as basic building
blocks and relates these activities to each other
through mathetnatical relationships. For instance,
alternate ways to accomplish a basic programming block can be as sociated with various estimates of completion dates and costs. The
mathematical model sum.marizes the relationships, and also relates the different activities to
each other through equations and inequalities.
Manpower and financial requirements appear as
dependent variables, whereas the titne phasing
of the various activities as independent variables.
In order to avoid the neces sity of manipulating
a large number of param.eters, sub-optimizing
techniques are introduced. For instance, it might
be required that certain types of sub-programs
be accomplished at a minimum cost and this policy
can be embodied in a system of equations through
mathematical progranuning techniques. By such
relationships, the majority of the variables of
the system can be tnade to depend on a few control variables. With the aid of mathematical
models and sub -optimization techniques it become s
pos sible for the players to manipulate only a few
of the major variables and still examine a large
number of alternate plans.
Equipment Requirements
There is no equipment on the shelf today that
can carry out in all its details the management
planning and control technique described here.
However, there is equipment available, which
with minor modifications ~ould possess the capability required. A detailed study of the RamoWooldridge Polymorphic Computer System and
Display Analysis Console, for instance, shows
that essentially all the required features could be
made available in a short time. This computer
system has been described elsewhere,and in this
discussion equipment details will not be included.
Decision Gaming

At the termination of the gaming exercise all
the implications of the final program are recotnputed with greater accuracy. It is not expected
that this re-cotnputation will result in major
changes, but only that the re-computation will
provide an accurate, acceptable and detailed plan.
All conununications between man and machine
are performed in a real-time manner and in
everyday English.
The File of Status and Alternatives
The Gaming Technique described here allows

Detailed Description of Decision Gaming System
The fundamental concepts underlying the
Decision Game are shown pictorially in a simplified form in Figure 1. Three displays enable
the players to communicate with the computer.
The first of these is a visual representation of the
time phasing of all the important missions and
goals. The information on this display is
schematically represented in the upper part of
Figure 1 and is to be displayed in the IIProgram
Network Tubet! of Figure 2 {projection capability

19
1.3
can be provided if desired so that a group of participants can analyze the data). Sufficient details
will be shown 'so that all milestones of importance
are displayed, but the data will not be so detailed
as to confuse the players. As the Game starts,
various questions will arise which will not be immediately answerable by the displayed material.
To meet this condition, back-up displays will be
stored which can be retrieved by the players as
requested. By this technique, it will be possible
for the players to go into any degree of detail in
the time phasing of the mis sions and goals without
making the presentation too cumbersome or confusing.
A part of the display on the "Program Network
Tube" is the visual representation of the utilization and loading of the different facilities ass ociated with the programs considered. This display
is shown by the third item from the top in Figure
1. All the previous comments made in connection
with the visual representation of Programs A and
B apply for the Facilities Loading displays, too.
Sufficient detail will be given so that the player
can appraise the state and progress of various
programs, and again sufficient back-up information will be available through retrieval.
The second display refers to dollars, costs,
manpower, and other resources. These are
represented graphically in the lower part of Figure 1 and are to be displayed on the "Resources
Requirements Tube" of Figure 2. The dollar and
manpower profiles as they unfold in time will be
represented in sufficient detail so that all the important information for the players will be furnished. In addition, when it is required, the
players will be furnished with hard copies of
printed financial information.
The display capability so far described furnishes the players of the Game with such pertinent
information as past history, status, and future
projections of programs. Particular emphasis is
placed on the preparation of this information in
such a form that organizational structure and
responsibilities are directly tied in to the information displayed.
The lower right corner of Figure 2 shows the
"Man Machine Communication Display". This is
the tube that offers choices of instructions to the
player in plain English. This tube is used mostly
for non-standard ty.pe of instruction to the player,
as ordinary instructions (say: "Machine Is Busy")
are provided through the illumination of status
lights.
So far we have described the display systems
and the type of information stored. We are now
ready to proceed to the description of how the
Decision Game is to be played. In order~to be
able to speak in more specific terms, we take
the hypothetical problem of a new requirement,
that a particular mission is to be accomplished
one year ahead of schedule. This new requirement requires the acceleration of a major program and a reorientation of the resources

available.
When such a problem arises, various discussions take place at different managerial levels.
We do not propose that the Decision Game is to
replace these conferences. However, after a
preliminary consideration of the problem the
appropriate management group gathers in the control room to play the Decision Game. By a step
by step procedure, they evaluate, modify and
sharpen the preliminary ideas that have risen in
connection with this problem of advancing the
completion date of a major mission.
When the group meets the first time in the
control room, the players begin by retrieving a
number of different displays to update and verify
their knowledge of the status programs. Such a
review cons ists of inspecting the principal displays
ass ociated with the problem and of retrieving'
various back-up information. After such a preliminary discussion, a proposed first solution to
the reprogramming problem is suggested and information defining the proposed change is keypunched into the computer.
At the instruction of the players the computer
begins to carry out the routine associated with the
particular reprogramm.ing problem introduced.
The computer consults the Data and Program
Reservoir containing the file of status and alternatives shown on the lower left-hand side in
Figure 2, and on the basis of stored information
and routines, computes dollar and manpower
requirements. In addition, facilities requirements
and loading are checked and computations are
made to determine whether the desired acceleration is feasible at all.
As the computer proceeds through its routine,
it might find that the proposed acceleration is
impossible or impractical. It is possible that
even if all projects are put on a crash basis the
mission could not be accomplished within an
acceptable date. It may be that for instance manpower is not available, even if more ~hifts are
employed. Under such conditions, the computer
will indicate that the plan is not feas ible and it
will display on the "Communication Display Tube"
a warning signal, which shows in detail why the
proposed solution to the reprogramming problem
is not feasible.
At this point, a group discussion follows to
determine whether by a higher order of decision
a solution could be found. For instance, it :r;night
be decided that another facility can be built or
made available, or that another contractor can be
called in. Information available to the decis ion
maker will not always be programmed into the
computer and, conseqlJently, feasibility indicated
by the computer will occasionally be considered
as tentative.

If, indeed, a need for such a new alternative
way of proceeding with the problem exists, this
information must be put into quantitative form and
fed into the machine. On the other hand, if the

20

1.3
computer indicates general feasibility, then the
players can immediately proceed to further evaluation of the proposed program.
When the program modification is feas ible, the
players are primarily con~erned with resource
requirements and with dollar and manpower profiles associated with the program. It is very
likely that the first solution proposed will not be
acceptable from the point of view of budgetary
considerations. It is likely that the costs at certain phases of the program will be beyond possib~e
funding, and perhaps at some other times there
will be an indication of surplus funds. This, then,
is the point where the players reconsider the
time phasing of the mission and goals and propose
an alternative. When the players agree on the
next trial of the program phasing, information is
fed into the computer and the computer proceeds
with computations to prepare a new program.
Again, the computer first explores feasibility and
then proceeds to the detailed generation of the
resource requirements.
It is seen that through a step by step process
of deliberation, discussion and computer routines,
the players will reach better and better solutions
to the reprogramming problem. It is envisioned
that programming computations will be carried
out first by a llquick and dirtyll method and then
by a more accurate routine. This will allow the
players to explore tentative alternatives rapidly
and there will be no unnecessary delay in waiting
for accurate computations which would not be
utilized in actual program plans. The computer
. will carry out accurate computations either
automatically (when co:mputing time is available)
or at the special direction of the player. This
approach allows the decision :makers to make
rapid changes and explore and evaluate dozens of
different program proposals. As the Decision
Game progresses, more and more satisfactory
solutions to the reprogramming problem will be
found. Towards the terminal phase of the gaming
exercise, the players may des ire highly accurate
estimates of the various program details. If
this is so, it may be neces sary to direct the
computer to carry out :more accurate special
program computations, and it may then be necessary for the players to wait for a longer period of
time to get the phasing of programs and the
resource requirements. Finally, the computer
is directed to develop and print a definitive program which will be used as a planning document.
Computation of such a program. may require hours,
and consultation with other agencies and
contractors.
So far, we have given only an outline of how
the Decision Game is to be played and described
only those phenomena that will be observed by
the players. Now we proceed to take a look inside the equipment and see how the various
logical steps, routines and computations are
carried out.

Illustration of Reprogramming Computations
The basic principle in carrying out reprogramming operations is to provide the computer with
data on possible alternatives and also with the
myriads of details on how these alternatives can
be combined into programs. The computer can be
programmed to go through a large number of calculations in an efficient fashion, and therefore
alternate programs can be generated by the computer in a matter of seconds. In order to illustrate the techniques, we will describe an extremely
simple but still significant reprogramming prdblem.
Figure 3 is a chart showing six different jobs
and the time phasing of the start and completion
dates of each of these jobs. In this simplified
programming Game, we are concerned only with
the monthly dollar expenditures which are shown
in the bottom of Figure 3. Suppose the player
desires (1) to accelerate by two months the accomplishment of Goal B (that is the terminal dates
of Job No.3 and 5); (2) to accelerate by three
months the final completion of the mission, that
is of Goal A; (3) leave all other goals unchanged.
The computer is to determine whether such an
acceleration in the program is feasible, and what
kind of dollar expenditures would be associated
with this accelerated program.
As this reprogramming information is keyed
into the machine, the machine examine s all jobs
to see which is immediately affected by the acceleration of Goals A and B. The computer selects
Jobs 3, 5 and 6 and evaluates the possibility of
accelerating those three jobs. It finds that the
time span of Jobs 3 and 5 are to be compressed
by two months and of Job 6 by one month.
At this point, the computer seeks information
on alternative ways of accomplishing Jobs 3, 5,
and 6. As the computer consults the file of alternatives, it finds for each job the time -cost relationship shown in Figure 4. The horizontal axis
shows alternative time spans allowed for the job,
the vertical axis shows the total dollars that must
be expended, if the job is to be accomplished in
the time specified. It is seen, for instance, that
a crash program--doing the job in the shortest
possible time--requires more total funds than a
more orderly and efficient execution of the task.
In the case of a stretch-out, due to overhead and
some other supporting activities, the total cost of
the job would also increase. The computer also
finds how these dollars would be expended in time.
(Dotted lines in Figure 4.) The file of alternatives
has curves of this type for each of the jobs and
therefore the computer can establish that the jobs
can indeed be accelerated to the desired time span,
but that a higher expenditure of funds is required.
Using this information, the computer can replace
the previous budgets for Jobs 3, 5, and 6 with the
new budgets and determine a new dollar profile
associated with the accelerated program. We see
that when the computer reprogram.s, it first proceeds through these computational steps and then
transmits the information to the display devices.
The player can visually observe the required

21
1.3
funding associated with the accelerated program.
We recognize that in a real problem we would
deal with a much more complicated set of routines. Manpower profiles would have to be computed, facilities loadings would have to be checked,
many other items of information on compatibility
would have to be considered. In the case of prototype production, or in other tasks where quantities are involved, relationships dealing with
"quantity made" would have to be included in the
analysis. However, basically, these considerations would only complicate (admittedly by a
great extent) the routines that the computer would
have to go through, but, conceptually, reality
would not add significant new difficulties to the
method of solution.
The time cost relationships as shown in Figure
4 form the basis of the file of alternatives that a
computer has to consult. As we already mentioned, there are types of problems where more
com.plex mathematical models form the building
blocks for the file of alternatives. However, for
purposes of our discussion, we will concentrate
on the concept of time-cost relationships and we
will show how such relationships can be generated.
We will show how the basic input data i,s to be
obtained and how these data can be built into the
appropriate files for representing various alternatives that the programming task may require.

secretary. This establishes his minimum effort
level and gives the "minimum effort'! point in
Figure 5. We connect the three points by a curve
and obtain a time-effort relationship and we
as sume that we could also operate at intermediate points on this curve. With the aid of the
curve shown in Figure 5, we can determine
the time-cost relationship shown in Figure 4.
All we have to do is to multiply the rate of effort
by the time required for the job, to get total
costs.
We see, then, that we have a technique to get
time -cost relationships, at least for relatively
simple jobs. However, if we want to extend
this technique to more complex tasks, we run
into problems. It is difficult or impossible to
find managers who have all the details of a complex job. Consequently, in order to make cost
estimates, the manager must w'brk with his subordinates and must combine in a complex fashion
many items of information. This combination of
data is a tedious and difficult job but is precisely
the kind of task that computers can execute with
great efficiency. Therefore, we propose to prepare time'-cost curves for complex jobs with the
aid of computers.
We will show how, with the
aid of mathematical models and sub -optimization
technique, one can construct time -cost relationships.
Sub -Optimization Cons ider ations

Concept of Alternatives
Let us reiterate the type of information we
seek. The player moves som.e of the gaming goals
in time and certain jobs must be performed within the time limits indicated by the player. We
need to find a way to determine the dollar requirem.ents associated with the various alternatives.
Let us begin by considering a relatively simple
job or task. Suppose that there is a single manager in charge, and let us as sume that this manager has a good grasp of all the details involved
o£ this particular task. The manager does his
own planning with paper and pencil and by discussions with his associates. We ask him to determine how much would it cost to perform this job
in an "orderly" fashion. After studying the problem, he estimates manpower, material, overhead
and dollar requirements. In Figure 5, the financial information is shown in a graphical form. In
the horizontal axis we show the tim.e allowed to
com.plete the task; on the vertical axis, we show
the associated effort (say dollars per week) required. "Orderly" performance of the task is
represented by the "most efficient" point in the
chart. We also ask the manager to determine
what it would take to complete the job on a crash
basis. He would need more men, more resources,
he would require a larger effort, but he could complete the job in a shorter time. This crash program is shown in our chart in Figure 5 by the
" m inim.um time" point. We can also ask him to
determ.ine the minimum level of effort required
to do the job at all. He needs two mechanical
engineers, an electronic expert, a technician, a

Let us take a simple combination of two jobs
which have to be performed in sequence. Various
alternate time spans are allowed either for Job
No. 1 or No.2. This implies a number of combinations of ways that the two jobs can be performed. In Figure 6 we show the problem in a
graphic way. Suppose tentatively we select a
certain duration for Job No.1, and we determine
the associated dollars required with the aid of
the time-cost relationship. In Figure 6 this timecost relationship is represented by point A. Now
by starting with this time span, we can as sign
different time spans to Job No.2. A possible
representation for Job No. 2 is point B. It is
seen that we can combine the two time-cost
curves in many different ways. In Figure 7, the
various poss ible time -cost curves for Job No. 2
are shown by dotted lines. Now we need a policy
to select, out of these many possibilities, the
desirable ones.
Suppose we agree that we want to complete
the two jobs within a given time span, but with
the least amount of money. Let us recognize
that when the combined time -span for the two
jobs is specified, still there are many ways to
do the two jobs; out of these many possibilities
there is one that yields the lowest cost. In
Figure 7, these low cost combinations are represented by the envelope of the dotted curves. We
say then that this envelope, corresponds to our
policy of minimum cost, and this envelope is the
combined time-cost relationship for the two jobs
to be performed. For instance, if we wish to
complete the two jobs at point P in Figure 7, we

22
1.3
draw the ve~tical line from point P until we reach
the envelope at point Q. This gives the combined
cost of the two jobs. Working backwards from
point Q, we can get point R which represents the
time and cost requirements of Job No. 1.
The policy we used here is to perform the two
jobs with the lowest possible cost. If there is
another policy such as say a constant manpower
requirement or the utilization of a facility, etc.,
each of these policies would have to be programmed into the computer. The important point,
however, is that even if complex policies are
formulated, due to the high-speed capability of
the computers, consequences of these policies can
be deduced efficiently.
Actually, the computer would not construct the
envelope of the curves, but would solve the appropriate mathemaJ;ical problem. It is easy to show
that the two jobs are to be combined in such a
fashion that the following equation holds:
(1 )

Here on the left-hand side we have the derivatives
of the time -cost relationship for the first task and
on the right-hand side, the derivative relationship for the second task.
The computer would compute these derivatives,
select the appropriate combinations of the tasks
and generate the new time -cost relationships.
In Figure 8, we show a somewhat more complicated problem when a sequence of jobs is to be
performed. Here it can be shown that the following equation must hold:

de.1


== _____ -----_ .. -- ---------- ------

rr----~-

1962
TIME
Figure 1.

1963
~

1964

1965

Basic Gaming Model

......

N

e.,),-....J

......

N

WOO

MAN-MACHINE SYSTEM

T
ITj

....

(JQ

~

11

PROGRAM

RESOURCES

NETWORK

REQUIREMENTS

TUBE

TUBE

(D

•

~

~
PJ
~
I

~

PJ
:::r'

()

r-I

s·

CONTROL PANEL

(D

CI.l

'<

til
..,.

T

(D

s
COMP1UTER
DATA a
PROGRAM
RESERVOIR

I

I

MAN-MACHINE
COMMUNICATION
TUBE

ELEMENTARY PROGRAMMING GAME

JOB NO.1
JOB NO.2

,I

+

t
I

+8
I

L::.J
,

:,

II

I

I

I

I
I
I

I
I

Fl

Fl
~

F1
------tl:::::.l

F1
l::::::.J

I

:

JOB NO.4

-------~!--~~
-'
------~~

JOB NO.5

------~§

JOB NO.3

JOB NO.6

~

+A

I
I
I

I
I

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

I .

r=1
~~---+I------------~
II

§~--~
I

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

E9
~

S/ MO.

1961

1962
Figure 3.

1963
TIME

1964

1965

Elementary Programming Game

.......

N
w(O

30
1.3

TIME-COST RELATIONSHIPS

$

I

2

Figure 4.

3
4
TIME IN WEEKS

5

6

Time-Cost Relationships

TIME-EFFORT RELATIONSHIPS

MOST EFFICIENT

$/WEEK

MINIMUM
EFFORT

2

3

4

TIME IN WEEKS
Figure 5.

Time-Effort Relationships

5

6

31
1.3

CONCEPT OF GAMING GOALS

$

I

§~--§
JOB NO.1 I

1=1

Figure 6.

TIME--I

I~I JOB NO.2
Concept of Gaming Goals

POSITIONING OF SLAVE GOALS

$

p

TIME
Figure 7.

~

Positioning of Slave Goals

32
1.3

SUBOPTIMIZATION TECHNIQUE

ex

xx

xx

x x

•
de;

CONDITION OF OPTIMALITY:

Tg

I
I
I
I

dC.
dli_ _ _II
_

T3

dC2
dT2

A

dlj

de3
dT3

TI

T3

T2

Figure 8.

Suboptimization Technique

POSITIONING OF SLAVE GOAL THROUGH
SUBSIDIARY GAMING

,
§)(

+

I
I

I
:

¥x

~

I

I

I

I

I

I

I

I

I

~

I

+
I

A

I

:
I

I

I

B

I
I

~~--_*~--X~I--~X~~H~X----X~~
Figure 9.

Subsidiary Gaming

=A

~

ETC.

MULTISTAGE MAN MACHINE DECISION SYSTEM

1 . - - - - - -,
I

I

I
I
I
,

I
I

,
I

I
I

FILE OF
AL TERNATIVES

L_____

r------l
I
I

I

MAN

I

I

I

I
I
I

I
I

FILE OF
ALTERNATIVES

I
'--- - -

I
I

----'

I
,

---1

,-- ----,
I

MAN

r-------,

I"

I

I
I
I

I

I

I

I

I
I
I

:

:

I
I

I

I
I

I
___'__

I

L _ _ _ _ _ _ -'

Figure 10.

I ALTERNATIVES
I
L _ _ _ _ _ --1

Multistage Gaming System

I-'

W
WW

34

1.3

CONFIDENCE LIMITS FOR GOALS
GOAL
NO. I

GOAL
NO.2

Figure 11.

GOAL
NO.3

Confidence Limits

CASCADING PROBABILITIES

.£\..
I

Jl
I

A

I

COMPOSITE
PROB ABI LITY

0-

2

=
Figure 12. Cascading Probabilities

'il~[@!~Wji'
!~~~; li ~ j~i j~ fiI I K~~:g!~D Ij::::~::::::::::::::::::::::::
~j~ ~I !i~ I Ijil S~~~:L lii~!~U~I~SJI ~~!:~T~~: 1111~ii~lmli~i~!1
::::::::::::::::~::::::::::::

START
ROUTINE

::::::::::::::::::::::::::::::'

IN PUT

r~i:::::::~~:i:::::::~:1

CANCEL

NO. I .

::::::::::::::::::::::::::::::

::::::::::::::::::::::::::::::

II-I- - - - - I I

I: : :~: : ~: : :~: : ~: I

PROCEED
WITH
ROUTINE

BAC K

r~:::::ii::i::::::;;:1

SPEC IAL

1::i::ii:~::~ii::;~~~:1 ~~E~~OOR~ I:*::::~±::~~::::::~I

~i@llllil R~~~~~ l~illiliilli~ I~~~!. l@~i~! i~j~j~! !i.~i I=~~:. ' ~@ii!MI~i~ K{~~i£~CD ~li:I@!~!!!!:

11
... - - - - - 1

LEFT

+ ri: : : 1: @lt i:1
I ~~~~~~~~~~~~~~~

rm:iii:Mi:1

RIGHT
TUBE

~ ~ ~ ~ ~ ~ ~ ~ ~ f~ ~ ~

BOTH
TUBES

II

~~~~~~~~~~~~~~

OPERATOR:

IMi:m::t~:1

~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ . IN-~~:;~SS ~ ~ ~ 1~ ~ ~ ~ ~t~ ~ ~

ADVANCE

I

r*~ : :~: ~t: :'~: 1

r::i i:'i : i: : :1
~1!!i!I:~MI

~ fi.i{I~!i1i!

BA CK
SPACE

1~~~~~t:~~§~(1
ff!lli~i!i

I~~~~~~*"{:'I
!f~f'lI.jiM

OPERATOR'
ERROR'

I~~~~~::::~~:I
~m!i~ItfI!!i

I
COMM.
TUBE

ri!!l!i!!j!!i!jl
::::::::::::::::::::::::::::::

....

P R I NT

r~i:~@:::::::::~:1

III ~ II II II O::~~::R:I.

Ilt------I
TUBE

II

I!i~:iii:~m::ml

4

1!liijJi.~lljl ::~;~:~ fim1£ii
:::::::~::::::::::~::~::~::

S LID E

::::~~:~:::::::::::::::::::

JPROGRAM

,11'
1!111!1
:::::::::~:::::~:::::::::::
J

T~~T Imiti~:-ril
:::::::::::~::::::::::~::5

-

Figure 13. Control Panel

.......

(".)

•(".) en

..... w

em
W

EXPENDITURES (Million Dollars)
Total
1961
D and P

1. ATLAS

Sys-conn

MINUTEMAN

B-70

1964

58

16

316

398

367

258

183

2437

81

46

74

223

1245

10

39

102

1335

828

620

453

3115

4

4013

Sys-conn

Sys -Conn
D and P

4.

B-52

Sys-conn

5.

B-58

Sys-conn

D and P

____120__ W1~

75

D and P
3.

1963

65

D and P
2.

1962

95
1060

214

25
1149

120
1134

1134

1134

9931

309

2748

58

25

159

324

309

309

Sys-conn

915

741

510

197

D and P

163

87

Sys=conn

174

230

83

D and P
6.

B-47

7.

KC -135

8.

GAM-87

D and P

2363
250

230

230

230

64

26

33

505

3

446

Sys-conn

1200

1200

1200

1100

1100

WEAPONS SYSTEMS

4591

4795

5173

4367

4002

LIMIT AT IONS

4600

4500

NON-SYSTEM
TOT AL FOR ALL

Figure 14.

Weapon Systems Expenditures

2084

Total in Program

276

Total Delivery to Date

8

o

Units per Squadron

12

Active Squadrons

Go Ahead Date

May 1958

1st Alternate

Last Delivery

June 1964

3rd Alternate

2nd Alternate

EXPENDITURES
All System.s

ATLAS
Squadrons
Projected

D&P

Sy. Con.

49

4

65

316

138
1962
1st Alternate
2nd
3rd

11

75

230
1963
1st Alternate
2nd
3rd

19

58

Year

To Be
Delivered

1961
1st Alternate
2nd
3rd

~

Total

Limitation

1200

4591

4400

398

1200

4795

367

1200

5173

Non-Sys.

--

Figure 15.

Atlas Program

......

w

w""

39
1.4
l1PPLICATION OF DIGITAL Sn·IDLATIOH T.8ClITUQ.UES
TO HIGInlAY DESIGN P:t10DLii!l':rs
Aaro:c. Glickstein
l·lid".·Test Re:Jearcn Institute
Kansas City, Missouri
S. L. Le-vJ
l·Iid'·lest Research Institute
ICal1cas City, Hissouri

i\. study of the operatinG chara.cteristics of the driver-vehicle combination has yielded a General diGital siillv~atiolJ. i:.1odel. Tllis sil.1ulation lLlodel, 1iJ.lich can duplicate ".:.raffic flO'll
on a 17,OOO-ft. sectiol1 of a rree'·iB.~T illCludiIl;:;;
two on-rar:lps and tl'lQ off -rarU.ps, can be used to
econoI1ically eval'llate alterIlute desien cri-ver:'a.

The slluulated venicle in the Bodel,
follOl·ring decision rules based on actual traffic
behavior, is alloiTecl ~GO maneuver throuQl the section of free,·ray Ul1cler study. The effects of
cl1anGes in traffic volv,me, traffic velocity,
freeivay configu.ratior., e-c,c., can then be evaluated by notinG chanGes in the computer outl)ut of
traverse time, waitinG time on ramp, volUJ:.:J.evelocity relait:Qnship, vleavinG cOl:lplexity, etc.
The con:.put.er simulatio:1. thus creates a
duplication of the real situation at a' s::-_lall
fraction of the cost of studyinG the real system.
Furthenllore, the siI;iulation allow's: (1) the evalv.ation of various freei-ray configurations "rithout
the expense of their construction, and (2) the
perfor~ce of controlled experll:lents ~npossible
to perfom. ".'lith the actual traffic.

1. \'i11U-G s110uld De tile lenGth of an
acceleration lx..e:·
2. Hho:c is tile effect on traffic floil
of locatin::; int;erchal.1z,e areas at various di8t&"1.ces fra.,l one another.
3. ';':nat is tile effect of various speed
m.iniJllW,lS ar::.cl UlXir,::w.J.s on the trai':2ic f101.[;

DurinG t~e past decade a n~,mber of descriptive theor~es concerninG vehic~~ar traffic
have been nut fOrlm.rd. In a;,: atte;:rfPt to classify
these theories I-IaiL:2TG 7 pro:9osed th~e t;y:.oes.
The first, an analytical and detenninistic r.loa..el cOl1side:rs tne c;eometrical and physical characteristics of tiJ.e autoaobile and the assur,leo.. behavior of t}1e driver. ~lhen specific
values of vclocit,y, acceleration, and. vehicle
f0110vlinC behavior are postulated, it is possi-ole
to descri-be the 1:"..otion of a n'l1:::ioer of cars in one
lane. The resulJcs, altlloU£,;h precise, are lL'liteo..
to a SL.la11 nUltiJer of vehicles. Hork i~ this area
has been carried out by Pipes; 15 Chaxl.dler, Eerl:J.a::.,
and Hont:roll; 1 aYJ.d Eenuan, I-lontroll, Potts, and
Rothery. 9

Introduction
In rece:c.t years, because of t;he Grmrth
of the hi01'i·rclY buildi:c.C; program., traffic enGineers have become concerned over t.he lac::C of
lmoi-Tledce ab01;;:'c, the nature of traffic flo'<7. This
lack of YJlovTledGe has haIf!.pered the desi[...:r.. enGineer in his ability to d_esicn an e;~pressi'i'ay ul1icn
lull 1:10Ve :traffic s;:,loothly and efficiently at a
llunilm.1rll cost. All too fl~equently after a hiGh~my
has been opened it. has been fOlli1d that traffic
:9roolems arise which requi:re ezterlsive :reo.esiQ1
and construction.
For exaraple the fol10vnnG questions
have still to be ans'lvered.

A second ap:?J:'oach has been to consider
trei'fic aD a stochastic process and to treat i~c
by the theory of Queues. Tl;tis apnroach has been
utilized by 1';e'l"eli,13 Tanner. 16 The tl-:eory of
queues can, hm·Tever, only be applied ',Then vehicles r.love at esnentially the same speed and ',T::-ler.l
all vehicles enter t:1e syste;:il at one poin"G.
Reasonable results have been obtained when traffic is act"J.ally Q.ueued, velocity 1..1lliforiJ. a:r~cl tiJ.e
driver has fe-"'l free decisions.
The third approach descriues traffic
flov1 in a continuum. This approacn has been utilized by Nev1ell,11,12 Tse Sun CiJ.ovl,13 Greenbers,6
aYJ.d Lighthill u.nd HhithD.n.10 ..Ul three approaches are limited by the fact that it is

40
1.4
assumed that vehicles do not overtake or pass
each other.
Simulation Models
It is, however, possible to construct a
simulation model for the study of traffic flow
which is much more general than the three types
mentioned. Simulation has been defined by
Harlin88 as tithe technique of setting up a stochastic model of a real system which neither
oversimplifies the system to the point where the
model becomes trivial nor incorporates' so many
features of the real system that the model becomes untractable or prohibitively clumsy."
Early work in the field of traffic simulation was carried out by Gerlo1,lgb.,2 Goode, 4
Wong,19 Trautman, Davis, et al. 17 Gerlough in a
pilot study simulated two lanes of a highway system one-fourth of a mile long. In this simulation model the vehicles were allowed to choose
their speeds from a three increment distribution
of desired speeds. This program run on the SWAC
cQnputer required approximately 35 to 38 seconds
of computer time to simulate one second of real
time.
Goode 5 in his simulation model represented a Single intersection of a simple type.
Provisions were made for the Origination of cars
in lanes approaching the intersection, and for a
stochastic mechanism for the determination of the
direction of travel of any particular car entering the intersection (i.e., ri@lt, straight
ah~~, gr left). The measure of effectiveness
used ivaS the average delay per car. The ratio of
computer time to real time v18.s 3H:':.
Although these early models were of
rather limited extent, they indicated that simulation procedures could be used to run controlled
experiments in order to gain a better understandinG of the nature of traffic flow.
The Expressway Interchange Model
In 1958 the Midwest Research Institute,
a grant frQu the Bureau of Public Roads,
started a research program on the simulation of
expressway traffiC flovT. The first phase of this
proGram was devoted to the simulation of an onrarnp area of the ex-pres sway .14 That model simulated a 1,700-ft. section of a three lane expressway including one on-ramp. Based on the ex~rience of this pilot study a more general model
was developed. That simulation model is described in this paper.
Th~der

The study Area
The portion of the expressway under
study is set up in a 4 x N matrix (N.! 999),
Fig. 1. The four rows represent: (1) the three
through lanes 1, 2, and 3, and (2) the ramp, acceleration, and deceleration Lane 5. Each of the
N blocks represents a 17-ft. section of freeway,
the approximate length of an automobile. For the
simulation runs on the computer, any value of
(N ~ 999) can be utilized. Locations of interchanges are designated as follows:
A = ramp input location
B = nose of on-ramp
C - end of acceleration lane
D = beginning of deceleration lane
E = nose of off-ramp
F = off-ramp output location
The program is very fleXible and permits the onand off-ramps to be located at any point on the
section of freeway under investigation.
Input Factors for Simulation Model
In order to obtain a true duplication
of actual traffic behavior on the freeway the
Simulation model should contain all factors which
influence traffic behavior. In this model the
folloVTing factors are considered:
1.

The volume of entering and exiting

2.

The distribution of vehicles to

3.

The velocity distribution of

vehicles.
lanes.
vehicles.
4. The gap acceptance distribution of
merging and weaving vehicles.
5. Tile acceleration of entering
vehicles.
6. Tne deceleration of exiting
vehicles.
7. The distribution to lanes of exiting vehicles.
In addition, all vehicles are allm'Ted to shift
lanes in order to pass slower mOving vehicles in
front of them.
Simulation yrocedure
Basically, the procedure consists of
simulating the arrivals of cars into the section
of highway under consideration and then controlling the action of the vehicle by a set of decision processes. During each second of real time
each vehicle in the matrix is examined. The

41
1.4
vehicle is allovred to advance, vreave, merge, accelerate, decelerate, or exit according to logical rules describine the behavior of actual vehicle-driver combinations. Just prior to examining all vehicles at each second, each of the
input locations is evaluated. Inspection starts
at vehicles closest to the end of the section of
highvay under examination and proceeds to vehicles in the input location.

7. The
traverse times.
8. The
traverse titles.
9. The
each lane.
10. The

distribution of through-vehicle
distribution of ramp vehicle
average vehicle velocity in
number of veaves from:

a.
b.

A description of the over-all logic involved in processing a vehicle through the system
is given in the flov1 diagram (Fig. 2). Detailed
flo\V diagrams for the computer can be obtained in
Reference No.3.

c.
d.

Lane
Lane
Lane
Lane

1
2
2
3

to
to
to
to

2
1
3
2

Input Data Preparation
Computer Operating Note

The folloving parameters are used in
the flow' diaGrams:

v = total volume in Lanes 1, 2, and 3;
Vi

vehicles per hour
per hour in the ith lane

= vehicles
3

L

En
iBn-l

iBn

v
iVn

\f

V· = V

i=l :l= block number of vehicle under
inspection
= block number of vehicle in ith lane
preceding vehicle under inspection
= block number of vehicle in ith lane
parallel to or behind vehicle
under inspection
= velocity of vehicle under inspection
= velocity of vehicle in ith lane,
parallel to or behind vehicle
under inspection
= time gap

Simulation Output
The present model \Vas programed so that
the follovTing infonnation can be obtained about
each simulation run:

1. The volume of vehicles traversing
the system in each lane.
2. The volume of vehicles entering the
freeway through each on-ramp.
3. Ti~e volume of vehicles exiting the
system at each off-ramp.
4. The number of vehicles \Vhich stop
on the acceleration lane.
5. The len6th of the queue on the ac ..
celeration lane.
6. The number of vehicles that desire
to exit but cannot.

To operate the program, an IBl,l 704 computer is needed vThich has at least 8,192 words of
core storace. One magnetic tape unit is needed
(Logical Tape 0), that one being used to produce
an output tape for printinG an an off-line printer.
On the Canputer Console, Sense Svitch 4
should be depressed if it is desired to obtain in
the problem output a listing Of all problem input
parameters.
To run the program, assemble the proGraLl. deck with a series of Control Cards immediately follOwing the deck. The munber and type of
Control Card.s vTill control the analyses to be run.
Tnere are 13 different Control Cards
that can be entered, which insert the necessary
setup data d.escribinG the system confi2;u.ration
and velocity and vTeaving distributions. Appropriate Control Cards should immediately follOif
the progr&1. decl~. A Problem Card, containing
data peculiar to each analysis then follow's and,
after this ProbleEJ. Card, chane;e cards for any of
the control cards and other Problem Cards may be
entered to control the running of a series of
traffic problems.
Control Card Fonruat
The 13 Control Cards are divided into
four groups. Five cards provide the velocity
distribution data for Lanes 1, 2, 3, and the tvTO
possible on-ramps. A second group of five cards
proy-ide data concerning gap acceptance for veaving operations. A single card specifies the geometry of the model; the length of the through
lanes, the ramp entrance and acceleration lane
geometry for both on-ramps, and the deceleration
lane geomet~J and ramp exit for both off-ramps.

42

1.4

A fourth group, containing two cards give data
concerning the a~iting decision process; one card
ror off-ramp No. 1 and the other card for offramp No.2.
T'.ne Problem Card contains a test of
identification number, the lensth of tin~ the
analysis is to be run, the volume of throU£:-:h
traffic and the input volume to each of the two
on-ramps.

Distribution of volume to lanes:

In

all experiments carried out, the traffic volumes

were assigned to the three lanes according to the
follOwing relationships:
Pl = 0.43693 - 0.22183a- + 0.05730cx 2
_ O. 00046 «3
P2

(1)

0.48820 - 0.03136lX + 0.000060(2
+ 0.000240[3

Detailed description of Control Card
forrllO.t is Given in Reference No.3.

(2)

0.07487 + 0.25319 D::' - 0.05736lX 2
+ 0.00382«3

study of Interchar.ge

Confie~ation

Various types of controlled experbnents
can be carried out using this simulation model.
For example, experiments on the effects on traffic
floi-1 of (1) various on-ram3? vehicle volun:es, (2)
various acceleration lane lengths, (3) various velocity distributions, (~) various eeometric confiGUrations, and (5) combinations of the above,
C8..'YJ. be carried out with this model.
EX3?erliaents have been carried out on
the effect on traffic flm! of spacing betvleen an
011-ra:J.1J and an off-rati:J?, under var'Jing traffic
voluaes.

(3)

where
proportion of total volume in the ith lane
DC

total freeway volume in thousands of vehicles per hour

Velocity distributions: Three different velocity distributions were utilized in the
simulations. One velocity distribution was used
for the on-rrunp vehicles, a second for the vehicles in Lane 1 and a third for vehicles in Lane 2
and Lane 3. These velocity distributions are
presented in Table I.
TABLE I

Interchange Configuration
Input Velocity Distributions
Ti-TO interchange configurations (Fig. 3)
,-,ere exanrlned. Each configuration was 200 blocks
or 3,400 ft. long and contained one on- and offramp cOillbination. Each acceleration and deceleration lane viaS 595 ft. long. In Configuration I,
exiting vehicles have 2,465 ft. to travel to the
nose of the off-ramp vThile in Configuration II
-the distance is 3,230 ft. In Configuration I,
the distance bet"leen the acceleration and deceleration lane is 34..-0 ft. and 1,870 ft. in Configuration II. All exitinG vehicles ,{ere deSignated at
block number 199.
Input Data
Volume of traffic: For both confi~ura­
the input to the simulation "lvas for 750
rarap vehicles per hour. For Configuration I experiments ,·lith through-lane volumes of 2,000,
3,000, ~,OOO, 5,000 and 6,000 vehicles per hour
\·;rere conducted. For Configuration II experil"!lents
i-lith throuen-lane volumes of 1,000, 2,.000, 3,000,
4,000, 5,000 and 6,000 were conducted. T"lO tests
I'Tere conducted at each volume.
tior~s,

Vehicle
Velocity
(blocks! sec. )
l.50
2.00
2.38
2.81
3.25
3.69
4.13
4.56
5.00
5.44

Cumulative Per Cent
Ramp Lane ~ Lanes 2 and 3
0.026
0.061
0.116
0.314
0.683
0.888
0.979
0.994
0.998
l.000

0.001
0.010
0.040
0.180
0.435
0.737
0.899
0.989
0.999
1.000

0.003
0.015
0.034
0.064
0.138
0.322
0.759
0.970
0.998
l.000

Gap acceptance distributions: Three
different gap acceptance distributions i-Tere utilized in the simulation tests. The distributions
for merging vehicles (both stopped and moving)
are presented in Table II. The gap acceptance
data for vehicles i-leaving between lanes are presented in Table III.

43
1.4
TABLE II

Gap Acceptance for MerginG Vehicles

%of

Size of
Gap
(sec. )

Total
of Moving
Vehicles
Accepting Gap
r~o.

0.00-1.00
1.01-2.00
2.01-3.00
3.01-4.00
4.01-5.00
5.01-6.00
6.01-7.00
7.01-8.00
8.01-9.00
9.01-10.00

7~ of Total
No. of Stopped
Vehicles
Accepting Gap

12.5

0.0
2.8
17.4
36.0
64.9
95.0
100.0
100.0
100.0
100.0

56.8

77.0
95.1
97.0
98.0
100.0
100.0
100.0
100.0
T.4BLE III

Gap Acceptru1ce Distribution
for v,leaving Vehicles
Length of Gap
(w)
(sec.)

Cumulative
Probability of
J:"cceptance

0.00-0.25
0.26-0.50
0.51-0.75
0.76-1.00
1.01-1.25
1.26-1.50
1.51-1. 75
1. 76-2.00

0.00
0.00
0.00
0.00
0.10
0.30
0.60
1.00

Exiting vehicles: For both freeway con10 per cent of the through vehicles
were designated as exiting vehicles. These exiting vehicles were further allocated to the three
lanes according to the follovnng schedule:
fi~urations

1. 90 per cent of exiting vehicles in
Lane 1 at Block No. 199j
2. 9 per cent of exiting vehicles in
Lane 2 at Block No. 199j and
3. 1 per cent of exiting vehicles in
Lane 3 at Block No. 199.

Experirllental Results
The controlled experiments described in
the previous section I'lere carried out on a 704
digital computer. The ratio of computer time to
real "time varied from. 1 to 5 for low traffic volumes, to almost 1 to 1 for the higher volumes.

The over-all results of the experiments
indicated that there ,,,ere no significant effects
on the traffic flow patterns of the freeway as a
result of this change in geometriC configuration.
A comparison between the through vehicle traverse
times for Configurations I and II are shown in
Fig. 4. The effect of increased vol~e on the
number of ,\-Teaves between lanes is sho\ID in Fig.
5. The computer output also indicated that the
number of vehicles stopping on the acceleration
l~~e increased with increased voluraes of traffic,
FiG. 6.
Conclusions and

Reco~nendations

This study has shO\ID that digital simulation can be useci to faithfully duplicate actual
traffic flow in an on- off-ramp area of a freeway. The output of the Simulation, furthenllore,
gives measures of effectiveness Iv-hich can be used
to evaluate alternate highway designs.
The simQlation experiments performed
indicate the need for research and experL~enta­
tion in a wide -variety of areas to answer questions such as the following:
1. What is the effect of various vehicle distributions to lanes on the traffic flow?
2. ~'lhat is the effect on traffic flow
of various distances betw'een adjoining on-ramps'.
3. What is the effect of various desired velocity distributions on traffic flow':
4. At what volurJ.e of traffic do weaving movements between lanes become hazardous:'
5. ~'lhat is the effect on traffic flow
of various volumes of commercial vehicles?

The simulation model developed in this
study can serve as an efficient tool to answer
these and other problems in the continuous quest
for means of moving traffic safely and efficiently.
Acknowledgement
The authors would like to acknm"ledge
the valuable "lork contributed by lvir. L. Findley
in programing the computer.

44
1.4
Bibliography
1.

Chandler, Herman and Montroll, "Traffic Dynamics; Studies in Car Following," Operations Research Journal, J:.'larch, April, 1958.

2.

Gerloug1l, D. L., "Simulation of Freeway Traffic by an Electronic Computer," Proceedings
of the Highvmy Research Board, Vol. 35,
1956.

3.

Glickstein, A., Findley, L. D., and Levy,
S. L., "A Study of the Application of Computer Simulation Techniques to Interchange
Design Problems. Paper presented at 40th
Heeting of Highway Research Board, January
9-13, 1960.

4.

Goode, H. H., and True, If. C., "Vehicular
Traffic Intersections, II paper presented a.t
the 13th National Meeting of the Associat
tion for Computing Machinery, June 11-13,
1958.

5.

Goode, H. H., "The Application of a HighSpeed Computer to the Definition and Solution of the Vehicular Traffic Problem, II
Operations Research Journal, December, 1957.

6.

Greenberg, H., "Analysis of Traffic Flm'l,1f
Operations Research Journal, January,
February, 1959.

7.

Haight, F. A., "Towards a Unified Theory of
Road Traffic," Operations Research Journal,
November, December, 1958.

8.

Harling, J., If Simulation Techniques in Operations Research,1I Operations Research Journal, May, June, 1958.

9.

Herman, Montroll, Potts, and Rothery, "Traffic Dynamics: Analysis of Stability in Car
Following,1I Operations Research Journal,
January, February, 1959.

10.

Lighthill and Hhitham, "On Kinematic Weaves,
I and II,II Proceedings - Royal Society of
London, May, 1955.

11.

Newell, G. F., lI£<1athematical Models for
Freely FlovTing Highvmy Traffic, II Brown
UniverSity, December, 1954.

:2.

Newell, G. F., and Bick, J. H., "A Continuum
l,iodel for Traffic Flow on an Undivided
Highway, II Division of Applied Mathematics,
Brown UniverSity, July, 1958.

13.

Newell, G. F., "Statistical iillalysis of the
Flow of Highway Traffic Through a Signalized Intersection," Q,uarterly of Applied
Mathematics, January, 1956.

14.

Perchonok, P., and Levy, S. L., "Application
of Digital Simulation to Freeway ON-RAMP
Traffic Operations," Proceedings of the
Highway ResaRrch Board, Vol. 39, 1960.

15.

Pipes, L.
II Operational AnalYSis of Traf1'ic Dynamics," Journal of Applied Physics,
1.1arch, 1953.

16.

Tanner, J. C., II A Problem of Interference
Betvleen Two Queues, It Biometrika, June,
1953.

17.

Trautman, D. L., Davis, H., Heilfron, J.,
Er-Chun Ho, Iviathewson, J. H., and Rosenbloom, A., "Analysis and Simulation of
Vehicular Traffic Flow,1I Institute of
Transportation and Traffic Engineering,
University of California, Research Report
No. 20, 1954.

18.

Tse-Sun Chm'l, 1I0perations Analysis of a
Traffic Dynamics Problem,1I Operations
Research Journal, November, December, 1958.

19.

HonG, S. Y., ItTraffic Simulation ''lith a
Digital Computer, It Proceedings - Hestern
Joint Computer Conference, February, 1956.

A.,

BLOCK
BLOCk
BLOCK
BLOCK
BLOCK
BLOCK
LANE
LANE
LANE
LANE

3999
3998
2999
2998
1999
1998

BLOCK 3000

-----'I

o
BLOCK 1000

3
2
I
5

WHERE A- RAMP INPUT LOCATION
B - NOSE OF ON-RAMP
C - END OF ACCELERATION LANE
o - BEGINNING OF DECELERATION LANE
E - NOSE OF OFF-RAMP
F - OFF-RAMP OUTPUT LOCATION
I - THROUGH LANE INPUT
o - THROUGH LANE OUTPUT

study Area Matrix for Computer Simulation of an Interchange Area

I-'~
•
~

(Jl

.......

~

~O'l

START

STOP--------~~~~~~
__------------~~~~~~~
IF NO MORE CARDS
LI~RE~A~D~DA~trA~C~A~R~D~S~I--------------------------------------------~

DATA

DETERMINE

CLEAR OUTPUT
DATA REGISTERS

MOVE ALL CARS AHEAD BY ONE TIME INCREMENT
AND PERFORM NECESSARY WEAVE OPERATIONS

i

QUEUE DATA

YES

Flow Diagram of Over-all Computer Logic

47

1.4

CONFIGURATION I
AB

1

;;H
70N

I
BLOCK
1 = 200
A = 150
B • 1~5
C • 110
o • 90
E = 55
F • 50
o = 00

C

EF

0

~ ~
fIJ

TRAFFIC FLOW

0

~
~

I

DISTANCE
3~0 FT
2550 FT
2~65 FT
1870 FT
1530 FT
935 FT
850 FT
000 FT

CONFIGURATION II
lAB

~
I:

0-=

0

EFO

~

~

~

,..
TRAFFIC FLOW

BLOCK
200
A= 195
B- 190
C- 155

r

C

~5

E = 10
F :r 5
o • 00

DISTANCE
3~00 FT
3315 FT
3230 FT
2635 FT
765 FT
170 FT
85 FT
00 FT

The Configuration of the Two Sections of Freeway Being Examined

...... ,t:.

• co
,t:.

LANE I
VOLUME 2000 V.P.H.

(I)

(I)

w

..J

LANE 3
VOLUME 2000 Y.P.H.

ex
0

w

100

100

:::E

~

:z:
w
>
C!l

90

80

:z:
z:

70

0

~

0
LU

t

CONFIG. I~/

90

80

I

~."!

CONFIG.

CONFIG.

70

IT

60

60

50

50

~O

IJ.()

CONFIG. IT

(I)

C!l

z:

(I)

ex
>
«
ex
LU

~

(I)

30

LU

..J
0

:I:

20

W

>

LL

10

0
~~

90

TIME IN SECONDS

10

20

30

~O

60

70

80 90

TIME IN SECONDS

Cumulative Distribution of Through Vehicle Traverse Times at a Freeway Volume of 2,000 V.P.R.

49

1.4

y

CONFIGURATION I

180
160
114-0
en

LIJ

>


<

120

LIJ
~

u..

100

0

0

:z:

80
60
14-0
20
L-----~

_______ L_ _ _ _ _ _

2

~

3

_ _ _ _~_ _ _ _ _ __ L_ _ _ _ _ _~_ _ _ _ _ _ _ _~x

5

6

VOLUME (THOUSANDS OF VEHICLES PER HR.)
The Number of Weaves in 5 min. on a 3,400-ft. Section of
Freeway as a Function of Total Through Volume

50

1.4

y

CONFIGURATION I
RAMP INPUT 750 V.P.H.

UPPER LIMIT

~

z:

:: 35

o

~

(I)
(I)

L&J

...J

~
:c

25

L&J

o

> 20

o

A.

%

cc
0::

o"~

15

y=-

-8.3100 + 0.0059X

10
5
L-----~----~------~--~~------~----~----------~x

2

~

3

5

6

VOLUME (THOUSANDS OF VEHICLES PER HR.)

y.--------------------------------------------~-----------.
CONFIGURATION n
RAMP INPUT 750 V.P.H.

~5

UPPER LIMIT

~ ~O
z:

:: 35 .
o

t-

(I)
(I)

30

L&J

...J

~ 25

L&J

>

y =-

A.

~
0::

o. 295~

+ O.0035X

"15
o
~

10
5

LOWER LIMIT
L------L----~--

2

____

~

3

__

~~

______L __ _ _ __ L_ _ _ _ _ _ _ _ _ _
5

6

VOLUME (THOUSANDS OF VEHICLES PER HR.)
The Relationship Between the Per Cent of Vehicles Stopping
on Acceleration Lane (Y) and the Total Through Volume (X)

~x

51
1.5
THE USE OF MANNED SIMULATION IN THE DESIGN
OF AN OPERATIONAL CONTROL SYSTEM
M. A. Geisler
W. A. Steger
Logistics Department
The RAND Corporation
Santa Monica, Calif.
Summary
This paper describes the general features of
the planning and operations phases of a new weapon
system. The uncertainties inevitable in planning
mean that considerable effort is made during the
operations phase to adjust the weapon system and
its resources to the actual environment it finds
so as to attain the desired level of operational
capability. The adjustment mechanism is called
an operational control system in this paper.
Elements of such an operational control system are
described.
The proposal is made that a better control
system can be designed if simulation is used to
help design it during the planning phase. The
use of simulation will not only produce a better
control system earlier but it will permit the
planners to adjust the other resources provided
for the weapon system so that they are compatible
with the environment and the control system. An
example of such a study is described in this paper.
I.

Introduction

The military structure is going through tremendous technological change. Successive generations of new weapon systems are coming along at
a fast rate, and each contains such different
characteristics that in many ways past experience
is of little use. Furthermore, the competition
among nations in military technology is causing
efforts to reduce the time .length of the defense
planning cycle, which complicates the necessary
process of coordination and interaction that is
required to produce consistent and compatible
hardware, operational plans, and support plans.
These two factors, compreSSion in planning and
rapid technological change, mean that the planning
process is becoming more difficult and its results
therefore less dependable. The purpose of this
paper is to suggest that the design of better
operational control systems is-one-way of contending with this planning diffICUltY; and further,
that the design of such systems will be considerably aided by the use of simulation techniques.
In developing this thesis, this paper covers first
a discussion of such a system, the use of simulation for designing a specific system, an example
of such a designed system, how the simulation of
the operational process was used to improve the
planning decisions, the role of computers in the
simulation and finally, some of the limitations
of the technique.

II.

Need for an Operational Control System

It takes a long time to develop, procure,
and install modern weapons that are needed to
satisfy a future operational requirement. In
Chart 1, we try to define in general terms the
process that leads to the creation of such systems.
As can be seen, there are two phases: a planning
phase and an operational phase. Tne planning for
new weapon systems takes place over several years.
It usually begins with a formulation of the future
environment for which a new weapon is to be developed. This environment may describe the operational Situation, or requirements foreseen, and
possible enemy intentions and capability. This
environment is then used as a guide to define the
hardware characteristics of the projected weapon
system, the operational plans for use of the
weapon, and the logistiCS or support plan of the
weapon for supporting its operation. This planning involves many organizations, many people,
and as we have said, takes a long time, perhaps
several years, because it is an involved process.
Necessarily, many of the factors developed
for planning use are subject to much uncertainty,
and it is very difficult to define all the relevant relationships affecting both the operating
and support characteristics and functions of the
weapon system. Nevertheless, these plans are used
to determine resource and budget estimates for the
new weapon system. These estimates then go
through a budget and procurement cycle which takes
additional years. We thus see that the planning
for new weapon systems involves many organizations,
and requires many assumptions about the uncertain
future, technical capabilities, and enemy intentions.
The operations phase begins 'when the resources
in the form of weapons, people, equipment, etc.,
resulting from the planning cycle are produced.
These resources are now confronted with the actual
environment. For many reasons, there are many
difficulties involved in fitting the actual environment and existing resources together to achieve
a satisfactory level of operational capability.
For one thing, the predicted environment and the
actual environment differ Simply because of predictive errors. The enemy capabili~ies are now
different from what we expected; the operational
situation is either more or less demanding than
that predicted; furthermore, the actual resources
. may differ appreciably both in quality and quantity
from those planned. The budget cycle provided
funds for less resources than expected; the

52
1.5

reliability of the weapon system is not as great
as projected; and the cost of the weapons and
other resources are greater than planned.
Thus, the major activity during the operations phase is to mesh and adjust the actual
environment and the existing resources so that the
resulting operational capability is as great as
possible. The system performing this meshing and
adjusting is called an operational control system.
It performs this activity, first, by using ~
resources available at anyone time to obtain as
high a level of operational capability as possible, and second, by specifying the adjustments in
future resources that will improve operational
capability. Thus, the operational control system
enters directly into the determination of both
operational capability and efficient resource
determination and utilization. In this respect,
the control system can be treated as another
resource of the total system; and therefore, its
contribution to operational capability and its
cost can be weighed against that of all other
resources.
With this in mind, there are three main
points we should like to make about an operational
control system. First, it is an integral part of
the process that determines operational capability,
and its use is an important means for contending
with planning uncertainty. Second, it is'a complex system in its own right, and its creation
may take a long time and much resources. Third,
if one has a means of predicting the operational capability, this information can be used
during the planning cycle to adjust resource requirements. A more responsive, accurate, and comprehensive control system may require less total
resources to achieve a given level of operational
capability.
It is the premise of this paper that simulation early in the life of a weapon system can
predict (to some degree) the contribution of the
operational control system to the preferred mix of
resources (including the control system) for the
operational weapon. If the composition of this
preferred mix can be known during the planning
process, such information can influence resource
policies and future reqUirements.

operational control system. At any given point in
time, the status of the environment and the existing resources represent what we call the current
system status. These include such data as the
operational condition of the weapon, work assignments of personnel, equipment location and uses,
etc. The operational control system works on this
status to improve the future capability of the
weapons system. Since the actual status may be a
complex of vast geographic distances, thousands of
personnel, many million dollars of weapons, equipment, and faCilities, the control system creates
an abstract representation to help it to Inanipulate and to improve the actual system.
This abstract representation is obtained
through an information system which receives inputs as changes in the current system status occur,
such as weapon malfunctions, personnel availability, etc. "These inputs are then put through a
data-processing system which contains a series of
rules and procedures for manipulating the resulting
data, using appropriate computing equipment to
produce relevant outputs. These outputs are given
in the form of reports or displays to appropriate
managers in the control system. These may show
weapon statuses, personnel shortages, equipment
downtime, etc. Thus, a major function of the
information system is to boil down or transform
the very complex real system to a few elements
that can be grasped and evaluated by a management
group or organization. This evaluation leads to
decisions which then result in management actions
that change the future system status in a way that
is intended to increase operational capability.
Thus, the control system consists primarily of a
set of poliCies or decision rules for achieving
increasing operational capability, some of which
are automated as part of the information system,
an organization for decision making and taking
management actions, and an information system.
IV. Difficulties in Designing
Control Systems for the Future

The control systems that are being developed
or conceived by the Air Force to enable major
commands to control lower echelons have characteristics which far transcend the hardware elements that such systems must necessarily possess.
They are man-machine systems, and although the requirements and availability of hardware to be used
in such systems is obviously very relevant, the
demands upon the man in the system are at least as
important to determine.

In other words, an essential requirement in
the design of control systems is the need to define
the role of man and his relationship to the hardware. The difficulty in doing this arises from
several circumstances. First, the control systems
require several years to develop and produce, the
most ambitious taking from 5 to 10 years. It is
therefore necessary for the deSigner to project
his requirements at least that far in advance.
With the rapidly changing military technology, the
environment within which the control systems must
operate will be radically new and will create conditions that are very different from those existing
today. Much higher alert requirements are likely,
with a further need for fast response to critical
Situations, under conditions of wide dispersion.
Radical weapons, including missiles, possibly
space vehicles, and nuclear powered vehicles are
in the offing.

Chart 2 helps to describe a simplified version of the elements and functions of such an

Not only will the environment be very different, but because of the potentialities of new

III.

General Description of an Operational
Control System

53

1.5
communicating hardware, such as data processing
and communications equipment, and the developing
research in organization and information theory,
the designer finds it very difficult to project
himself readily into this highly complex and different situation in order to take maximum advantage of these new factors in the design of the
system. The relationships involved in these complex man-machine systems are often far too involved for intuitive or analytical treatment alone.
Although we can be sure that the environment
of the next 5 to 10 years will be very different,
we do not know the exact specifications of it.
There are many uncertainties in the rate at which
new weapons will be developed, both of an offensive and defensive nature, and the size and kind
of enemy threat. These conditions as well as
many others, can differ enough so that the demands
on the men and machine in the control systems can
only be estimated with uncertainty. However, the
designer needs to know in detail the impact of
this uncertainty upon the specifications of the
control system so that the hardware manufacturers,
the programmers, the training activities, etc. can
carry out their assignments in a timely, coordinated, and comprehensive way. Thus, the core
problem in the design of control systems is planning complex man-machine systems, requiring several
years to produce, under conditions of uncertainty.
In the remaining portion of this paper, we will
discuss some of the general aspects of the design
of control systems, and how simulation, as used
in RAND's Logistics Systems Laboratory, can help
in treating this core problem.
V.

Use of Simulation

The definition of the elements of a control
system in some detail produces the design of a
feasible and compatible operational control system.
The creation of this design is no simple task
because it requires the deSigner to specify with
some precision just how the decision maker in the
operating system will act to increase his operational capability. The deSigner, therefore, has
to visualize how the operating system would work.
If he had an operating system to study, he could
obtain such a visual aid, but the one he is designing will not exist for several years, and he cannot wait. Typically, in the past this lack of an
operating system to study and manipulate has led
to inappropriate, incomplete, and infeasible
system designs. The gaps and defects were then
left to actual operations to correct. This is
both an inefficient and lengthy procedure which
results in delays and difficulties in obtaining
the maximum capability with the weapon system.
The point of this paper is that if the operating
systerncanbeSimuIa'ted in appropriate detail and
under reaI'fstic conditions, the simulatIOii"Can-help fill the void at least partially, and perhaps,
ther'e'fO're, "1ielpprodUce better designs with resulting less-stresses for the real world adjustment.
--- -- - - --- ---In recent work at the RAND Logistics Systems

Laboratory, efforts were made to use simUlation to
help in this planning process and to design an
operational control system. We reasoned that if
we could simulate the environment of a future
weapon system in sufficient detail we could then
take the proposed hardware characteristics, operating plans, and support plans and determine their
mutual compatibility, and, furthermore, that this
would help us to estimate the operational capability they would produce for the weapon system.
We used simulation because no real-world opera~ng
system existed, and yet we wanted to study an operating system early enough in the planning process
to be able to help evaluate and possibly influence
the proposed plans. Since we realized that our
simulation of a future environment was subject to
uncertainty and error, we varied it over a range
of foreseeable conditions and hoped, in this way,
to allow for our ignorance of the future. Second,
we exposed the policies of the system to the
range of simulated enVironments, and in this way
determined their compatibility and overall performance. Third, we also used the simulated environment to help specify the details of the control
system design, and by using this control system
in conjunction with the other policies, we were
able to determine how the particular control system which we employed modified the other poliCies,
and therefore affected the performance effectiveness and resource requirements of the total system.
We believe that this study produced many useful results, both in the specific context of the
weapon system we used, and in the techniques that
eVOlved, which we believe have an application to
many other such weapon systems, since each of
these weapons is characterized by the planning
difficulties we have described. We should like,
therefore, in the remainder of this paper to
describe the simulation experiment that helped to
produce a design and an evaluation of an operational control system. This experiment is known
as LP-II, and one of its primary objectives was
the design and evaluation of a control system for
an ICBM weapon system of the 1963-1965 time period.
The control system that was developed is for a
tactical unit, such as a squadron or wing, for two
reasons: the ultimate effectiveness of the weapon system depends on performance of this unit,
and second, the bulk of the resources and cost
of the weapon system are also in the tactical
unit.
VI.

Description of Simulated System

Chart 3 contains a schematic diagram of the
simulated system. The first requirement was to
define the hardware of the weapon system under
study. Engineering study and analysis gave us
estimates of the missile configuration likely in
the 1963-1965 time period. The long time required
to convert hardware specifications into field
eqUipment gave us some assurance that we had captured the essence of the missile hardware, and
further, that for practical purposes it would be
most realistic to assume that this part of the
system was basically fixed. We found that we had

54

1.5
to describe the hardware in what we called "Support Unit" terms in order to capture the hardware characteristics in sufficient detail to permit interaction of the operations and support
activities in the tactical unit. A support unit
was defined as a module, black box, or a functionally integrated component. For example, the
checkout and malfunction detection equipment was
basically designed to detect failures within
support units, so that the support unit normally
would be that module or component to be removed
an& replaced on the missile to correct a malfunction. To describe the missile, its ground
support (launching and monitoring) equipment, and
launch facilities in this level of detail, we
found took 1,500 support units: 800 on the missile, 250 on the ground support equipment, and
450 in the facilities.
To represent adequately the operationssupport interactions of this tactical unit in a
minute-to-minute way, we found required about a
hundred different parts characteristics for each
support unit. These characteristics covered such
elements as the skills and equipment needed to
perform remove-and-replace activities; missile
system checkouts; diagnosis of an ambiguous malfunction as reported by the checkout equipment;
calibrations, and so forth. A very significant
characteristic of each support unit ,was its reliability. Since this is one of the most uncertain
parameters, its value was represented by a range
from minimum-acceptable to maximum-likely reliability. In the course of the experiment, we
examined the effect upon the control system of
varying the reliability over this range.
A "failure II model had to be constructed which
would be consistent with the reliability values,
and which would define for each support unit the
probability that it would fail under a given type
of stress. We defined six types of stresses
likely to cause failure for each of the 1,500
support units. The operational and maintenance
statuses were defined for each of the 40 conditions in which the missile system could be placed:
various degrees of alert, launching, countdown,
checkout, replacement, etc. The description of
each status defined the stresses that applied to
each of the 1,500 support units by 15-minute intervals of time. Therefore, given a particular missile system status, the probability of failure for
each support unit could be established for each
15-minute interval, and by making random draws
from a probability distribution, a malfunction
generation model was created. This model generated
a significant portion of the workload for the
squadron (identified by the dashed line in Chart 3).
The squadron contained several launch complexes and a squadron headquarters. A full
squadron complement would normally contain several
hundred people, but for the purposes of our study
we were interested primarily in the management
personnel. We identified the launch complex
commander for each launch complex and the operations, maintenance, and supply officers of the
squadron headquarters as being the key management
personnel. Therefore, they were the "people"

of the simulated squadron, and ~ersonnel to staff
these six positions were brought to the laboratory
from Strategic Air Command missile organizations.
This squadron organization was given a set of
operational requirements, policies and resources
by the embedded organizations who represented
higher headquarters. The embedded organizations
were staffed by laboratory personnel. These embedded policies specified the operations, maintenance, supply, manning, and data-processing
reqUirements to be met by the squadron. The
resources given to them were initially derived
from Air Force manning documents, equipping documents, supply tables, etc., that were in effect
at the start of the experiment. These resources
were represented by punch cards, which could be
assigned to jobs in the performance of operational
or support activities by the managers of the
squadron.
The squadron managers also had access to a
data-processing and analysis center. One function
of this center was to maintain the maintenance and
supply-status data of the squadron, and to perform
various logistical functions, such as automatic
resupply notifications, required engineering
changes, maintenance personnel and eqUipment
availability and utilization, etc.
The experiment was operated by glvlng the
squadron a set of operational requirements to meet,
such as maintaining a specified amount of target
coverage and a number of missiles on alert. The
squadron was also given certain preventive maintenance and engineering tasks to perform. As the
squadron managers tried to satisfy these operational and logistical requirements, they had to
schedule the missiles to be in various situations
or statuses. These statuses in turn caused
stresses to be applied to the support units of the
missile system, which caused malfunctions. The
squadron managers then used the resources of the
squadron to correct these malfunctions; and,
throughout the entire exercise, the squadron tried
to schedule its activities with the given resources
so that maximum operational capability was attained.
As the experiment proceeded through several
runs,* the operational and support policies were
changed, the assumed reliabilities were varied,
the resources provided were modified, and the
organizational structure was altered. At the
same time, the decisions made by the managers
were studied, both by observation of the results
of their decisions and by interviewers. Such
analysis produced inSights into the nature of the
decision processes, and suggested possibly preferred decision rules. For some of these decision
*Each run lasted one simulated week. He
found that such a time period was sufficient to
permit most of the stochastic elements in the
system to "settle down." A simulated week took
about one week of real time, but time compression
was achieved since we simulated the 24-hour days,
7-day weekS, during a real working week of about
35 hours.

55
1.5
processes, we used mathematical models or heuristic
programs to develop the decision rules. The use
of such rules in turn suggested required information and organizational arrangements for making
decisions. This led to an evolving operational
control system, which is the result of experience
by Air Force people in the simulated environment
plus analysis of the results obtained by various
techniques tried during the experiment.
VII.

Resulting Operational Control System

Referring to Chart 2, the control process
illustrated is the classical one of information
and feedback.* The novel problem in LP-II was to
adapt this process to the military environment
posed by ballistic missiles in the 1963-1965 time
period. We can foresee a situation in which
missiles are dispersed over a wide geographic area,
maintained on a high level of alert, and responsive to sudden emergencies. The cost of these
missile systems is very great, and so every effort
will be made to keep the resources and costs they
need to do their job to the minimum. One means of
achieving this goal is to be able to shift resources (within the constraints of possible sudden
war) from one place to another responsively, since
the demands for resources at anyone place are
typically sporadic and loW. Further, the support
costs in equipment and facilities are so large a
fraction of total system cost that minimum missile
system downtime is desirable not only for military
reasons, but also for economics, since support
resource levels are almost fixed, in the short
run, for a missile base. Since the cost of a
weapon down is so great (if we value it at system
cost per time period), the major problem for a
tactical unit is to use its fixed resources to
maximize alert.
ThiS, then, is the environment in which the
simulated control system had to function. The
scheduling rules assigning missiles to various
degrees of alert, and assigning the resources to
repair missiles requiring preventive or other
prescribed maintenance were found to be the critical decisions in the squadron. The formulation of
precise scheduling rules by mathematical analysis
was very difficult because of the many combinations of conditions and constraints that had to be
considered. In effect, the implications of the
schedule had to be determined by support units, and
there were over 50,000 such support units on the
operational missile system in a tactical unit.
Further, the schedule was affected, to some extent,
each 15 minutes as a missile or a resource changed
status.
The useful scheduling rules have been developed by trying many rules of thumb during the experiment. The rules that seem to work best are
*RM-213l, Management Information for the
Maintenance and Operation of the Strategic Missile
Force, D. Stoller and R. Van Horn, The RAND Corporation, April 30, 1958, describes the requirements of management and information systems for a
missile environment.

called "opportunistic scheduling." Under this
rule, all prescribed maintenance is deferred as
long as the policies permit. Then, when a missile
malfunctions, all possible prescribed maintenance
as can be done concurrently is accomplished at
that time. In this way, missiles are taken off
alert as little as possible for prescribed maintenance.
Clearly, such a rule requires a highly
responsive organization and information system
because it must be ready to react instantaneously
to any missile malfunction, change in resource
status, etc., under conditions of wide dispersion.
The structure of the information system that
has been designed to meet this requirement of
unscheduled and rapid responsiveness contains the
following: the current status of each missile
and resource, a complete listing of all missile
and maintenance situations and the resources required to accomplish them, and a listing of all
prescribed maintenance awaiting performance on
each miSSile. As a miSSile or resource changes
status, the information system is updated through
appropriate inputs. As a new situation occurs,
each of the managers interrogates the information
system for the implications it has for his function (SUCh as operations, supply, or maintenance),
revises his schedule accordingly, and assigns
available resources to the urgent activities.
During the experiment, the information system
was able to respond instantaneously to such interrogations. This was made feasible by plaCing all
the necessary information in a very large randomaccess memory that could be interrogated very
quickly. Since the system status changed very
rapidly, an input system had to be devised which
would quickly update the status, and this further
required the minimizing of input data. This was
done by requiring only a single input of each
necessary data element, which served all managers.
Since the organization had to respond very
quickly to emergencies, a good deal of functional
integration was required. Operations, maintenance,
and supply in the squadron headquarters had to
work as a close team since each was affected by
such emergencies. In addition, close communication
between the launch complexes and the squadron
headquarters had to be maintained. Report formats
evolved which helped to present to each manager
the consequences of another function's decisions
on him, and the consequences of his decisions on
them. In this way, each manager was able to react
rapidly to dec~sions that might affect his plans.
Further, the squadron headquarters found it
could not centrally handle all of the deciSions,
since many of them were localized to a particular
launch complex. It therefore worked out conditions
under which the launch complex could act without
prior approval of the squadron headquarters, and
when such prior approval would be necessary. These
conditions have also been further developed by the
laboratory staff into rules of thumb, since the
relationships appear to be too complicated for
analytic solution. Such heuriaU-e techniques welle

56

1.5
found to be very useful during the experiment.
'Ihe data produced by the squadron managers in
the course of their minute-to-minute operational
decision making were also used by them for longerrun or planning decisions. 'Ihus, these data were
used to estimate resource demands and utilization
which provided the basis for determining required
manning, equipping, and stockage of the launch
complexes and squadron echelons. 'Ihese data reflected the particular decision rules used to
schedule the employment of the resources, and so
there was a direct relationship between the shortrun operational decision rules, like scheduling,
and the longer-run planning deciSions, like determination of the resource requirements of the
squadron.
Thus, the simulation helped to produce and
evaluate a control system that might fit the
peculiar environment of the 1963-1965 missile era.
The nature of the deciSions, information system,
and organization for such a control system has
been developed in considerable detail, and as such
it may provide a good guide to the design of such
a system for the real world. Further, this blueprint is available early enough to permit its
appearance in tactical units by 1963. We cannot
say that this is an optimal control system, but
it appears to be ~easible and compatible with the
type of environment experienced in the laboratory.
Simulation thus has helped produce a reasonably
detailed design of a control system for a future
radically different environment.
VIII.

Computers and the LP-II Experience

'Ihere are two aspects to computers and a
simUlation like LP-II: first, computers are used
in the modelling process and runs; and, secondly,
we learn something about the computer needs of the
management system under study.
Only a few words need be devoted to the first
aspect, here. In addition to standard punch card
equipment, two general purpose digital computer
3ystems were used: one with magnetic tapes capable of microsecond arithmatic speeds (in our case,
an IBM 704 system) and a data-processing machine
with relatively slow processing speeds but with a
large disk memory capable of rapid access to any
of several million characters (in our case, an IBM
305 RAMAC). In the laboratory, we have, occasionally, used the high-speed computer more or less
"on-line" with the simulation -- men making decisions which are fed into the computer which in
turn quickly "plays the decisions" through various
models and returns to the relevant men the system
impacts of these deciSions, which in turn causes
new decisions, etc. But in LP-II, most of our
operations on the high speed computer were not
"on-line". 'Ihey were either for analysiS purposes or for computing lists of "potential failure
'rates" which were then interpreted by laboratory
personnel to reflect the minute-to-minute decisions
of the laboratory partiCipants (see the description of the failure model above).
As a

result, the entire project consumed only a few
hundred IBM 704 hours, a relatively small portion
of the total cost of the operation, which also
required more than 70 man-years. In other manned
simulations where the participants and the highspeed computer "talked" to one anather many times
daily, the computer costs were a considerably
larger proportion of the project's total costs. l
The computer with the large memory (the RAMAC)
on the other hand, was used primarily on-line,
minute to-minute, during the runs, and-Served
mainly2 as an information storage and interrogation device for the laboratory participants.
Study of this computers' programs and use permitted
us the statements we could make about the kinds of
computers such a tactical unit might need if it
were to obtain the same effectiveness as our simulated unit: The RAMAC operation was basically a
"status of resources" system and served to provide
the resource manager with "up-to-the-minute" data
needed to make decisions on the assignment of
these resources. Hence, all of its internal files
were "status files" in the areas of maintenance,
supply and operations. 'Ihe inputs to the system
were changes in status (personnel movements, maintenance actions, etc.). 'Ihe outputs were status
reports (generated in response to the manager's
inquiries) and a comprehensive set of history
cards. The history cards served as a historical
file, and were also used to generate certain summary and history reports. 'Ihe system had two
basic components; the disc files which were the
information acted upon, and the stored programs
which did the acting. 3
The system contained both t'on-line" and "offline" programs. Off-line programs operated independently of each other and of any central control.
They were used to perform functions of an occaSional, isolated nature such as loading the initial
files. Tb use an off-line program, it was necessary to interrupt the normal 305 operative mode.
On-line programs were all controlled by a Master
Routine. These were the programs which processed
the normal input cards. All on-line programs (and
the Master Routine) were stored in the disc files.
The Master Routine was normally on the drum and
operating, and the processing routines were read in
and operated as needed. Several of the on-line
programs used subroutines. 'Ihe subroutines were
stored on the disc files and were read in and controlled by the programs which used them.
IWe have been able to do this several times
during the course of an 8-hour day, which simulated
several daily or weekly "time periods tl •
2It was, also, invaluable for analysis purposes, later, following the runs.
3The remainder of this section is largely due
to the efforts of J. Tupac and K. Labiner. See
their RAND Research Memorandum, RM-2572, The LP-II
Data ProceSSing System, September, 1960.

57

1.5
The LP-II data system did not unduly tax the
machine's stor~ge capacity, uSing less than 90%
at peak usage.
However, the machine's slow processing and output speeds did influence the system's performance considerably, and suggested the
infeasibility of dOing all that had been hoped for
with little or no information lag. To quote from
the report on the LP-II data system: 2
'Slow processing and output speeds strongly
influenced LP-II's system design. As mentioned,
almost all machine time was used in processing the
"on-line" status changes and interrogations. As a·
result, we observed three consequences. First, the
machine did not generate all the summary reports.
Had we used a faster machine, all summary reports
could have been generated in step with its other
operations and thus furnished managers with some
summary data on an inquiry basis. Greater directprinting capacity and speed would also have been
required. Secondly, the machine did not make
decisions. We found it was difficult to define
decision rules when the system was being designed.
In addition, a machine with much higher internal
processing speed and computing capability than
that available for the experiment would be necessary in order to incorporate decision functions.
This would have been true even if the rules could
have been established at the outset. Finally,
there was not as much internal checking of data as
one would have liked."
To summarize the major outputs of LP-II with
respect to control system design are as follows:
1. An integrated control system for a
large missile tactical organization, combining
the organization, policies, data system, with the
novel operational and weapon system environment of
future years.
2. Estimates of workloads, resource
utilization, delays, missile alert perfonnance,
etc., resulting from the interactions of environment, policies and decision-making of the organization.

3. Frequency of, and kinds of techniques used by the management personnel for making
critical decisions, such as targeting, scheduling
of mis'sile status; and resource assignment: af',
tailored to the special demands of the large tactical organizations.

4. The information requirements for
making critical decisions and p~rforming management activity. The reports requested by the
partici9ants changed considerably as the organization was exposed to different Situations, SUCh
as higher alerts, engineering changes, reduced
resources and revised requirements thus generated
provided an explicit estimate of the data displays
IThis was partially due to the scale used
in the simulation models, however.

2RM -2572, op.

cit.,

PP.37-38.

that the new organization may require.

5. A detailed description of the data
processing system for providing the required
information. This system was based upon several
new concepts that imply greater automation, new
methods of reporting data, and a different type of
data organization. These concepts were made operational in the simulation so that many of the specific techniques used may be directly transferable to
a real-world control system. Also, estimates were
obtained for the frequency and urgency with which
different data elements were required. These may
be useful in developing specifications for data
processing hardware.
IX.

Limitations and Future Research

The number of alternatives facing the deSigner of the information portion of a management
control system are enormous, particularly during
the planning stages. Chart 4 gives some idea of
the nwnber and kinds of choices that a system
designer must resolve. For a given budget, he is
expected to choose the right man-machine resource
mix for each of the alternatives mentioned. Anyone who has ever helped design an information
system can supply examples of the kinds of choices
that Chart 4 represent.
We have said, so far, that manned Simulation,
properly combined with all-computer simulation
models, can help the designer make these choices.
LP-II, if nothing else, should increase the designer's belief that this is a possibility. However, lots of problems which remain must be satisfactorily resolved before this becomes part of the
everyday tool-kit of a system design program.
1. The presently high costs of a simulation like LP-II. The sizeable scope, great detail and the use of humans in Simulations are all
factors which make this sort of experiment useable,
primarily, for design~ng larger; complex t1 man machine" systems. Of course, some of these costs
would have to be borne by any Sizeable study of a
control system but others can be reduced by turning to suitable compiler programs, and using only
the degree of detail necessary to obtain the desired results.

2. Experimental DeSign Problems.
Manned simulation experiments cannot produce, for
the same research budget, the same number and
length of runs that their all-computer brethren
can. This makes it difficult to perform the desired amount of sensitivity testing or run long
enough to wipe out) fully, the effects of initial
conditions. Work is proceeding at RAND, and elsewhere, on experimental design techniques to mitigate these effects.

3. Inability to Produce O~timal Designs.
Simulations, by their very nature, produce at best
highly preferred -- but not optimal .. - policies.
Significant betterment in management control design
is, of course, nothing to belittle. The development of all-computer simulation and analytic

58

1.5
models based on the manned simulation as a breadboard model should help, considerably, to produce
even better results.

4. Programming Task Synchronization.
Coordinating the many tasks through time that have
to be performed by the computer portion of the
simulation has, in the past, ordinarily required
setting the basic time unit of the simulation
equal to the minimum time unit of all tasks. The
possibility of using variable time synchronization -- different tasks running at different
time units -- is being investigated. This would
permit running the simulation activity at differential speeds as a function of the question the
particular simulation run is addressing. l
5. Defining a Suitable Level of Scope
and Detail. The system designer might wish the
simulation activity to produce answers to highly
specific questions but this requires the simulation activity to represent great amounts of
detail and many functions and organizations.
This is costly and makes analysis more difficult.
Therefore, we must better understand how to deal
with management control problems in an abstract
way for simulation purposes, yet in a realistic
enough manner to produce valid design factors for
real-world application.

x.

Conclusion

Notwithstanding these difficulties, we nevertheless believe that manned simulation will become a very useful technique for those seeking
to make choices in a system context, between
alternative management control resource mixes.
The problem described in this paper is probably characteristic of many large system-design
problems. Certainly the need for such operational control systems is expanding in the military, and in both civilian and military the need
for improved management systems has been ever
present. Simulation techniques can prove of much
help in these problems by providing a means of
pooling and integrating knowledge from many
sources and by providing the opportunity to iterate and vary the many variables and parameters
that compose such systems. Although most published simulation experiences have involved allmachine models, we have found much value in manmachine simulation when the problems have involved
organizational interactions, the design of information systems, and conflicting or interacting
decision rules, since these undergo considerable
development during the simUlation process.
IJack Little of RAND's Computer Science
Department is working on this problem.

59

1.5

BaZ'dware
~

Characteristics

P
L

ProJected
Future
Envil'Omaent

A

OperatiOD&l
Phase and
Policies

.

~

"1\

Resource

-

PlaD. and

Policies

-

-

-

-

-

-

....

-

System
Requ:1rements

N
N
1
N
G

Budget and

P
H
A

Program

Cycles

-

-

--

II

Actual

Env1romDeD.t

~7

Operational
Control
System

L-

S
E

Available
Resources

1
Operational
Capabillty

Chart 1

PLADIRl-OPERATIONS RELATIONSHIPS FOR A FUTURE WEAPON SYSTEM

o
P
E
R
A
T
I

o
N
S

P
H

A
S
E

60
1.5

Actual
EnvirGnment

f---

;--

Available
Resources

,
r---'=

Current
System
Status

~

.Management

System
Inputs

~

Decisions

Actions

Operational
Capability

Chart 2

ELE1'fJENTS OF AN OPERATIONAL CONrROL SYSTEM

Data
Processing

. Reports
and
Displays

61

1.5

Hardware

Operational

and Parts

Failure
totxlel

Characteristic
Data

end

Maintenance
Problem

I

t

j,

Malfunction
Generator

,

- - -

10-

-

-

\v
)..1

Background

Research
Hardware
Operations
Maintenance
Manning

7

T

1
I

'1"-

~

I

Launch
Complex

7

I

\If

Supply

Data Processing

.r
1

Squadron

.'"

"J

1

Data Processing
and
Ana.l:ysis Center

l;-

I

I"

I
I

_ _ _ _ _ _ -.J

Chart 3

MJDEL OF SIMUIATED WEAPON SYSTEM

Higher
Headquarters

62

1.5

CHART 4

FEATURES OF ASSUMED INFORMATION SYSTEM DEVELOPMENT

INPUTS
l.

2.

3.
4.

DATA PROCESSING STRUCTURE

1.

SPECIFICATION OF
DATA ELEMENTS
LEVEL OF AUTOMATION
MINIMIZE REPETITION
RETAIN BASIC RELATIONSHIP

2.~
11

3·
4.

.~

OUTPUTS
l.

2.

3·
4.
5.

CONTENT FLEXIBILITY
INTEGRATION OF REPORTS
DISPLAY FORMATS
LEVEL OF AUIDMATION
SPECIFICATION FOR
EXCEPTION REPORTING

X

_.,
DATA PROCESSING HARDWARE
l.

2.

.

ON -LINE VERSUS
ANALYSIS NEEDS
INTEGRATION ACROSS
FUNCTIONS, ECHELONS,
TIME AND DISTANCE
ALTERNATIVE PROGRAMMING
TECHNIQUES
MECHANIZATION OF DECISIO~
RULES

3·
4.

DEVELOPMENT AND TESTING
MEMORY CAPACITY
COMPUTING SPEED
INPUT AND OUTPUT
CHARACTERISTICS

63
2.1
A SURVEY OF MICROSYSTEM ELECTRONICS
Peter B. ~ers
Motorola Inc., Semiconductor Products Division
Phoenix, Arizona
Summary
The history of micro system electronics is
briefly traced through the successive stages of:
miniaturization; subminiaturization; microminiaturization; thin-film integrated circuits; semiconductor integrated circuits; and finally morphological integrated circuits or functional blocks.
The term integrated circuit is defined as the
combination, on or within a single chunk of material, of multiple electrical elements to perform
a desired circuit or systems function. Improvement in reliability, decrease in size and weight,
decreased power consumption, and lower cost are
discussed as motivations for Integrated Circuitry.
Fabrication techniques, interconnection, and
accessibility are identified as major problems. A
catalogue of techniques for fabrication of Integrated Circuitry is presented.
What Is Microsystem Electronics?
We define Microsystem Electronics as that
entire body of electronic art which is connected
with or applies to the realization of extremely
small systems. As such it includes power supplies,
input and output transducers, and various forms
of memory where applicable, as well as digital
and analog electronic circuitry.
This paper surveys the techniques of microsystem electronics that apply to the arithmetic
or logical elements of computers and to the electronic circuitry of communications equipment. The
electronics industry has not yet agreed on a name
for this part of micro system electronics but the
term Integrated Circuitry seems to have wide and
growing acceptance as a generic name.
Integrated Circuitry
Definition
We define an integrated circuit as a combination, on or within a single chunk of material,
of a number of basic electrical elements to perform a circuit function.
The concept and terms used can be illustrated
by a simple four-box picture of electronic technology. Let us arbitrarily divide electronic
technology into the following four mutually exclusive and exhaustive categories: Materials,
Components, Circuits, and Systems. If we investigate the emissive properties of a heated tungsten
wire or the dielectric constant of polyethylene
we are clearly studying the properties of
Materials. If we combine a tungsten wire, some
bits of nickel wire and sheet, put it all in a
glass envelope which we then evacuate, we have
made a vacuum tube -- a Component. If we properly
connect our vacuum tube, an Inductor, a capacitor
and a battery, we have an oscillator -- a Circuit.

Finally, if we combine an oscillator, a modulator,
an amplifier, a power supply, and an antenna, we
have a transmitter -- a System.
Note that in the above example we were engaged in four distinctly different types of
activity. In the specialization of today's electronic industry these activities are performed by
four different men. In the example, and more
generally in our four-box picture of electronic
technology, we find communication and interaction
-- bi-lateral feedback -- between each type of
activity and its immediate neighbors. The components man, for example, takes the output of the
materials man, asks him for materials with specific characteristics, and tries to feed back the
results of his experiements in such a way that the
materials man can modify or optimize his output.
The components man also communicates with the
circuits man, listening to his requirements and
providing or designing components to fit. He
receives feedback from the circuit designer as to
how well his components work and how they should
be changed.
Similarly, the circuits man goes to the components man for his building-blocks and to the
systems man about the finished product he hopes to
supply. Each man, in addition to his own problems
and language, must be able to communicate with his
neighbors sufficiently well to exchange the information necessary for his work.
Note, however, that in this simple picture
there is no need for the materials man to talk to
the circuits man or to the systems man, and in
practice he hardly ever does. In fact, they have
so little in common in our conventional electronic
technology that they don't really speak the same
language. Likewise the components man has no real
need to understand the problems of the systems
man, and the latter tends to think of components
only as black boxes with certain failure rates
which limit the reliability of his system.
This partitioning is a result of the need for
specialization in today's technology and of the
finite amount of time and interest the average
technical man has for exploring outside his own
speciality. It is unfortunate in that it leads to
inefficiency and lost opportunity in a growing
number of instances where a new requirement is met
by an adaptation of an old technique instead of by
a fresh examination of the wide range of unexploited possibilities.
Integrated Circuitry removes the internal
partitions and integrates the boxes into a single
field. Integrated Circuitry still has its
specialists, but each must have a working knowledge of the problems and objectives of areas which
yesterday would have been the exclusive domain of
other specialists. For example, the circuits man
may no longer limit his consideration to the black
box characteristics of existing components. He
must be aware of the characteristics of the

64
2.1
materials which go into his components, of the
limitations imposed by the laws of the solid
state, and of the possibilities it opens up. The
components man may no longer accept blindly the
re~uirements set by the circuit designer.
He has
a responsibility to examine the overall re~uire­
ments on the system to find the most satisfactory
way'of accomplishing the desired function. As an
example, a conventional circuit designer might
build a power supply with a transformer, a diode,
a choke, a couple of capacitors -- or a simple
RC filter, if you prefer -- in either case he has
used a minimum of five components. Our integrated
circuit designer might choose to build the same
power supply from a single piece of silicon.
Alternating current flowing through one part of
the silicon encounters resistance and generates
heat which filters through the silicon to a
thermoelectric area, where Seebeck Effect produces
filtered dixect current. This particular example
happens to be one in which the Integrated Circuit
bears little resemblance to its conventional
counterpart. The basic electrical elements used
include resistance for the generation of heat,
transistance in the conversion by Seebeck Effect
of heat to direct current, and thermal and electrical insulation at appropriate places.
Basic Electrical Elements
New as it is, Integrated Circuitry is made up
of the same six basic electrical elements which
long have been the ingredients of the electronics
art. They are insulation, conduction, resistance,
capacitance, inductance, transistance, and some
special combinations of these. By insulation we
simply mean the prevention of current flow, or
the isolation of electric or magnetic fields.
Conduction signifies the free flow of electric
current and resistance, capacitance, and inductance have their customary meanings. Transistance
is a new and highly useful term which describes
the gain of active elements, or their ability to
achieve precise control. It applies to all conventional forms of transistors, diodes, and other
Solid State active elements. Finally, by special
combinations we mean elements which can only be
approximated by networks in our conventional
technology. The prime example is the distributed
resistance-capacitance network realized in
Integrated Circuitry by a thin-film resistance
deposited upon a dielectric layer, which in turn
has been deposited on a conducting medium.
Evolution of Microsystem Electronics
To gain perspective in our survey of microsystem electronics let us look at its history and
evolution.
The trend was set by miniaturization which
made use of smaller forms of conventional components interconnected by conventional wires and
solder. In every case the discrete nature of the
individual component and usually its conventional
form has been maintained. Miniaturization represents the first step, chronologically, in the
attempt to make electronic e~uipment smaller.
The next step was sUbminiaturization with
still smaller forms of conventional components.

In most cases the conventional form of the component has ~een maintained but the size and weight
are reduced to the point at which thin wire leads
provide ade~uate support for mounting. The
cordwood techni~ue, in which cylindrical axiallead components are stacked like cordwood with
their leads fed through matched holes in parallel
printed circuit boards in front of and behind the
stack, 'is a good example. Various forms of
printed or etched circuitry, both rigid and
flexible, are used as the combination wiring harness and support.
Microminiaturization is the name given to
the ultimate size reduction of the individual
component. It differs from subminiaturization in
that the conventional shape and form factor are
generally lost, leads are often left off, and a
supporting board or matrix is always necessary.
Microminiaturization still permits the circuit
designer uninhibited freedom of choice in the
selection of his individual components.
Thin-film Integrated Circuitry represented
a major advance by dispensing with separate
mechanical supports for each component and combining multiple thin-film components on a single
glass or ceramic substrate. Overlapping or
touching films form internal connections. At
present a hybrid form of thin-film Integrated
Circuitry is necessary since none of the many
companies working in this field has perfected a
way to fabricate workable thin-film diodes and
transistors on a glass or ceramic substrate.
Current thin-film integrated circuit technology
involves deposition of thin-film passive elements
followed by the appli~ue of any re~uired active
semiconductor elements.
Semiconductor Integrated Circuitry means the
combination of thin-film and semiconductor circuit
elements on and within a single crystal semiconductor substrate. Connections are made by
deposited thin-film conductors, and by physical
juxtaposition. Of all the existing forms of
Integrated Circuitry, the Semiconductor version
permits the widest variety of active and passive
elements ana the greatest potential flexibility.
Acti ve element'; can be formed wi thin or on the
substrate wher~ needed, and either thin-film or
semiconductor techni~ues can be used to form the
passive circuit elements. When semiconductor
technologists have perfected a means of depositing thin-film semiconductor active elements on
glass or ceramic substrates, the passive substrate
approach may well prove more 'flexible than the
active substrate approach. This comes about
because part of each active element is unalterably connected to a common semiconductor crystal
in the active case. Although careful placement
of active elements and tailoring of the shape
and thickness of the active substrate between may
allow ~uite complicated circuits to be built into
a single semiconductor substrate, more complex
circuits could undoubtedly be achieved if both
active and passive elements could be deposited on
an insulating substrate.
MOrphological Integrated Circuitry or what
some people call the functional block, represents
the combination of solid state materials to perform a desired circuit function, although neither
individual components nor precise electrical

65
2.1

circuits are necessarily identifiable. The power
supply mentioned earlier is a Morphological Integrated Circuit. Another example is the familiar
standard fre~uency ~uartz crystal, which is a
homogeneous slab of material although it acts as
a combination of resistance, capacitance, and
inductance. The lack of a physical counterpart
in conventional circuitry makes this type of
Integrated Circuit the most difficult to design.
However, there is a long list of unexplored
physical effects awaiting our attention, and the
possibilities are unlimited.
In the rest of this paper we shall limit our
consideration to Integrated Circuitry, as defined
above, since it represents the most significant
departure from conventional technology. The
following sections first review some of the
reasons for developing Integrated Circuitry and
then present a catalogue of the basic applicable
fabrication techni~ues.
Philosophy of Integrated Circuitry Development
Having briefly examined what it is and the
evolutionary stages through which it is developing, let us consider some of the motivations for
the development of Integrated Circuitry. Although
there may be disagreement as to the order of
importance, it is not hard to identify the following areas of concern as having motivated the electronics industry in this field: reliability,
size and weight, speed, power consumption, accessibility and cost.
Reliability
Everyone talks about reliability, and it is
certainly true that without a definite minimum
reliability neither microsystem nor any other
kind of electronics can long endure, but is there
any significant reason to expect a better reliability from Integrated Circuitry than from conventional component electronic circuitry? Yes,
say its proponents, for at least two reasons:
first, because the number of discrete point
connections is greatly reduced and connections
have always been one of the sources of failure;
and second, because the overall reliability of the
system is no longer the same complicated function
of the individual reliabilities of each individual
component. This latter claim stems from the
common mode of fabrication of Integrated Circuits
where usually all of a given type of electrical
element, say capacitor, are created at the same
time. This common fabrication may seem to have
more bearing on process yield, and hence on cost,
than on reliability but there is a connection with
reliability buried back in the reduction of total
process steps and the conse~uent greater attention
that must be paid to each process step.
It is undeniably true that our electronic
systems, be they computer, communications, or
weapons control, are getting progressively more
complex. With increased numbers of components
all having to function together, we seem to be
approaching an asymptotic barrier where additional complexity can only be achieved at the cost of
decreased reliability. While we are still a long
way from the machine with infinite complexity and

zero reliability, our modern systems are already
uncomfortably close to the complexibility-reliabili ty barrier.
Size and Weight
At least three sources of concern can be
grouped under the heading of size and weight. For
missile and other airborne applications the size
and, often more important, the weight of an electronic system has a cost in propulsion machinery
that provides a strong motivation for reducing
both. While reliability is of very great importance in some of these applications, reducing
size and weight to particular values can mean the
difference between the possible and the impossible.
A second size and weight motivation is
found in the realm of portable products where the
size of a poteutial market, both industrial and
military, often depends on whether a system can
be created and packaged within certain limits.
A third motivation for size reduction is the
economic, where an e~uipment must be housed or
protected or otherwise maintained in certain conditions. If the cost of providing a re~uired
environment is high enough and is proportional to
the volume re~uired, considerable effort can be
justified in reducing the volume. A final motivation is shared with the next area of concern.

For a long time i~creasing the speed of
electronic systems was a matter of improving the
components. Recently the speed of certain systems has reached a point where propagation delays
in the wiring, rather than the characteristics of
the devices, prohibit faster operation. This
point is often called the speed-size ceiling and
it has important bearing on computer Bystems. If
a particular system has reached its speed-size
ceiling, it is impossible to expand it functionally without decreasing its speed of operation,
and vice-versa. The only way through a given
speed-size ceiling is to decrease the physical
size of a given system and thus shorten the
propagation delay through it. Microsystem electronics, and particularly Integrated Circuitry,
shows promise of substantial improvement in
speed-complexity product by significant reduction
of propagation delay within and between integrated circuits.
Power Consumption
Information, in an electronic system, must
be related to some form of energy if it is to be
processed. In general the amount of energy
contained in a given piece of information is very
small in comparison to the energy expended in
processing it. In part this is because of the
inefficiency of the processing e~uipment, but
often a much larger drain is deliberately introduced in order to assure continued operation in
the event of drift or change in certain component
characteristics. Where many different processes
are involved in fabricating the components for an
e~uipment, many different and often uncorrelated
drifts and changes may be expected. To assure

66
2.1
proper operation under these conditions may require ten or one-hundred times the power needed
in the absence of change. The advantage of Integrated Circuitry in this respect is based on the
reduction in number of different processes involved in fabricating the equipment. If all resistors
have the same temperature coefficient and aging
characteristic, for example, variations in bias
and operating point of associated active elements
will be much reduced.
As we shall show in the catalogue of integrated circuit techniques which foxms the major
part of this paper, not all resistors useful to
Integrated Circuitry are made by the same process.
However, many of the resistors in a given integrated circuit will not only be made by the same
type of process but will also have been made at
the same time from the same material and under a
single set of conditions.
Accessibility
There are a number of facets to the problem
of accessibility of Integrated Circuitry. First,
it is apparent that those parts of electronics
systems that must interact with a human operator
must be matched to him in physical size. Knobs
must be such that fing~rs can grasp them, dials
of a size that the eye can read them, and so
forth. Input and output equipment must be
matched to the environment from which it receives
and to which it gives information. These are
problems of accessibility where limits of useful
size reduction can be detexmined and beyond which
it is neither reasonable nor desirable to go.
The second set of problems revolve around
the need for replacing defective parts of the
system upon failure. It is no longer feasible to
think. of "repairing" a
failure. When some part
of an integrated circuit as we have defined it
stops working, it cannot be operated upon in the
field. The service man must treat the entire
circuit, perhaps even a group of related integrated circuits, as a single "component" which he
replaces by a new "component" from an inventory
of known working spares. But we cannot afford to
throw the radio away just because the dial cord
is broken, and hence we have the problems of
suitable modular subdivision and of accessibility
for replacement. The problem further breaks down
into one of size and one of interconnection.
Th~ rroblem of size in accessibility for
replac~ent is much like the first problem of
matching to a human operator. The technician
must be able to get a grip on the faulty module
to remove it.
The problem of interconnection is much
deeper and more fundamental. It reaches across
and can nullify the motivations of reliability,
size and weight, and speed as well as accessibility. Many present day electronic systems use
almost as much volume for interconnecting wiring
behind the panel as they use for componen~s out
in front. Unless the connecting means from
circuit modules to wiring harness is more reliable than the Integrated Circuitry itself,
system reliability will be hurt. If the wiring
harness is physically long, the propagation delay

in getting the signal from one part of the system
to another reduces the maximum speed of operation.
Finally, unless the interconnecting means is
flexible and cleverly arranged, it may be almost
impossible to get at the connections to replace a
module, since the motivations of reliability and
size tend to rule out the use of plugs and sockets.
The final set of problems of accessibility
are concerned with the removal of heat. The
potential saving in power through an allowable
decrease in operating margins has already been
mentioned, but even so a significant part of
integrated circuit design is thermal in nature.

The cost picture for Integrated Circuitry
is a complicated one. Initially some of the
specialized appl~cations for which there are no
other possibilities will probably support Integrated Circuitry regardless of cost. Any sort of
general acceptance, however, will require a cost
competitive with other forms of circuitry. Since
individual components can no longer be selected
after manufacture, both control of process and
process yield will have to be high. An aid in
this respect is the fact that the same process
run usually forms all of a given type of circuit
element on a particular integrated circuit substrate and thus if one part is good all tend to
be good. This simultaneous fabrication of all
diffused resistance regions or all deposited
capacitors at the same time is one of the advantages by which its proponents hope to lower
the cost of Integrated Circuitry.
Catalogue of Integrated Circuit Techniques

Our survey of micro system electronics may
logically end with a catalogue of existing and
proposed means for obtaining the elementary
circuit functions. Since the assembly of complex
circuit functions called for in a designable
Integrated Circuit Technology must rest fundamentally on the precision and process control
attainable in the basic circuit functions, major
continuing effort has been put on the study of
means for obtaining greater control over these
individual processes.
The ten subdivisions which appear below are
the ten basic elements needed by almost every
circuit regardless of techniques used to build it.
Method of Catalogue
There are a number of ways in which Integrated Circuit Technology may be catalogued. We
may catalogue it by process, by materials used,
by physical foxm or arrangement, by the basic
circuit parameters, by the energy forms and transformations, by generic circuit function and,
finally, by specific circuit function in a representative system. For simplicity we have chosen
to subdivide by basic circuit parameters:
1.

2.

3.

Insulation
Conduction
Resistance

4.
5.
6.
7.
8.
9.
10.

67
2.1
Capacitance
Inductance
Special Networks
Active Elements and Substrates
Encapsulation
Mechanics
Design.

The above catalogue is by basic circuit parameters rather than one of the other means of
classification because for a survey such a breakdown seems to be most meaningful and concrete.
In developing these basic elements we sometimes use the bulk properties of the body of the
substrate, sometimes operate at or close to the
surface by alloying or diffus~on, and sometimes
build the element on top of the substrate by _
epitaxial growth of semiconductor materials,
evaporation, plating, or other means of deposition
of thin films of material. The mechanical, thermal, electrical, and chemical interactions resulting from these new approaches to electronic circuits have introduced both new problems and
indeed a new science.
Insulation
Electrical insulation is required in all but
the most trivial Integrated Circuits. Extrinsic
techniques of insulation include operations external to the substrate. The development of
extrinsic electrical insulation includes anodization, vapor phase deposition of dielectrics by
pyrolysis, evaporation, and plasma deposition
techniques as well as the conventional mechanical
coating. Intrinsic insulation includes any
method of isolating fields or current flow within
a semiconductor or other substrate. Among these
are thermal oxidation, fabrication of an isolating
layer of intrinsic semiconductor material, and
creation of one or more back-biased junctions.
Anodization is the formation of an insulating
oxide over ~ertain elements, usually metals, by
electrolytic action. The most commonly anodized
materials are tantalum, aluminum, titanium, and
niobium. Anodiaation is a particularly useful
form of insulation where protection of a conductor
is required since the base metal can form the
conductor and the anodized surface layer can form
the insulator. Since anodized films of tantalum
can be controlled to a high degree and possess a
dielectric constant of approximately 25, the
process finds wide application in formation of
capacitor dielectrics. The high dielectric constant becomes a disadvantage as the frequency of
the energy to be insulated increases.
Pyrolysis as used here is the thermal decomposition of a volatile chemical compound into
nonvolatile and volatile byproducts. It is generally carried out in an inert carrier gas and
this fact distinguishes it from vapor plating,
which generally uses an active carrier gas such
as hydrogen or steam. Silica, silica-based glasses,
and silicone polymers have been deposited successfully.

Evaporation is the deposition in high vacuum of insulation thermally liberated from a
parent source. Silica films of low optical absorption have been produced by electron bombardment of the parent oxide. The technique has
important possibilities in the area of direct or
mechanically masked insulation deposition.
Plasma Deposition involves the spraying of
highly excited atomic particles of the insulator.
Heat, commonly obtained by electric arc or hydrogen flame, is the source of excitation and almost
any elementary material can be applied to almost
any surface by the method. The disadvantages are
the grossness of the spray process and the high
temperatures involved.
Mechanical Coating covers the spraying,
painting, or other physical application of organic
and inorganic insulators and finds certain applications in the grosser insulation of Integrated
Circuitry.
~ means of extrinsic techniques broad area
Jielectric film insulation has been ~eveloped
with field strengths in excess of 10 vOlts/cm.
These insulating films can be reproducibly obtained to thicknesses in excess of 10 microns.
Thermal Oxidation is here used to denote the
formation of a self-oxide upon the exposed surfaces of a semiconductor. The intrinsic insulation of silicon has been developed to a high
degree by this method. Oxidation of parent material can be used to insulate broad areas and
also PN junctions.
Intrinsic Layering refers to the method of
deparating two regions of conductive semiconductor
by a region of near intrinsic semiconductor material which differs enough in resistivity from
the adjacent regions to serve as an insulator.
"The epitaxial growth of silicon is a promising
method of providing layers of intrinsic material.
Back-biased Junctions may be formed within
semiconductor substrates to provide a type of
insulation. With reasonable values of reverse
bias the capacitance across a graded junction can
be reduced to the point of insignificance for lowfrequency applications. The division of semiconducting supstrates into two or more regions
isolated by reverse-biased junctions opens the
way to an eventual increase in the number or
complexity of the functions that might be built
into a single substrate. A disadvantage of the
technique is the collection of any minority
carriers in the area.
Conduction
Conduction is used here to indicate electron
current on or in substantially ohmic material of
negligible resistance. One of the goals of Integrated Circuitry is to achieve integration of
circuit functions on and within substrates so that
ohmic connections from one location on the substrate to another are drastically reduced or

68
2.1
eliminated. In cases where this is impossible or
impractical, extrinsic conducting paths are formed by evaporation of metals, by painting or plating of conducting stripes, by sputtering, by
pyrolysis or by mechanical coating. Intrinsic
conduction can be accomplished by development of
degenerate regions within the semiconducting or
substitute substrate. We can do this by creating
alloyed, epitaxial, or melt grown regions during
or following crystal growth.
Evaporation of conductors covers the vaporizatioll of metals at high temperatures in vacuums
of 10- mm Hg or better. The metallic vapor moves
in substantially straight lines onto a substrate
which may be mechanically masked to limit the
deposition to desired regions. The substrate must
be very clean and must generally be raised to an
elevated temperature in order to assure intimate
contact of the particles with the substrate after
impact before solidifying. Alternative means of
producing conductive patterns involve coating an
entire surface and subsequently removing unwanted
conductive material by one of a number of techniques. Evaporation may be of a single metal or
of an alloy. In the latter case the deposition
may be made simultaneously from a common source
or sequentially from separate sources, after which
the substrate may be heated to produce an alloy on
the surface of the substrate. If the alloy penetrates the substrate, as in the case of a semiconducting substrate, the result is classed as
intrinsic.
In the study of conductors by deposited
techniques, resistivities below one order of
magnitude higher than metal conduction have been
achieved. Where mechanical or thermal considerations make adhesion more important than achieving
the lowest ohmic resistance the introduction of a
chrome-gold alloy has been useful.
Sputtering differs from the evaporation
process discussed above in the conditions under
which the conductive material reaches the substrate. Sputtering is the result of a glow discharge between an inert anode and a bombarded
cathode of the desired conducting material.
Because the presence of gas at 10- 2 to 10-4mm Hg
is necessary for the generation of the ionized
bombarding molecules, the sputtering process is
inherently harder to keep clean. Even so the
process has certain advantages over vacuum
evaporation in the relatively low temperatures
needed or generated in the system, and thus the
lower chance of contamination from the evaporative source. It finds one of its most important
applications in the production of thin films of
tantalum for resistors or capacitors.
Pyrolysis as used for deposition of conducting films is the same process discussed previously
for insulators. It has the advantage of flexibility of materials and conditions of deposition
but the disadvantage that the deposited material
cannot be masked by mechanical shields as satisfactorily as the vacuum evaporated materials.
Pyrolysis finds application where an entire
surface can be coated as for electrostatic or
magnetic shielding, or where the unwanted material

can be removed selectively after deposition.
Pyrolysis may also be carried out on selective
regions under certain circumstances by employing
a catalyst.
Plating of conductors includes electroplating) chemical or electroless plating, and
vapor plating. Electroless plating tends to cover
everything as does vapor plating. Selective
removal of unwanted material can be combined with
electroplating to build up conducting paths. Both
electroplating and electroless plating suffer from
danger of contamination from the wet chemistry
involved. Vapor plating is capable of achieving
high-purity deposition.
Mechanical Coating includes conducting glass
pastes which can be painted onto gross terminal
pads to bridge irregularities that vapor deposition cannot manage. It also includes various
solders that may find temporary use in making
conductive paths to external terminals. The
conductive glasses have the advantages of bonding
well to many substrate materials and of reasonable
match in thermal expansion coefficient.
Degenerate Regions are regions within a
semiconductor substrate in which the conductivity
is very large and in which carrier transport is
sssentially ohmic. These regions are generally
produced by alloying near saturation concentrations
of a doping impurity into a specific pattern.
Because of this method of formation the paths are
generally at the surface of the substrate. By
control of the intrinsic techniques mentioned,
resistivities of 10-3 to 10- 4 ohm-cm can be obtained and in general may be considered to be
feasible for internal conduction.
Connections from one Integrated Circuit to
the outside world or to another substrate are
treated in the Section on Mechanics as in conduction of heat away from dissipative el~ments
within the Integrated Circuit.
Resistance
Resistance may be provided in Integrated
Circuits in at least four ways. The three
intrinsic ways are by bulk resistivity, by transverse conduction in thin diffused back-biased
regions within a semiconductor substrate, or by
an epitaxial layer of opposite impurity backbiased with respect to the parent substrate. The
extrinisic way is by deposited thin-film resistors on top of the substrate. Each method has
particular advantages for certain applications
and a complete Integrated Circuit capability
requires mastery and evaluation of each. Both
methods may be sensitive to their ambients.
Anodized Tantalum, Titanium, or Aluminum
films provide an attractive method for obtaining
highly precise resistance values. The initial
deposit of these metal films is relatively thick
and is easily formed by the vapor plating and
vacuum deposition techniques previously cited.
After deposition, the metal film can be trimmed
to a precise value by anodizing the outer surface
of the metal film to its oxide which is an

69
2.1
insulating semiconductor. Since the oxide thickness is a direct'function of anodizing voltage,
a high degree control of the remaining metal film
thickness can be obtained. Resistance films in the
10 to 500 ohms per square range can be obtained
with±lj4 per cent reproducibility. Accurate
layout of these resistance films is achieved by
the application of photoresist techniques to define areas of anodization. The unanodized film
'can be utilized to form conduction elements.
Tin Oxide films have reached sheet resistance
values of over 5000 ohms per square. With conventional resistance patterns, values of up to 1.0
megohm can be achieved. The films are produced by
hydrolysis in a technique that has been brought
to a high degree of development.
Indium Oxide films have been produced by a
two-step metallizing-oxidizing process involving
a vacuum deposition of pure indium in a low
pressure pure 02 atmosphere followed by a low
temperature (below 200 0 C) thermal oxidation for
several hours. This low temperature technique
has application in instances where higher temperature resistance fabrication would damage
other temperature sensitive thin-film functions.
Nichrome films can be evaporated directly on

~lean substrates by volatizing the alloy from a

tungsten heater. These films show excellent
adhesion when substrate surfaces are heated to
300o C. Nichrome resistance films can be reproducibly deposited and show relatively high stability on standing. They possess an average
temperature coefficient of resistance of 6 x 10-5
ohms per ohm per degree centigrade over the temperature range -50 C to 150 C. However, these
films have low resistivities which limit their
application to 500 ohms per square. In addition,
the surfaces of unencapsulated Nichrome films
are prone to a certain amount of corrosion and
oxidation which limit compatibility with other
thin-film circuit element fabrication. Specific
geometries of Nichrome resistance elements are
usually achieved by the use of mechanical masks
during evaporation. These films are quite amenable to forming ohmic contacts with other metal
films.
Bulk Resistivity makes use of the gross
geometry and parent material of the semiconductor
substrate. It is easily controlled but is suitable only for relatively low values of resistance
based upon maximum usable resistivity of ~500
ohm-cm. Since bulk resistivity of 500 ohm-cm is
not compatible with present semiconductor base or
collector technology, a more realistic upper
limit is 100 ohm-cm with the most compatible
value probably lying between 10 and 30 ohm-cm.
Bulk resistivity has the additional handicap of
a complicated temperature coefficient. This
temperature dependence can occasionally be turned
to advantage in circuit applications for temperature compensation of specific types of circuit.
Diffused Back-biased Regions cover the class
of resistors formed by transverse conduction within a thin diffused back-biased layer of semi-

conductor. They cover a much wider range of values and have additional advantages in their
possible range of temperature coefficients. By
control of the diffusion profile it is possible
to achieve positive, negative, or substantially
zero temperature coefficient at room temperature.
The difficulty in using diffused resistors in
practical circuits lies in their great sensitivity
to the back-biasing voltage including the selfgenerated component of back bias caused by the
voltage drop in the resistor. Where very high
resistance of non-critical value is required,
diffused resistors may have great value. Their
other attractive application is in circuits
needing an electrically alterable resistance.
Epitaxially Grown Resistive Layers can form
resistive regions with useful characteristics.
The method is similar to the diffused back-biased
resistance described above but the epitaxial
process promises better control of the junction
characteristics. The epitaxial resistor requires
a compatible masking technique during layer
growth if mesa techniques and wet chemistry are
to be avoided.
Capacitance
A number of methods have been developed for
providing Integrated Circuit capacitance. Capacitance may be provided intrinsically by reversebiased semiconductor junctions, by self-biased
junctions or extrinsically by deposited thin-film
capacitors using gold or some other conducting
film as a counterelectrode.
Anodized Tantalum, Titanium, JUuminum, or
Niobium can be used to form the lower conducting
layer and dielectric of deposited thin-film
capacitors. After the desired thickness of
dielectric has been formed, a counterelectrode
of some conducting material is deposited to
complete the capacitor. The films are generally
deposited by evaporation or sputtering'. Subsequent anodization can be controlled to a high
degree and pinholes cleared before deposition of
gold as the counterelectrode. Ratings of 5.0
volt microfarads per square centimeter at 50% of
breakdown voltage are now state of the art.
Titanium, Aluminum, and Niobium, can also
be anodized with useful characteristics. Aluminum oxide has a dielectric constant less than
25% that of tantalum but has almost twice the
working voltage for the same forming voltage.
Unless the counterelectrode is of the same material as the anodized material, this type of
capacitor is polar.
Deposited Metal Oxide Glasses can be sandwiched between deposited conductors to form an
alternative thin-film capacitor. Here the dielectric constant is much lower than either of
the anodized dielectrics mentioned above but the
use of low melting silicate type glasses avoids
some of the problems of the anodized capacitor
process. Deposited metal oxide glass dielectric
capacitors can be fabricated at lower temperatures
than are required in the deposition of tantalum.

70

2.1
Deposited Ferroelectrics offer attractive
possibilities as dielectrics for thin-film capacitors. Chief among the attractions is a dielectric constant for barium titanate three orders
of magnitude greater than that for the silicate
type glasses. Such a dielectric might also have
important contributions stemming from its nonlinearity and polarizable nature. The chief disadvantages appear to be a limited operating range
of temperatures and instability or deterioration
of electrical properties. Work with deposited
ferroelectrics is in the early exploratory stage.
Diffused Back-biased Junction capacitors
depend upon the depletion layer as dielectric and,
for step junctions at low voltages, can provide
on the order of 1 volt-microfarad/sq.cm. However,
since the width of the depletion layer varies with
the magnitude of the reverse bias, such capacitors
are electrically alterable with the magnitude of
their effective capacitance depending on their
bias. This voltage dependence can be an advantage
or disadvantage depending on the use to which the
capacitor must be put.
Reverse-biased junctions act as capacitors
whose value depends not only on the junction area
but also on the width of the depletion region at
the junction. This width depends upon the impurity concentration grauient on each side of the
junction and upon the reverse voltage applied
across the junction. Since the normal diffusion
process produces a graded junction, capacitors
formed by diffused junctions tend to have lower
values of capacitance at the same bias than
either the alloyed or the epitaxially grown
junctions discussed below.
If current flows parallel to the junction
outside the depletion region, the bias on the
junction will not be everywhere the same. In
particular, if current flows through the bulk
resistance of the substrate beneath a shallow
diffused but otherwise unbiased junction, the
floating junction will conduct to equalize potential at one end and will thus back-bias the
rest of the junction. The result can be useful
in circuits where a small capacitance is needed
to bridge a load or coupling resistor. Note that
this configuration is really a passive network of
distributed resistance and capacitance.
Alloyed Back-biased Junction capacitors tend
to have a much steeper impurity gradient across
the junction on one side while maintaining practically the impurity concentration of the parent
substrate on the other. The approximation to a
step junction is much better than in the diffused
case and much higher values of capacitance at low
reverse bias are possible. Because of the relatively low impurity concentration of the substrate, the total depletion layer is very nearly
as wide at large values of reverse bias as in the
diffused-junction case.
Epitaxial Back-biased Junction capacitors
hold promise of the ultimate attainable in
reverse-biased junction capacitors. In addition
to being more amenable to control than the alloy
process, they are able more .closely to approximate

the step junction from high impurity concentration
of one type to high impurity concentration of the
other. This means that they should be capable of
having the maximum capacitance, at low values of
reverse bias, of any junction capacitor. It also
means that their voltage dependence should be more
predictable.
Thermal Oxidation on the surface of a semiconducting substrate can provide the dielectric
for a hybrid type of capacitor in which one plate
is the semiconductor substrate and the other plate
is a thin metallic film. Such hybrids can have
occasional applications as special networks.
Where precise and voltage-stable capacitors
of less than a microfarad are needed, deposited
thin films of tantalum, titanium, aluminum, or
niobium can be anodized with good control. Highvoltage or nonpolar capacitors with small values
of capacitance can be fabricated reliably using
silicate glass type dielectrics sandwiched between
metal film electrodes. Voltage variable capacitors
or non-critical coupling capacitors can be formed
in semiconducting substrates by back-biased
junctions.
Inductance
From an Integrated Circuit point of view,
inductance is without doubt the most difficult to
obtain of all the basic circuit parameters. The
major difficulty is the requirement of a volume
for the storage of magnetic flux which does not
lend itself readily to Integrated Circuitry.
Furthermore, the coupling of energy to the volume
poses the requirement of a coil that is not easily
achieved by deposited film techniques. Both a
s~uare-loop and a linear response are required
for a general systems capability_
Deposited Nickel-Iron Films are among the
extrinsic means available for storing energy
resulting from the flow of current. Deposited
nickel-iron films of 82%-18% composition and
1000A to 4000A thickness have been used for
storing energy for logical matrices and as smallvalued low-frequency inductors. A requirement is
the presence of a magnetic field during the deposition process to orient the magnetic anisotropy.
Unfortunately this limits the magnetic film
configuration to simple forms. If the driving
magnetic field is applied in the direction of the
"easy" magnetization of the domains, a squareloop B-H curve results which is useful for memory
and for magnetic logic applications. When the
magnetic field intensity is applied normally to
the direction of "easy" magnetization, a more
linear B-H curve results which is useful for
linear systems and impedance transformation.
Deposited Ferrites have possibilities as a
second extrinsic technique for obtaining inductance. Glass, which is a mixture of metal oxides,
is currently being deposited by the pyrolytic
decomposition of su~table metallic-organic esters.
The reaction temperature required to form the
ferrite material from mixed oxides is in the
order of 300 o C. Since ferrites do not have the

71

2.1
same magnetic anisotropy as thin nickel-iron
films, no magnetic field is needed upon deposition
and the form factor is not limited as it is with
metallic films. Magnesium-manganese ferrite material provides a s~uare-loop B-H curve and is,
therefore, suitable for magnetic logic elements.
Manganese-zinc ferrite provides a high MQ linear
material usable to about 500 kc. Nickel-zinc
ferrite provides a highM Q linear material suitable
for the fre~uency range 0.5 to 100 mc. By proper
masking methods it is possible to form thin-film
solenoids which surround such depostted ferrite
materials and provide a means for coupling energy
into and out of the material. Finally ferrite
materials possess variable permeability which is
a function of the applied dc magnetic field, and
this can provide a control element for ac magnetic
flux. Such variable inductors can serve as electronic tuning elements or other control elements.
Air Core Geometries are suitable for r.f.
coils and other high-fre~uency small-valued
inductors. It is possible to deposit "air core"
pancake-type windings of thin-film conductors.
When associated with thin-film insulators of low
dielectric constant the pancake-type winding can
be formed in multilayers to increase the total
inductance of the element.
Inversion of Capacitance-by Active Element
networks is among the intrinsic techni~ues for
obtaining inductance. Field effect semiconductor
devices can provide an impedance inversion
function. By this means a low-Q capacitor can be
made to appear as a high-Q inductor for circuit
applications where resonance is not re~uired.
Ferrite Substrates provide the possibility of
using a single or multi-aperature ferrite material
both as an inductive core and as a substrate for
other Integrated Circuit elements. The coupling
to the ferrite can be achieved by thin-film
conductors. These elements, because of the dependence of their permeability on the applied
field, can also serve as a means for controlling
magnetic flux.
The Inductance Diode is another intrinsic
approach to the problem of providing inductance
for Integrated Circuits. Although this device is
still in the early experimental stages a number
of workers in the field have succeeded in measuring inductive behavior of germanium diodes. The
maximum effect is found with alloyed or near-step
junctions in which the injection efficiency is
close to unity. Chief difficulty in measuring or
applying the observed inductive behavior is the
instability of the inductive effect, both with
current through the device and with temperature.
The problems surrounding the incorporation
of inductance in Integrated Circuitry are severe
and first efforts at Integrated Circuit design
will probably try to accomplish the desired e~uip­
ment functions without the use of inductance.
Special Networks
A few basic circuit parameters are of a hybrid
nature and do not fall logically into any of the

pure parameters discussed above.
Deposited Distributed R-C networks are
among the extrinsic examples of functions that
cannot easily be duplicated by lumped constant
circuits. A simple example of this is an anodized tantalum resistor upon which has been deposited a counterelectrode of gold or some other
metal. Each part of the resistor has a capacitive
relationship through the anodized dielectric to a
unipotential conducting plane. The result is a
series resistance with distributed capacitance
acting all along its length.
Deposited L-C networks are in the same class
of difficulty as ordinary thin-film lumped inductance. About the only simple example is the pancake air core coil deposited as the counterelectrode of a deposited and anodized tantalum
capacitor.
Transformer-like thin-film configurations
have been proposed but both distributed L-C and
thin-film passive impedance transformation for
Integrated Circuitry are in the early experimental
state.
Diffused Back-biased R-C networks are similar
to the thin-film extrinsic kind except that they
are voltage sensitive as to their capacitance and
sometimes even their resistance. If the currentcarrying diffused layer is thin enough, or if the
back-biased junction is between the parent substrate and a thin current carrying epitaxial
layer, variation of the reverse-biasing voltage
will affect Doth the series resistance of the
thin layer and the shunt capacitance to the substrate. Such a configuration can form a tuning
unit in a phase-shift oscillator or feedback
amplifier.
Bulk Resonance effects such as the piezoelectric resonance of specially cut ~uartz
crystals are classed as intrinsic special networks
even though they normally occur in other then
semiconducting substrates. In the case cited,
the mechanical motion of the crystal would probably be disastrously damped by using it as the
substrate for other Integrated Circuits, but it is
conceivable that thin-film circuits could be laid
out entirely along nodal lines.
Active Impedance Transform Networks cover a
wide and largely undeveloped field. 'rheir scope
extends from the impedance transformation characteristics of ordinary non-integrated devices
such as transistors and field effect devices, to
vague and esoteric proposals_for active delay
networks and hybrids of active and passive
phenomena.
By definition, the functions included in this
category do not in general exist in a one-to-one
correspondence outside the solid state. For this
rea$on they are perhaps closer than some of the
other basic circuit functions to being Integrated
Circuits themselves. The members of this category
are expected to increase as new integrated phenomena are developed.

72
2.1
Active Elements and Substxates
The active semiconductor elements useful in
Integrated Circuitry include the logic diode, the
breakdown or zener diodes, various multilayer diodes, the mesa transistor structure, the flat or
oxide-masked mesa, the unipolar transistors,
various other field effect devices, and compound
transistor-like elements. The list of non-semiconductor active elements includes ferromagnetic
and ferroelectric elements and many others.
Vapor Deposition of Semiconducto~ Active
Elements on a passive substrate is undoubtedly the
most important extrinsic technique in this category.
To date, no one has reported on the deposition
of large-area nondegenerate single-crystal thin
films of semiconducting material on an insulating
substrate. When this technique is fully developed
it will be possible not only to build complex
active functions but to build them on either
active or passive substrates as the situation
dictates. One of the most important consequences
following on the eventual achievement of the thinfilm deposition of active semiconductor elements
will be the removal of the topological-electrical
limit to the number of active elements that can be
built into one semiconductor substrate. Another
will be a decrease in the minimum capacitive
coupling attainable in an Integrated Circuit.
Vapor Deposition of Nonsemiconductor Active
Elements on a passive substrate can make useful
many previously overlooked solid state effects.
Both ferroelectric and ferromagnetic materials
are unique in that certain of their basic properties, such as dielectric constant or permeability, can be readily changed by the application
to the material of an electric or magnetic field
of the proper magnitude. This property has been
utilized in microcircuits for the purpose of
circuit tuning. When the problems of the deposition of coherent films of these materials are
overcome, an entirely new generation of active
circuit elements will be possible, utilizing the
nonlinear properties of ferroelectrics and ferromagnetics to actively tune circuits, act as bandpass filters, and so forth.
Applique of Active Elements has been the
universally used extrins1c expedient since thinfilm semiconductor active elements do not yet
exist on insulating substrates. The active elements are attached to passive substrates to make
a hybrid structure. _ They 'IIJB.y either be conventional or special microminiature elements in cans
with leads brought through the passive substrate
and soldered, or they may be unencapsulated or
self-encapsulated devices mechanically and electrically affixed on the passive two-dimensional
substrate.
Conventional Active Element Design can in
general be transferred directly to Integrated
Circuit application. Intrinsic techniques for
creating active elements within semiconductor substrates include doped or rate-grown junctions in

crystals as pulled from the melt, alloyed junctions, diffused junctions, vapor-grown junctions
as epitaxially oriented overgrowth and hybrid
junctions where one side of the junction is the
result of one process and the other of another.
§pitaxial Techniques of growing silicon
layers on silicon parent stock are among the
most promising tools in the area of threedimensional Integrated Circuits. Vapor phase
epitaxially oriented overgrowth was first developed in Europe and has since been intensively
studied and applied by American materials and
components manufacturers.
The epitaxial process permits a number of
desirable improvements over conventional active
element technolOgy. The previous requirement
with active semiconductor substrates that the
active elements be fabricated within the substrate becomes within or on the substrate. The
substrate can now be a lower resistivity than
could be tolerated when it had also to form the
collector junction of our active elements. Largearea step junctions are now much more nearly a
reality and the only diffusion effects are those
occurring at the junction while the epitaxial
layer or layers are being grown.
There are many facets of the process that
need exploration and development but the prospective rewards are great - both from the standpoint of new and hitherto unfeasible active and
passive elements and from the standpoint of
logical extension to automated, low cost, and
highly flexible Integrated Circuit assembly
processes.
Form Controlled Growth. In the extension of
their studies of semiconductor crystal growth a
number of companies are developing methods of
ribbon crystal growth. The basis of all this
work is the ability to control the shape and
structure of a growing semiconductor crystal as
it is drawn from the melt,. In some cases the
ribbon is a dendrite pulled from a supercooled
melt, in other cases different means of obtaining
ribbon shape are used.
Ferrite Substrates. The utilization of
ferrite substrates offers a unique method of
introducing an inductive element into the circuit.
The requirements of such a substrate are those
which must be met by any substrate, such as
surface smoothness, compatibility of thermal expansion and thermal conductivity with other
active elements and physical strength. The use
of a magnetic substrate imposes additional requirements. Magnetostrictive effects might
present the problem of maintaining a bond be~ween
the substrate and deposited films, as well as
changing the electrical properties of the films
through their electrostrictive properties.
In order to take fuli advantage of a magnetic
substrate, the magnetic properties of the material
must be matched to the particular requirements.
For inductive purposes a low loss, high permeability material would be required, suitable for
use at the frequency of interest. The temperature coefficient of permeability might be

73

2.1
important where large fluctuations of the ambient
temperature are anticipated.
The geometry of the substrate would depend'
upon the method used to couple the magnetic field
into the circuit. One method consists of laying
down the coil as a thin film on the surface.
Another consists of providing holes in the substrate
through which, or around which, wires can be
wound.
Encapsulation
The encapsulation problem is a critical one.
Achievement of desired performance, control, and
reliability of semiconductor active elements, and
to a lesser extent passive components, is dependent upon reliable and well-founded solutions
to the protection of surfaces from the migration
of contaminants in the ambient. These include
moisture, radiation, and many other forms of contamination.
Hermetic Seal has been the most widely used
and untlI recently the only satisfactory means
of protecting semiconductor devices from contamination. In the initial work with Integrated
Circui~ry the hermetic seal will undoubtedly
continue to be used. A number of companies have
made great strides, however, in developing other
means of encapsulation and the hermetic seal
should be considered only as a temporary expedient.
Low Melting Inorganic Glasses have shown
remarkable success in improving device perfor~
ance and reliability. The glass may act as an
ion getter in cleaning up surface contaminants
and it protects against 100% humidity for over
10,000 hours. Protective films can be deposited
from the vapor phase over large surface areas:
Their application is limited, however, by the
facts that they may be soft at room temperature,
that sulphur in the glass reacts detrimentally
with certain thin-film materials, and that thermal expansion in mating with lead materials may
cause difficulties.
Inorganic Film Pyrolysis offers a number of
advantages: it is free from pinholes, does not
affect the device, has high dielectric strength,
has high chemical and physical durability, has
low volume permeability to moisture, and re~uires
only low device temperature during encapsulation.
Although the work with pryolitic glass is not
widespread, there are indications that this
approach will have wide applicability to the
physical and chemical protection of Integrated
Circuit Functions of the future.
Anodization has already been discussed under
Insulation, Resistance, and Capacitance. It uses
the parent metal in the oxide formation.
Accelerated Thermal Oxidation is a moderate
temperature reaction carried out at atmospheric
pressure which can form a completely protective
glass type layer by conversion of in situ material.
A characteristic of this process is that it applies
only to material capable of being converted to
glass, and hence is most applicable to silicon

surfaces. In the self-encapsulation of a silicon
device, therefore, the process leaves substantially untouched the leads or contacts. This feature
has advantages for straight device fabrication
but is of restricted applicability where thin
films and semiconductors are combined in Integrated Circuits.
Surface Stabilization is an attractive
possibility involving the deliberate addition of
something to the surface of an active element
which ties up all the possible loose ends and
leaves the surface indifferent to subsequent
contamination.
Mechanics
The mechanics of Integrated Circuitry are as
much a technique as the realization of any circuit
element. They include form factor, masking,
thermal conduction, access, and interconnection
of the Integrated Circuits with each other and
with the outside world.
The Form Factor includes the size and shape
of the substrate and must take into account such
things as heat removal, replaceability, accessability for test, position of leads, and so
forth. The initial form factor must allow for
hermetic seal of each Integrated Circuit. A
second step will probably be individual surface
passivation or self-encapsulation with a number
of related circuits in one can for protection.
Ultimately, surface passivation will probably
also constitute physical protection and the
individual circuits will be entirely selfencapsulated.
Masking is another area falling within the
field of Mechanics. Masking is chemical insulation during the fabrication process. Chemical
insulation as used here refers to means of
restricting the action of chemical etches and to
masking portions of the substrate during evaporation, alloying, and/or diffusion. Metallic
masks are a standard part of the art for evaporation as are the photoresist techniques and wax
coating for etch protection. Recently added as
standard art are self-oxidation and vapor deposition of glass which give promise of simplifying
mask technology by an order of magnitude and
reducing complicated micro layout to a photographic reduction of standard art work.
The Interconnection of two or more Integrated Circuits is one of the most important parts
of the program. It may well be the most important
consideration facing all of Integrated Circuitry
at this time. Interconnection covers not only
the communication of the Integrated Circuitry
with associated equipment but also the communication and supply of power between substrates. As
such it includes some of the most important and
perplexing problems in the field. If size, weight,
and reliability must remain tied to our present
method of plugs and interconnecting wiring, much
of the impetus of Integrated Circuits is lost.
Present efforts make use of ohmic connection,in some cases in the form of strips of

74
2.1
conductor sandwiched between isolating ground
planes. Such low impedance strip lines have a
number of advantages from the circuit standpoint
but the difficulties of removal and replacement
of a particular circuit are formidable. Ultimately, other means of coupling (in addition to ohmic
connection) will be used. Magnetic, capacitive,
photon, and ultrasonic coupling are possibilities
as means of conveying information and power.
Designability
The development of an Integrated Circuitry
depends not only on being able to fabricate and
interconnect the basic circuit parameters discussed above but also on knowing what techniques
are compatible with which, what range of parameter values are realizable, what reliability individual circuit functions may be expected to have
and what control is possible in the constituent
processes. It depends on the design of processes
that fit into automated manufacturing methods.
It depends on building into the manufacturing
methods, from initial conception onwards, sufficient flexibility that the output of the product
line can be changed easily and quickly to yield
a different Integrated Circuit.
Finally, the widespread development of Integrated Circuitry depends on being able to accomplish all the necessary functions and operations
economically. In the last analysis, it is cost
that will determine the acceptance and use of
Integrated Circuitry. If a particular fabrication
process cannot be made economically competitive,
it is the wrong process.
Conclusion
Microsystem electronics has struggled through
a variety of stages on the way to maturity. Undoubtedly it has a long way yet to go. But the
need, significance, and accomplishments to date
are such that it can no longer be ignored by the
computer industry. For a while it will be used
mainly by those for whom nothing else will do,
primarily because of size and weight. After
reliability is proven there will probably be
another period when the main use will be in military applications. Developments during this
period will determine whether Integrated Circuitry
spreads throughout our electronics industry or
remains with a few specialized applications. The
determining factor will be cost. If the ultimate
cost can be brought down to that of conventional
computer circuitry, Integrated Circuitry will take
over large segments of the electronic computer
lndustry.

Such an acceptance of Integrated Circuitry
will force another integration. During the proving period computer manufacturers will probably be
willing to design their equipment around available
~ntegrated circuits, but the time will come when
the computer design engineer will demand greater
freedom of choice in selecting the particular integrated circuits from which he builds his machine. When this happens, and it surely will
happen, our industry will go through some interesting gyrations; computer equipment companies will
be pushed into the semiconductor and solid state
materials and components business, and the semiconductor products manufacturer will be forced into
the systems engineering and equipment fabrication
business. The end point will be remarkably similar for the two, even though there will always be
differences in accent and degree. When dynamic
equilibrium is reached, not only the electronic
circuitry but the people who create and use the
new circuitry will be integrated to an extent
unknown today.
There will always be specialists, as there
always have been, but the new specialists will
have a broader base which will frequently extend
into more than a single scientific or technological
discipline. The new specialist will understand
and speak the languages of the associated disciplines and will contribute his point of view in
matters that previously were considered the exclusive concern of others. The chemist and the
metallurgist will take their rightful places beside the solid state physiCist, the circuit designer, the systems engineer, and the statistician.
The successful integration of computer circuitry will depend upon demonstrating adequate
control of designable, compatible, and above all
reliable techniques.
Acknowledgment
In a survey paper such as this, the work and
techniques described are the result of effort by
many people in many companies. In particular I
would like to mention the help and cooperation of
my colleague, James R. Black of MOtorola's Solid
State Electronics Department, in supplying information on thin-film integrated circuit techniques.
Finally, to all technologists whose contributions
are included in these pages, I wish to express
appreciation for sharing their discoveries and
experience. The rapid growth of microsystem
electronics depends upon just such sharing and
cooperation.

75

2.2
TESTING OF MICROLOGIC

ELEMENTS

Robert H. Norman and Richard C. Anderson
Fairchild Semiconductor Corporation
Palo Alto, California
Summary
The acceptance tests of integrated logicfunction'elements require a departure from the
notion that the parameters of the individual components of a circuit must be known. A test program
based on external characteristics only is made
necessary by the complexity of the circuit; it is
contended, in addition, that such a test has more
valid significance than does measurement of constituent parameters. It is more significant, for
example, when testing a compatible set of digital
functional blocks, to make a measurement which
gives a direct estimate of fan-out than it is to
measure the gains, leakage currents, and resistor
values in the output stages.
This paper will discuss the validity of this
concept and describe a program of tests based on
the concept.

characteristics in such a way that we can accurately determine that its present operation is
satisfactory and, as important, that we can
expect continued operation over wide environmental conditions. If this cannot be done, ther~ is
little point in continuing with an integrated
circuit program.
We will support the assertion that this can
indeed be done by discussing a test program which
is based upon the supposition that the assertion
is correct. This will not, of course, constitute
proof; it will give, we hope, an intuitive confidence that the probability of success is high. A
theoretical proof of the adequacy of an "external
characteristics only" test would be both difficult
and unacceptable. The only ultimate verification
of the adequacy of such a test procedure will be
its success in operation after millions of devicehours.
Micrologic

Introduction
In the production testing of integrated logicfunction circuits, we face two familiar problems:
first, the test must insure that the device operates properly over a wide range of environments,
yet it is desirable that the test be performed at
room temperature. Second, the test must not only
reject those which are initially inadequate, but
it must reject those which are predictable future
failures. In addition to these usual requirements, in testing functional blocks a third
problem soon becomes evident: the integrated
logic circuit contains some active and passive
constituents whose individual parameters cannot
be measured, due to the circuit in which they are
incorporated. Further, the device contains
several internal circuit nodes which are not
accessible for measurements once the device is
packaged and complete. In short, in each circuit
some device parameters would be unmeasurable if
we could get to them . • . and we can't.
This, then, is the new problem posed by the
functional-block circuit: the complexity of the
device dictates that no meaningful reliability
tests and evaluation can be performed unless they
be tests of external characteristics. We can no
longer be concerned, for test and evaluation purposes, with the parameters of every active and
passive constituent of the circuit--with the gains
of the transistors, for example--or even, for that
matter, with the distributions of those parameters. We must be able to test a completed integrated circuit by examining only its external

The remarks above are general, readilyapparent comments concerning the entire genus of
integrated function-block devices; the tests and
methods now to be discussed pertain to a particular species, the micrologic element.
Micrologic is a compatible and sufficient
set of integrated semiconductor logic-function
elements. Each micrologic element consists of
from one to five DCTL NOR gates; the circuit
components--resistors and planar transistors--are
diffused into a single slab of silicoq, and metal
intraconnections are deposited on top of the slab.
The device is then packaged in an eight-lead TO-5
or TO-18 package. The logic and schematic diagrams of a typical element, the flip-flop, are
shown in Figure 1.
Effects of Parameter Distribution and Drift
Before the micrologic development program
was undertaken, some assurance was required that
the problem under discussion could be solved.
Since the circuit configuration chosen can tolerate a wide variation of parameters, and since
these parameters might not initially be measurable, assurance was required that those parameters
far from the expected value would not drift with
time, causing subsequent failures. Intuitively,
the inherent stability and reliability resulting
from the planar process gave this assurance.
To investigate these questions, flip-flops
were made up of conventional components, using
the circuit configuration proposed for the future
integrated 'T" element; the transistors used were

76
2.2
incorporates the test principle we have been
discussing. It is a sequence of DC measurements
of the overall input-output characteristics of
the element with no attempt made to measure
internal pa;ameters. It is designed to determine
the present and expected operability of the element.
A go, no-go indication is given for each element,
and all measured data are recorded.

I

@
2

7

6

3

4

5

3.

The LAB TEST is a measurement of all of

the internal device parameters, such as
Rc' that are accessible.

~,

rb,

These parameters are

measured and recorded over the specified temperature range.
Figure 1.
reject devices.
I

Schematic and Logic Diagrams of
the Micrologic "F" Element.
Approximately half were V
and
EBO
The

ALL
pLOGIC
ELEMENTS

SALES

rejects and half were random rejects.

CBO
V
rejects gave a 20% yield of good flip-flops,
EBO
the I
and random rejects gave an 80% yield,
CBO
indicat i.ng that the DCTL circuit configurat ion is

REJECTS

LIFE AND
ENVIRONMENTAL!V::::=::::!!:.-.,Il
STRESS
TESTING

relatively insensitive to wide parameter variations.

One hundred fifty of these circuits were

then placed on 150°C storage test, together with
25 similar circuits made up of good transistors.
After 7000 hours at 150°C ten of these flipflops have failed to operate under load. Seven of
these failures were caused by leads and three by
EB shorts. There were no other modes of failure.
Environmental tests indicate that these failure
modes will not be a problem with the depositedmetal connection method developed for micrologic
elements. The results to date in this life test
give us, therefore, a reasonable degree of ass~r­
ance that transistor parameters which are outsLde
specifications will not drift with time sufficiently to cause operational failure in the micrologic DCTL circuit.
This first-step program of building simulated
micrologic elements thus indicated that, indeed,
individual-constituent parameters are not the
primary criteria for estimates of operability, and
that circuits which are initially good, although
they may have "poor" constituent parameters, will
stay good.
Micrologic Test Program
The test program for micrologic elements is
illustrated by Figure 2, which shows the flow of
completed elements through the three types of
measurement phases:
1. The LOGIC SORT classifies elements as to
type and rejects catastrophic failures. It is a
go, no-go test with no data recording.
2.

The PRODUCTION TEST is the phase that

Figure 2.

Flow Diagram of the Micrologic
Test Program.

It is relatively easy to set up a production
test that evaluates present element operability;
to devise a test that maximizes future operability
over a long time span, and under wide temperature
variations is a more difficult matter. Only with
time can the accuracy and significance of such a
test be verified. For these reasons the life-test
production-measurement loop has been set up, as
shown in gray in Figure 2. Sample quantities of
the elements which are accepted by the Production
Tester are fed into this loop; the elements of the
sample are run through the lab test and all
parameters are recorded. These elements are then
placed on operating and storage life tests. At
given intervals, they are again run through a
production test, samples are again lab-tested,
and the group is returned to the life test.
It is this loop that generates the feedback
information shown in the figure. If the production test indicates a drift with time, the sample
sent through the lab test is increased in size,
and the parameter causing the drift is studied.
If the lab tests, made over the temperature range,
indicate that the room-temperature production test
requires modification, then the production-test
parameters, or criteria, or even test methods are
altered.
The methods, parameters, and criteria of the
production test, then, are initially set up
according to our best first guess. The recirculating life-test, lab-test loop generates the

77
2.2
feedback information that tightens or relaxes the
production test.
As the rate of predictable failures in the
loop decreases, our confidence in the validity of
the production test measurements increases.

Let us now consider in greater detail each of
the phases of the test program.
Logic Sort. The Logic Sorter, shown in
Figure 3, uses a four-stage counter to generate

D U T
OUTPUTIS)

GO

S WIT CHI h G
STANDARD
OUTPUTlS)

L: -"5

TAN-O-A-R-O"

NO - GO

-EL EM-E-N Ts-i

00000;

L _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ .J

Figure 3.

Block Diagram of the
Logic Sorter.

binary inputs which are applied to six "standard"
elements and to the switching network. As the
switch steps through the test sequence, the device
under test is tested first as an F, then as a G,
then as an S element, the correct inputs being
supplied by the switch, and the outputs being
logically compared to the outputs of the corresponding "standard" element.
The need for a logic sort test arises from
the fact that micrologic elements may come from
the assembly process with no classification as to
type. Since the only differences among the processes for the various elements are in the masks
used, no attempt at classification need be made
from the time the wafer is diced. When an element
of unknown type is logic-sorted, the tester steps
through its sequence, and, when a correct comparison with a "standard" is found, the element is
sorted as that type. If the sequence ends without
a comparison, the element is rejected. The sorter
rejects any device that does no~ give a logically
correct output when loaded by a fan-out of two at
room "temperature. Thus it acts both as a convenient sorter and as a measure to prevent elements
with major flaws--such as short-circuits--from
reaching the Production Tester, where they might
cause overloading.
Production Test. As mentioned above, it is
the Production Tester that embodies the principle
of rejecting incipient failures from a series of
measurements of external characteristics. The
tester automatically steps through a different

sequence of tests for each element, setting up
worst-combinations of applied conditions, recording the resultant D.C. measurements, and providing
go, no-go indication and recording.
The establishment of a production test is
essentially the definition of the criteria of what
constitutes a satisfactory device. That is, the
setting of acceptance limits implies that these
limits are the valid and significant criteria by
which one can evaluate the usability as well as
the reliability of the device. For these reasons,
we need first to discuss the factors that govern
our selection of the general criteria of useful~
ness and reliability and then to discuss the DCTL
circuit characteristics and limitations that
determine the specific criteria; then we can
return to discuss the production test conditions
and acceptance limits.
General Criteria

•

The general criterion of acceptance of a
"good" micrologic element is based on the following definition: if any given system operates
correctly, utilizing a given element, then that
element is considered to be good. This definition
is, of course, a very general one, usable only
under certain conditions, namely that "correct
operation" of the system can be clearly defined
and that the set of all circuit environments of
an element in any given system is a limited one,
permitting description of all possible circuit
configurations or of the most stringent configurations.
The nature of the micrologic family is such
that these conditions are met. "Correct operation" of the arbitrary system can be definitely
specified because it is a binary system; "correct
operation" is, therefore, synonymous with errorfree operation.
The set of all possible circuit'environments is known because micrologic elements alone
are used in any micrologic system, and their use
in the system is constrained by the logic design
rules. These design rules specify that no more
than six transistor collectors may be connected
to a given node (i.e., fan-in), and no more than
five micrologic loads may be driven by a given
node (i.e., fan-out). The most severe inputoutput conditions for any element can, therefore,
be. readily derived. In fact, by analyzing several micrologic digital systems, one can determine the frequency-of-occurance distributions of
the many combinations of fan-in and fan-out. A
study is now underway to determine these distributions in a typical small general purpose computer and in several logic subassemblies.
In summary, a micrologic system consists
solely of micrologic elements, and the set of all
possible combinations of these elements is limited
by the design rules, permitting accurate description of the limits of the circuit environment that
will be encountered by any given element. These
factors permit the use, for micrologic, of this

78

2.2
general criterion of acceptance: the element is
good if any arbitrary system utilizing it operates
correctly.

Here transistor Q fans out to N loads, one of
1
which is a simple inverter, and (N-1) of which
are multiple-input gates, all of whose transis-

Direct-Coupled Transistor Logic

tors are in saturation.

The basic circuit, and the basic logic block,
used in micrologic is the DCTL NOR gate, shown
schematically and in logic symbol in Figure 4.

The simple inverter has

the V
/ IB characteristic marked Q on the
BE
2
curves, but the other (N-1) loads have the input
characteristics of Q3' because their collectors
are clamped by the parallel transistors.

+
C = (A+B+---+M)

Ao--_~

For a

given node voltage V , therefore, the inverter
BE
base will draw current IB(Q )' while each of the
2

B

A

M~---1~

Figure 4.

The ~TL NOR Circuit and
Logic Symbol.

The circuit is a simple one, consisting of a
transistor for each input and a single collector
resistor. The output of the gate is a "one" (a
positive potential) if none of the inputs is a
one.
DCTL was selected for use in micrologic
because of its simplicity, low power consumption,
high speed, and insensitivity to wide variation
of certain parameters. One limitation of DCTL is
the requirement for a transistor type which has a
narrow distribution of base current for a given
value of "on" base voltage. This requirement is
necessitated by the "current hogging" characteristic discussed below. One phase of the micrologic program was the development of such a suitable DCTL transistor.
Current Hogging. This euphonious title
applies to the situation shown in Figure 5.

other (N-1) bases draws a far larger current,
IB(Q ). There is then a good possibility that
3
the heavy current drawn by the (N-1) loads will
not permit the voltage V
to rise high enough
BE
to fully turn on Q2.
If, however, the base input characteristics
are closely uniform and if the base input resistance is increased moderately, then the disparity
in input currents is greatly reduced, as shown by
the dotted curves. When these input characteristtcs are incorporated in the DCTL transistor, the
current-hogging problem is minimized, and the
advantages of DCTL are made available.
Specification of Threshold Levels. In the
design of a logic circuit some specification
must be made of the signal levels representing
lIs and O's. In DCTL circuits, a one is represented by a positive potential, the level of
which is determined largely by the bases being
driven. A zero is represented by a near-ground
potential, the value determined by the saturation voltage of the output transistors and the
number of such transistors in parallel.
The degree of saturation in the "on" condition and the permitted amount of collector current in the "off" condition are, within limitations, up to the designer, and there are many
combinations of "one" and "zero" voltages that
might appear to be usable. Just what are the
limit at ions?
In DCTL, we are concerned, of course, with
a distributed system, consisting entirely of
closely similar gates. When we determine the
limiting values that the signal levels may take,
the solution must hold over every logic path in
any system.

OFF

\

V

I !l£ '----ISa-

L_!... __ _

Figure 5.

The "Current-Hogging" Condition.

The problem to be solved can be considered
in this manner: assuming an infinite cascade of
DCTL logic stages, what is the lowest "turn on"
voltage and the highest "turn off" voltage that
may appear at any base, under the condition that
all following stages in the cascade are alternately on and off. Considering Figure 6, what
is the highest voltage that we may apply to the
base of Q and still have the condition that Q
O
O

79
2.2
as a function of V
for a representative group of
BE
DCTL transistors with collector resistors; the
upper and lower limits are the 30' points of the
distribution.

The two-transistor feedback loop of

Figure 7 simulates the "infinite" cascade; transistor A has a VCE characteristic corresponding to the
upper 30- curve, and B is represented by the lower
Figure 6.

Cascaded DCTL Inverters.

and all even-numbered transistors are cut-off,
and that Q and all odd-numbered transistors are
I
on? Similarly, what is the lowest base voltage
{

applied to Q that insures that Q and all evenO
O
numbered devices are "on" sufficiently to hold

curve.
Assume initially that A is cutoff, B is
conducting and that the feedback loop is open; a
trial turn-on voltage, VBE(A)' applied to A will
result in VCE(A) at the collector of A and at the
base of B.

For VBE(B)

= VCE(A) ,

the output of B

is seen to be VCE(B) , which is lower than the

the odd-numbered devices cut-off?

initial assumed applied voltage.

This problem can be graphically analyzed by

The applied

voltage was not sufficient, then, to

s~itch

the

considering the circuit of Figure 7 together with

stable states had the feedback loop been connected,

the curves of Figure 8.

and it would not have been sufficient to drive a

The curves indicate VCE

long cascade.
In Figure 9 let us assume a higher turn-on
40~----r---r----r---"

Figure 7.

Two-transistor Feedback Loop.

40
3.0

I...J

~~

20
1.5

\



«

~

1.0

""

07

0
I-

t®VCE(Al

\

05

::-

03

~
\

\

0.4 JVCE(Bl_
f----

r\ "- ~ I' "---

VBE(B1d

05

0.6

0.7

B

~BE(Al

08

05

06

07

08

09

VBE - BASE VOLTAGE - VOLTS

\

I

02

I-I--

VBE

\

\

0

u

-

...

veE

' 1\

\

U

...J

Re

\"

0

::-

~

0.9

VBE - BASE VOLTAGE - VOLTS

Figure 8. Curves Showing 3a Limits of the
Distribution of VCE as a Function of V
for a
BE
Representative Group of DCTL Transistors with
their Collector Resistors. The First Trial
Solution for the Threshold Value of V
is also
BE
Shown.

Figure 9. Curves of VCE vs. V with
BE
Second Trial Solution (Dashed Lines) and Threshold Operating Point (Dotted Lines).
voltage VBE(A) for the same circuit.
the same steps (by proceeding from

CD

Repeating
to

® ),

we find that the resultant feedback voltage is
higher than the assumed turn-on voltage, so that-had the loop been connected--the voltage VBE(A)
was more than sufficient to insure a reversal of
stable states.

This solution is indicated by

dashed lines in Figure 9. By repetition of this
graphic procedure for intermediate values of

80

2.2
turn-on voltage, we can find the threshold value,

in the circuit shown, the base A wili not allow

VBE(on)' which is just sufficient to turn on A

VCE(B) to rise to the value shown on the graph,

enough that B is sufficiently cut off to cause its

but will clamp it to some lower value, VBE(on)A'

collector to rise to VBE(on)'

There is, however, useful information to be gained

This threshold

condition is shown by the dotted lines in the
figure.

from the value of VCE(B); it indicates the collector current

Since A has the poorest turn-on characteristic, we see that this is a worst-case condition
within the 3 (J" distribution and that all other
transistors are turned on harder than A for the
same minimum-required VBE(on)'

drawn by transistor B with VBE(B)

applied to its base.
I CEX

6V

=R

This current,

(See 1~ of' Figure 9)

c

Selection of this threshold value of turn-on

I LOAD

voltage also implicitly determines the maximum

Rc

voltage which may appear at the base of an off
transistor.

By selecting VBE(on) we have defined

the "on" condition as one in which VCE is VCE(on)
or lower, as shown in the "on" VCE region in
Figure 9.

(b)

Since this voltage is equal to the base

(a)

voltage of the following stage, we see that the
base voltage of an off transistor must be VBE(off)

Figure 10.

(a)

or lower.

(b)

These two values, then, VBE(on) and VBE(off) ,
define the threshold levels for the on and off
transistors in the feedback loop or in the cascade.
These same values would be obtained if we assumed
an initially-applied voltage to turn off B.

The Graphical Solution Considering Loading.
Consider again the two-transistor feedback loop
and the VCE vs. V curves of Figures 9 and 10.
BE
This time the output node of B has a current load.
We again assume the trial turn-on voltage, VBE(A)'
ing points

CD

through

To a first approximation, then, we can
consider transistor B to be a current generator,
drawing current I

from its output node, as
CEX
indicated in Figure lab. In order that the

assumed voltage VBE(A) be sustained while a load

All those readers with experience in logic
circuitry are now entitled to rise up in wrath,
shouting, "What, no noise rejection?" and "What
happened to fan-out!" together with mutterings of
"What did I tell you about DCTL!" And this would
be justified. The graphic solution above is a
preliminary one; it is for the purpose of giving a
basic understanding of the method of analysis of
cascade of DCTL networks by consideration of the
device parameter distributions. It does not consider loading effects or the necessity for noise
rejection. These we will discuss next.

which is higher than the threshold value.

Two-transistor Loop with
Load.
Equivalent Circuit for B.

Follow-

@ in the figure, we see

that the assumed applied voltage results in the
appearance of VCE(B) at the collector of B.

current is drawn from the B output node, the
following is true:
V

Iavailable

=

cc - VBE(A) _ I
R
CEX'

(1)

c

where Iavai1able is the current available from
node B to drive the base A as well as other
unspecified current loads at a voltage VBE(A)'
This current is easily calculated.
If this available load-driving current is
plotted as a function of V , we get a plot
BE
similar to the one of Figure 11, where the distribution of IB as a function of V
is also shown.
BE
Iavailable in this plot is the current available
from node B at a given voltage V , with the same
BE
voltage V applied at the base of A. It is seen
BE
that as the operating V point is reduced from
BE
a high value, the current-available gradually
rises, since more current can be supplied through

Now these VCE curves were plotted for an
unloaded collector node; it is quite apparent that,

Rc to a lower voltage.

As the VBE(on) operating

point is approached, however, the "leakage"

81

2.2
current drawn by the "off" transistor, B, becomes
dominant, and lavailable falls extremely rapidly.

acceptable "on" value of V
is established there.
BE
This value is defined as Yon; it, in turn,
specifies a maximum acceptable turn-off value of

To recapitulate, Yon' the minimum allowed
base "on" voltage, is the voltage required to turn
on all transistors sufficiently well that all
collectors are

Voff

or lower; this voltage is, in

turn, low enough that all transistors to which it
is applied have VCE equal to or greater than Yon
0:

\

\

I

/"

~ o7~------r---~~\--;~~'"/~I'~--~~
...J

8
I

V·

I \{

\

05

V)(

05
04

"1"

1

/I~ :

while supplying current to drive the test04

I'---~

..

~

specified fan-out.
Note that, whereas Yon is the voltage required

03

2~------~--~/~/_'~:1~-4'~~~~~--~02

//
05

06

07

to turn on a "poor" inverter, the loads

! :

08

~

in

equation (2) are those required by transistors with

09

clamped collectors.

VBE - BASE VOLTAGE - VOLTS

Figure 11. Collector Voltage, Base Current,
and Current Available as Functions of Base Voltage
for a Group of DCTL Transistors.

We have, therefore, based our

calculations on the most stringent DCTL fan-out
configuration.
In summary, we have graphically analyzed the

Knowing Iavailable and lBE as functions of
' we can quickly determine fan-out as a function

V
BE
of V ' Clearly, for any operating point of V
BE
BE
the number of transistor bases that can be driven
by a node is equal to the total current available
divided by the current required by each base.

fan-out as functions of V • By establishing a
BE
minimum turn-on voltage and a maximum turn-off
voltage which may be allowed to appear at a
transistor base, we have insured that the calcu-

Fan-out, N, can therefore be expressed by
equation (2):1
N = lavailable

distributions of VCE and IB as functions of V
in
BE
order to determine the available current and the

lated value of fan-out will be achieved at every
node.
(2)

where IB is the upper 3 cr value of base current at
the given V • Since this function has a maximum
BE
in the neighborhood of the maximum value of
lavailable' the specified value of the minimum
lA more accurate value for fan-out is
obtained by using the upper 3 cr- limit of the
distribution of N(l ) , as in equation (3). Here
B
N is an implicit function of Iavailable'

The methods of calculation of current available and fan-out discussed above are largely of
academic interest. They are included here to
give a better understanding of the tests we are
about to discuss and to illustrate the evaluation
procedures used during the development of the
micrologic DCTL transistor. They are useful for
evaluation purposes, but they are not used to set
test acceptance limits. The acceptance of an
element is based upon measurements which give
direct indication of sufficient current available
rather than upon calculated values.
The Production Tester

(3)

where N is equal to fan-out, iB is the mean value
of base current, and ~B is the standard deviation of the base current distribution. This gives
a higher value for N; the use, therefore, of the
simpler equation (2) results in a conservative
estimate of fan-out.

As can be seen from the DCTL circuit considerations and from the foregoing graphical analysis,
an adequate external production-test of micrologic
elements must consist of three basic types of
measurements: input current, output current
available, and output VCE(sat)'
Definitions. The terms to be used in
describing these tests are defined as follows:

82
2.2

Yon

1)

is the highest turn-on base voltage

which will appear in any system.

It is equal to

Vcc applied through the lowest value of collector
res istance.

mente

V
). This is the other output measureCE(sat
With Yon applied to the base of an output

transistor, the collector must pull down below

Yon is the lowest turn-on base voltage

2)

voltage appears at the base.

allowed to appear in any system.
the same manner as the V

-on

It is derived in

V
• If the base is not directly accessible,
sat
then the worst-combination of inputs is applied
such that the lowest possible internally-generated

that we discussed

earlier.

turn-on voltage appears at the base.

3) V
is the highest value of turn-off
off
base voltage allowed to appear in any system. It

The loading conditions and the acceptance
limits for these tests are set above those
required to meet the specified fan-out of the
elements. The margin of excess is one which has
been empirically determined to insure operation
with noise rejection over the specified temperature range.

is set high enough to permit an I
comparable to
CEX
that which will occur at high temperature.
4) Y

is the collector voltage applied
off
during measurement of I B• It is the lower limit
of the voltage appearing at the collectors of
three paralleled saturated transistors.
Vsat is the highest acceptable VCE(sat)
for a transistor driven by Yon at room temperature.
5)

It is appreciably lower than

Voff •

With these definitions we can define the three
types of tests performed by the production tester.
Input Current.

With applied voltages of Yon

at the base and Yoff at the collector, the base
current drawn by any transistor may not exceed a
maximum value, defined

as~.

See Figure 12.

This measurement is performed only on bases which

Once the general test methods and criteria
are established, the only remaining trick is to
determine the worst-combination of conditions to
apply at the other pins when performing a measurement at any given pin. Each element presents its
own problems and requires a separate solution;
the simplest element, "G," requires seven test
steps; the most complex element, "S," requires 16.
It might be helpful to discuss some of the
general considerations involved in the devising
of the tests. The tests of input characteristics
are rather straightforward, since only transistor
bases are connected to the input pins; the only
problem, is to insure that the most severe
collector-clamping conditions prevail. In the
case of the "G" element, where the collector node
is available and could conceivably have additional
transistors connected in parallel, Y
is
off
applied directly. See Figure 14. In the case of

12912

are connected to micrologic element input
terminals.

~

~ON

~OFF

VCE
( MEASURE)

7

Figure 12.
Input Current.
Current Available.

5

Figure 14.
Figure 13.
Current Available
This is a measurement of

element output characteristics.

With

Yoff

applied to the base of an output transistor, and
with current IK drawn from the collector node,
the collector node voltage must exceed Yon
Figure 13.

8
6

1;

See

If--as is often the case--the base is

not accessible, then a combination of inputs is
applied such that the highest possible turn-off

3

4

7

3

6

4

5

Logic Diagrams of the "G"
and "s" Elements.

the "s" element, the input-gate collector nodes
are not accessible, and the worst possible clamping for input current at pins 6 and 8, for
example, occurs when pin 7 is driven by V
on
For current-available measurements at output
terminals, a greater number of variables must be
considered. All transistors whose collectors
are connected to the given terminal must be
turned off as weakly as possible in order to draw
as much leakage current as will ever be experienced; all transistors whose bases are connected
to the given terminal must have the worst possible

83

2.2
collector-clamping in order that the bases will
draw the maximum current. Then, with as much
current as possible being internally drawn from
the node, the external load current is drawn, and
the node voltage checked to insure that it is
higher than Yon'
If a single transistor drives an output node,

TABLE I
Measure
&

Compare

Remarks

1

Yoff at pin 4
Yon at pin 6

Check input
current, Q
l

2

Yoff at pin 2
V at pin 8
-on

Check input
current, Q

the test of VCE(sat) is made by turning it weakly
on with Yon and measuring VCE'

Apply

Test

If several transis-

tors in parallel drive a node, then one is turned
weakly on with Yon and the others are turned off
with Yoff; the test is then repeated until each

4

Yon at pins 6
and 8

V4 < V
Check V
sat
CE (sat) ,

at pins 6
4 -on
V

Check V
V2 < V
sat
CE(sat)'

3

transistor has demonstrated its ability to hold

Q
l

the node voltage sufficiently low.
and 8

Test Example
As an example of the application of these

5

Yoff at pin 4
at pin 8
V
off
Sink IK at pin 2

6

Remove Y
at
off
pin 4;
other inputs
remain

general considerations, let's look at a specifiq
micrologic production test, the test of the "F"

Q4

V 2> Yon

Current available
at pin 4, QZ
clamped

element.
As summarized in Figure 15 and Table 1, the
first two steps of the "F" test are measurements of

V

4

< Vsat Check VCE (sat) ,
QZ

the input base currents with the specified Yon
applied at the base and with the collectors
clamped.

7

Same as test 6

Current available,
Q not clamped
2

The currents are measured, the values

recorded, and comparison made against the maximum

8

~ff

at pin Z

V4

> Yon

V

off at pin 6
Sink IK at pin 4

acceptable value, IB'

Current available
at pin 2,
Q clamped
3

+ 3V
9

10

Current available,
Q not clamped

Remove Yoff'
all others same

V >V
4
-on

Same as test 9

V 1e explain phenomena by reducing
them to other phenomena that seem to us,
someh:)w, simpler and more orderly. How
did I'~endel, for example, explain the
relative frequencies of his different
kinds of peas in successive generations?
He postulated (without any direct
observational evidence) underlying
dominant and recessive factors passed
on from parents to their progeny, whose
interaction determined the physical
type of the progeny. Only lTIany years
later vms any direct evidence obtained
of microscopic structures in the cell-the chromosomes--that could provide the
biological substrate for Mendel's
, factors.
Again, rlJ.organ's studies of
fruit fly populations led him to
postulate even tinier c·omponents of the
chromosomes--the genes. These had to
a',"lai t the electron microscope before
they could be shown, by direct observation, to exist; and even today, we
are still far from an explanation of
these biological structures at the next,
biochemical level.
f'

The goal, then, in Simulating complex human behavior is the same as the

Finally, when vie use computers to
state and test information processing
theories of thinking, we do not postulate
any crude analogy between computer and
brain. We use the computer because it
is capable of Simulating the elementary
information processes that these theories
postulate as the bases for thinking.
We do not assert that there is any resemblance between the electronic means that
realize these processes in the computer
and the neurological means that realize
the corresponding processes in the brain.
We do assert that, at a grosser level,
the computer can be organized to imitate
the brain.
Information

Proce~~ing Lan~~ages

There has been a strong, and not
accidental, interaction between work on
the computer simulation of human thinking
and research on computer programming.
T~e kinds of processes that computers
are called upon to perform when they are
Simulating thinking tend to be quite
different from the processes they perform
when they are carrying out numerical
analyses. A superficial difference is
that the former processes involve little
or no use of arithmetic operations. A

114
3.1

more fundamental difference is that
memory must be organized in quite
distinct ways in the two situations.
Within the past five years there
have been a number of reports to these
conferences on the general characteristics and specific structure of information processing languages specially
designed to facilitate non-numerical
simulation. 15 ,lb I shall not go over
this familiar ground again, except to
point out that when such languages are
used to build psychological theories the
languages themselves contain implicit
postulates--although rather weak ones-about the way in 1-'ihich the central
nervous system organizes its work.
One of the common characteristics
of all of these languages is their
organization of memory in lists and list
structures. By this means there can be
associated with any symbol in memory a
"next symbol--the symbol that follows
it on the list to which they both belong.
By the use of a slightly more complicated
device, the description list, there can
be associated with any symbol in
mem~ry a list of its attributes and
their values. If the symbol, for example,
represents an apple, we can store on its
description list the fact that its color
is red, its printed name is APPLE, and
its spoken name, APUL. The incorporation of these two forms of association-the serial order of simple lists and
the partial ordering of description
lists--in information processing
languages permits one to represent many
of the associative properties of human
memory in a quite simple and direct
way. We can use simple lists to simulate
serial memory--e.g., remembering the
alphabet--and description lists to
simulate paired associations--e.g., the
association between an object as recognized visually and its name.
11

A characteristic of the list
processing languages, which they share
with most other compiling and interpretive languages, is that they organize
behavior in hierarchical fashion.
Routines use subroutines, which have
their own subroutines, and so on. This
characteristic of the languages again
facilitates the construction of programs
to simulate human behavior, which
appears to be organized in a highly
similar hierarchical manner. The ·fact
that most investigators have found it
easier to write simulation programs in
interpretive list languages than in
machine language derives, in all likelihood, from the fact that the former

languages have already taken the first
steps in the direction of organizing the
computer processes to mirror the organization of the human mind.
Heuristic Problem SolVing Programs
The Program of Selfridge and Dinneen
The work of Selfridge and Dinneen
on pattern recognition,7 which I
earlier assigned to the se.cond category
of simulation programs--simulation of
sensory-perceptual processes--really
marks a transition to information
processing simulations. The SelfridgeDinneen program specified a set of
processes to enable a computer to learn
to discriminate among classes of patterns
presented on a two-dimensional Hretina.1!
The patterns could represent, for example,
English letters like nA· and "0 of
varying shape, size, and orientation.
II

In the Selfridge-Dinneen program,
recognition was accomplished by using
various operators to transform the
retinal stimuli--in general to simplify
and I'stylize" them--and then searching
for characteristics of the transformed
stimuli that grouped the various exemplars
of a given alphabetic letter together,
but separated the exemplars of different
letters. Although the program made use
of the arithmetic instructions of the
computer, the operations were basically
topological and non-numerical in nature.
Appropriate organization rather than
rapid arithmetic was at the heart of the
program.
The Selfridge-Dinneen program foreshadowed subsequent work in this area in
another important respect also. The
characteristics used to distinguish patterns were heuristic. They amounted to
rules of thwnb, selected by the computer
over a series of learning trials on the
sole basis that they usually worked-that is, made the desired discriminations.
In more traditional uses of computers it
is usually required that the programs be
algorithms--that they be systematic
procedures which guarantee" solution of
the problem to a desired degree of accuracy. The heuristics generated by the
pattern recognizing program provided no
,such guarantees. Since there are vast
ranges of tesks, handled every day by
human beings, for which no algorithms in
the sense just indicated are knoitm to
exist, the admission of heuristics as
program components opened the vlay to
simulating the less systematic, but often
effective, processes that characterize
much garden-variety, everyday human
thinking.

115
3.1

Subsequent work has tended to
confirm this initial hunch, and to demonstrate that heuristics, or rules of
thumb, form the integral core of human
problem-solving processes. As we begin
to understand the nature of the
heuristics that people use in thinking,
the mystery begins to dissolve from such
(heretofore) vaguely understood processes
as "intuitionl. and 11 judgment •
t,

Some Other

Problem-Solvin~ Pr~~rams

In the period 1956 to 1958 there
came into existence a number of other
compute~ programs that accomplished
complex tasks with a "humanoid" flavor:
composing music,17 playing checkers,18
discovering proofs for ~heorems in
10gic,19 and geometry,20 designing
electric motors and transformers,2l playing chess,22,23,24 and balancing an
assembly line. 2 5 The primary goal in
constructing most of these programs was
to enable the computer to perform an
interesting or significant task.
Detailed simulation of the ways in which
humans perform the same task was only a
secondary objecti ve--or vIas not
considered at all.
Nevertheless, it was discovered
that often the best program for doing
the job was a program that incorporated
some of the heuristics that humans used
in doing'such jobs. Thus, the music
composition program of Hiller and
Isaacson made use of some of the rules
of classical counterpoint; the motor
design programs and line balancing
program were generally organized in much
the same ways as the procedures of
experienced engineers, and so on. Hence,
to a greater or lesser degree, all of
these programs have taught us something
about the ways in which people handle
such tasks--especially about some of the
kinds oi' heuristics they use.
Among these programs Samuel's
checker program and the Los Alamos chess
program place the least emphasis on
heuristics, and hence provide valuable
yardsticks for comparison with heuristic
programs handling the same, or similar
tasks. These two programs make essential
use of the computer's capabilities for
extremely rapid arithmetic, for their
basic strategy is to look at all possible
(legal) continuations of the game for
several moves ahead, and then to choose
that move which appears most favorable
(in a minimax sense) in terms of the
possible outcomes. In contrast,
Bernstein's and the NSS chess programs
examine a small, highly selective subset
of all possible continuations of the

game and choose a move that appears good
in the light of this selective analysis.
Thus, the Los Alamos program, looking two moves ahead, will typically
examine a little less than a million
possible continuations, Bernstein's
program approximately 2,500, and the NSS
program almost never more than one hundred
and more usually only a handful. All
three programs play roughly the same
quality 01 chess (mediocre) t'J'ith roughly
the same amount of computing time. The
effort saved by the heuristic programs
in looking at fewer continuations, is
expended in selecting more carefully
those to be examined and subjecting them
to more thorough examination. Thus, the
more syste~atic, arithmetic programs
provide benchmarks agains t ~lhich the progress in developing heuristics can be
measured.
The General Problem Solver
All of the programs we have
mentioned fell short of human simulation
in one very fundamental respect--apart
from failures of detail. They were all
special-purpose programs. They enabled
the computer to perform one kind of
complex task, and one kind only. Only
in a few cases (the Checker Player 18 and
the Logic Theorist 26 ) did they enable
the computer to improve its performance
through learning. Yet we know that the
human mind is (a) a general-purpose
mechanism and (b) a learning mechanism.
A person ~'lho is brought into a relatively
novel task situation may not handle the
situation with skill but, unless it is
inordinately difficult, will not ftnd
himself at a complete loss. Whether he
succeeds in solving the problem that is
posed him, or not, he is able, at least,
to think about it.
~ve must conclude that if a computer
program is to simulate the program that
a human brings to a problem Situation,
it must contain t'i'l0 components: (a) a
general-purpose thinking and learning
program that Qakes no direct reference
to any particular task or subject
matter; and (b) heuristics that embody
the specific techniques and procedures
which make possible the skilled and
efficient performance of particular
classes of tasks. The program must
incorporate both general intelligence
and special skills.

The General Problem Solver (GPS)
vms the first computer program aimed at
describing the problem solving techniques
used by humans that are independent 01'
the subject matter of the problem. 2 7

116
3.1

Since GPS has been described elsewhere,
I shall say only a word about its structure. It is a program for achieving
the goal of transforming a particular
symbolic object (representing the "given
problem situation) into a different
symbolic object (the l'desired" situation
or goal situation). It does this by
discovering differences between pairs
of objects, and by searching for
operators that are relevant to reducing
these differences. In the form in
which i,t has thus far been realized on
a computer, GPS is not a learning program, hence still falls far short of
Simulating all aspects of what we would
call general intelligence.

ll

In its current computer realization,
GPS has solved some simple problems of
finding proofs for theorems in symbolic
logic (substantially the same task as
that handled b¥ the special-purpose
Logic Theorist). It has solved the wellknown puzzle of Missionaries and
Cannibals--finding a plan for transporting three missionaries and three
cannibals across a river without any of
the missionaries being eaten. Hand
simulation has demonstrated that it can
handle trigonometric and algebraic
identities. On the basis of other investigations that have not fully reached
the programming stage, it appears highly
likely that GPS will be able to solve
certain tactical problems in chess (e.g.,
to find a move leading to a fork of a
pair of enemy pieces), do formal
differentiation and integration, and
write codes for simple computer programs
in IPL V. Several possibilities for incorporating learning processes in GPS,
one of them using GPS in the learning
mechanism itself,2e have also been
explored.
The adequacy of GPS as a simulation
of human problem solving has been
examined, primarily in the task domain
of symbolic logic, by comparing the
computer trace with the thinking-aloud
protocols of college students solving
identical problems. 2 9 The evidence to
date suggests that GPS does indeed capture
the principal problem-solving methods
used by the human subjects. The
detailed comparison of its behavior with
the protocols has cast considerable light
on the processes of abstraction and on
the nature and uses of imagery in
problem solving.
Recent Advances in the
Simulation of Thinking
The remaining papers to be presented
in this session will describe a number of

heuristic programs that have been written
in the past two years, and which extend
very substantially the range of human
mental processes that have been simulated with these techniques. I shall
not anticipate the content of these
programs, beyond indicating what their
relation is to those I have already
mentioned.
Areas of

pSY~Eo~~gical Experime~tation

The simulations mentioned so far
all fall in the area that psychologists
call "higher mental processes.
As I
indicated earlier, these processes have
tended to be underemphasized in American
experiment~l psychology until quite
recently because we did not have tools
for investigating them in an objective
and rigorous way. If computer simulation
has shown itself to be a powerful tool
of research in an area as difficult
as the study of higher mental processes,
we might expect this tool to prove even
more powerful if applied to the simpler
phenomena with which experimental
psychologists have been largely concerned. The papers of this session
report some of the first evidence that
this expectation is justified.
1,

What are the kinds of tasks and
processes that have been most thoroughly studied by psychologists? Perception--the interaction of sensory
organs and central nervous system in
the discrimination and recognition of
stimuli--has been the subject of extensive investigation. A second, very
active, research area has been learning,
and particularly the rote learning of
serial material and of stimulus-response
pairs. A third area has been simple
choice behavior, especially choice among
a small number (usually two) of alternatives with systematic or intermittent
reward. Animal and human maze learning
experiments have been used to study
both rote learning and simple choice
behavior. Finally, there is a rather
varied assortment of work that is
usually classified under the heading of
"concept formation or "concept attainment."
II

No one supposes that the topics
I have mentioned--perception, rote
learning, simple choice behavior, maze
learning, and concept formation--are
mutually exclusive and exhaustive
categories. They are simply pigeon
holes that psychologists have found
convenient for classifying experiments.
It is almost certain that the mechanisms
required to perform tasks in one of
these areas are called into play in some

117
3.1

of the others. Hence, we would have
reason to hope that as heuristic programs
are constructed to handle one or another
of these tasks, the mechanisms employed
in the several programs will begin to
show distinct resemblances--and resemblances also to the mechanisms used in
problem-solving simulations. Such
resemblances and common mechanisms are
already beginning to appear.
Long-Range Goals of Simulation
The long-term research strategy
would again be gradually to replace a
multitude of special-purpose programs
with a more general program aimed at
simulating the whole man--or at least
the cognitive aspects of his behavior.
Although enormous gaps of ignorance still
separate us from that goal, the goal
itself no longer seems entirely Utopian
to the active researchers in the field.
Perhaps the largest single gap at
present--and one that is not filled by
any of the work to be reported today--is
in programs to explain long-range human
memory phenomena. I will venture the
personal prediction that filling this
gap will soon become crucial to progress
in the whole field of information
retrieval.
Another important gap that also has
significant practical implications lies
in the area of simulation of natural
language processes. Here, interest in
language translation and in the improvement of computer programming languages
has already led to exciting progress--as
illustrate~~ for example
by the work
of Chomsky
and Yngve.3 i
Heuristic Programs in New Areas
The areas of rote learning, simple
choice behavior, and concept attainment
are represented in the programs to be
described by Mssrs. Feigenbaum, Feldman,
and Hunt, respectively.
Rote Learning. The Elementary
Perceiver and Memorizer (EPAM) is a
theory to explain how human subjects
store in memory symbolic materials that
are inherently meaningle·ss.
The
typical learning materials are "nonsense
syllables --spoken or printed syllables
that do not correspond to English words.
By studying rote learning, we hope to
tmderstand, for example, how humans
learn to associate names with objects,
and learn to read by associating printed
words vIi th their oral counterparts.
I

t,

I,

Binary Choice. In the so-called
partial reinforcement or binary choice
experiment, the subject is instructed
to guess which of two events will occur
next. In variants of the experiment,
the actual event sequence may be
patterned, or it may be a random sequence.
The binary choice experiment has been
one of the principal situations used to
test the stochastic learning models
that have been develo~~d in psychology
over the last decade. ,32 Mr. Feldman's
Binary Choice program offers an alternative theory to explain these phenomena,
hence provides an interesting example
for comparing and contrasting heuristic
programs with more traditional mathematical models.
Concept Formation. In the simplest
form of the concept formation task, a
rat is given a choice of two gates,
one of which is labelled, say, with a
large triangle, the other with a small
circle. If the experimenter's aim is
to test the rat's attainment of the
concept "triangle, he places a reward
behind the gate labelled with the
triangle. On succeeding trials, the
symbols change in shape, size, or color,
but the gate labelled with a triangle
always leads to the reward. Within the
past year, several computer programs
have been written that simulate slightly
more complex concept learning behavior
in humans. One of these programs, the
Concept Learner, will be described
by Hovland and Hunt.
11

Conclusion
I have tried to outline the development over the past decade of the use
of computers to construct and test nonnumerical information-processing
explanations for human thinking and
learning. Such programs, which are
beginning to be validated by behavioral
evidence, are providing embryonic
theories for these phenomena in terms
of underlying information processes.
Hopefully, the elementary information
processes that are postulated in the
theories will, in turn, find their
explanation in neurological processes
and mechanisms. The papers in this
session describe a few of the programs
of this kind that have been constructed
to date, and provide some basis for
judging the prospects for this approach
to understanding the human mind.

118
3.1

References
1.

\tva 1 ter, Grey, The Living Brain,
Van Nostrand, 1953.

13.

2.

Ashby, W. R., Design for a Brain,
Wiley, 1952.

14.

3.

Fatehchand, R., Machine recognition
of spoken words, in F. L. Alt,
ed., Advances in Computers,
Academic Press, 1960, pp. 193-231.

Bush, R. R. and F. Mosteller,
Stochastic Models for Learning,
vliley-; 1955.

15.

Newell, A. and J. C. Shaw, "Programming the Logic Theory Machine,1t
Proceedings of the 1957 Western
Joint Computer ConTerence, pp. 230-

4.

s.

Fo,;rgie, J. vJ. and C. D. Forgie,
'Results Obtained from a Vowel
Recognition Computer- Prpgram,1l
Jour. of the Acoustical Society
orAinerica,-'3I: 1480-1489 ( November,
1959) .
Gold, B., lMachine R2cognition of
Hand-Sent Illor-se Code, IRE Trans.
on Information Theory, IT-);----1;-1: 17 -21+ (March, 19)9).

1 9 ) ) . - - - ---

21+0.----------------

16.

Shaw, J. C., A. Newell, H. A. Simon,
and T. O. ElliS, "A Command
Structure for Complex Information
Processing, 1, Proc. of the 1958
Western Joint Computer--Con-:ference,
pp. 119-T2IT ;--------- --_.----

17.

Hiller, L. A. and L. M. Isaacson,
Experimental t,1usic, IV!cGraw-Hill,
19:59-:-------~------ ,,"----

18.

Samuel, A. L., 'Some Studies in
rllachine Learning using the Game
of Checkers," IBN J. of Research
and Development;--J:2W":229

11

6.

--(.

8.

Blair, C. R., " On Computer Transcription of f'lanual Horse,' Jour.
of the Association 1')r Computing
i',lachinery, O:L:r~9-44c-rJuly, -19)9) .
Sel.2ridgG, O.G., "Pattern Recognition
and Nodern Computers! and
Dinneen, G.P., "Programming Pattern
Recognition," Proc. of the 19;))
Western Joint Computer Conrerence,
pp :- 91-10Cr:--------·--------Farley, B. G. and \<1. A. Clark,
'Simulation of Self-organizing
Systems by Digital Computer, IRE
Trans. of the Professional Group
on Information ThErory;--P-UIT-::-rr:7O.::m--n-ep-temo er-;--195 11- ) •
l'

9.

10.

Clark, W. A. and B. G. Farley,
'Generalization of Pattern Recognition in a Self-organizing System,'
Proc. of the 1955 Western Joint
Qom:t>~~e~ --cr®1~r-eriEe-;-pp. 86-91.
Rochester, N., J. H. Holland, L. H.
Haibt and 1.1f. L. Duda, Tests on a
Cell Assembly Theory of the Action
of the Brain Using a Large Digital
Computer,' IRE Trans. on Information Theory, IT-2;--ifj-;------(Septemoer,-I9j6) .

11.

Hebb, D.O., Organization of
Be~avi~E.' Wiley-;- 1949.

12.

Rosenblatt, F., "The Perceptron: A
Probabilistic Hodel for Information
Storage and Organization in the
Brain,' Psychological Review,
65: 386-40tr-(1'fovemb-er, 1958).

Johnson, D. Iv'l., The Psychology of
Thought and Judgment, Harper,

-

rjU1.y,---r~r)97-.

19.

Newell, A. and H. A. Simon, lThe
Logic Theory lVlachine, IRE
Trans. on Information TheorY
IT-2rff~"-pp:--6r::rr9-;-\ September,
1956) .
l'

20.

Gelernter, H. and N. Rochester,
"Intelligent Behavior in ProblemSolving Machines, I, IBr~l J. of
Research and Develof)me-n~~:2!336yr~-CocTc:iOer-;-T9~8) .

21.

Goodvlin, G. L., 'Digital Computers
Tap Out Designs for Large Motors
Fast,' PONer, April, 1958.

22.

Kister, J., P. Stein, S. Ulam,
W. Walden and M. \tJells, 'Experiments in Chess,' Jour. of the
Association for Computlngl%aChinery,

1f:-rrrr-r7r--(Ap-rlT; -1.9-57T:--

23.

Bernstein, A., M. de V. Roberts,
T. Arbuckle and IV!. H. Belsky,
'-A Chess-Playing Program for
the IBM '704, Proc. of the 1958
'IlJes tern Joint CompuTer--Conf'erence,
pp. 1)7-=T5Sr:----------- -------- - l'

24.

Newell, A., J. C. Shm·J and H. A.
Simon, "Chess Playing Programs and
the Problem of Complexity," IBIvl J.
of Research and Development, 2:320335 CO-C-£05e-r-;--T958T.-----

119
3.1

25.

Tonge, F. M., A Heuristic Program
for an Assembly Line Balancing
Problem, Prentice-Hall, forthcoming.

26.

Newell, A., J. C. Shaw and H. A.
Simon, "Empirical Explorations of
the Logic Theory Machine, Proc.
of the 1957 Western Joint Computer
Conference, pp. 218-230.
II

27·

Newell, A., J. C. Shaw and H. A.
Simon, itA General Problem-Solving
Program for a Computer, Computers
and Automation, 8:10-17 (July,
1959) .
II

28.

Newell, A., J. C. Shaw and H. A.
Simon, A Variety of Intelligent
Learning in a General Problem
Sol ver, in r,L C. Yovi ts and S.
Cameron, eds., Self-Organizing
s§stems, Pergamon, 1960, pp. 1531 9.
II

I,

29·

Newell, A. and H. A. Simon, "The
Simulation of Human Thinking,1I
in Current Trends in Psychology,
1959, U. of Pittsburgh Press, 1960.

30.

Chomsky, A. N., Syntactic Structures,
Mouton, 1957.

31.

Yngve, V., A Model and an Hypothesis
for Language Structure," Proc.
of the American PhilosophICaI
Society, 104:444-466, October,
1960.

32.

Suppes, P. and R. C. Atkinson,
Markov Learning Models for
Multiperson Interactions,
Stanford U. Press, 1960.

II

121
3.2

THE SIMULATION OF VERBAL LEARNING BEHAVIOR*
E. A. Feigenbaum
University of California
Berkeley, California
and
The RAND Corporation
Santa Monica, California
Summary

An information processing model of

symbolic learning is
given a precise statement as a computer
program, called Elementary Perceiver
and Memorizer (EPAM). The program simulates the behavior of subjects in experiments involving the rote learning of
nonsense syllables. A discrimination
net which grows is the basis of EPAM's
associative memory. Fundamental information processes include processes for
discrimination, discrimination learning,
memorization, association using cues,
and response retrieval with cues. Many
well-known phenomena of rote learning
are to be found in EPAM's experimental
behavior, including some rather complex
forgetting phenomena. EPAM is programmed
in Information Processing Language V.
elementary~human

of an attempt to state quite precisely
a parsimonious and plausible mechanism
sufficient to account for the rote
learning of nonsense syllables •. The
critical evaluation of EPAM must ultimately depend not upon the interest which
it may have as a learning machine, but
upon its ability to explain and predict
the phenomena of verbal learning.
I should like to preface my discussion of the simulation of verbal
learning with some brief remarks about
the class of information processing
models of which EPAM is a member.
a.

H. A. Simon has described some
current research in the simulation of
human higher mental processes and has
discussed some of the techniques and problems which have emerged from this
research. The purpose of this paper is
to place these general issues in the context of a particular problem by describing
in detail a simulation of elementary
human symbolic learning processes.

These are models of mental
processes, not brain hardware.
They are psychological models
of mental function. No physiological or neurological assumptions are made, nor is any
attempt made to explain information processes in terms of
more elementary neural processes.

b.

The information processing model
of mental functions employed is realized
by a computer program called Elementary
Perceiver and f'/lemorizer (EPAJ:v1). The
EPAM program is the precise statement of
an information processing theory of verbal
learning that provides an alternative
to other verbal learning theories which
have been proposed.** It is the result

These models conceive of the
brain as an information processor with sense organs as
input channels, effector org~ns
as output devices, and with
internal programs for testing,
comparing, analyzing, rearranging, and storing information.

c.

The central processing mechanism
is assumed to be serial; i.e.,
capable of dOing only one (or
a very few) things at a time.

d.

These models use as a basic
unit the information symbol;
i.e., a pattern of bits which is
assumed to be the brain'S
internal representation of
environmental data.

e.

These models are essentially
deterministic, not probabilistic.
Random variables play no fundamental role in them.

*1 am deeply indebted to Herbert A.
Simon for his past and present collaboration in this research. This research
has neen supported by the Computer Sciences Department, The RAND Corporation,
and the Ford Foundation. I wish to express appreciation for the help and
critical comments of Julian Feldman,
Allen Newell, J. C. Shaw and Fred Tonge.
**Examples of quantitative (or quasiquantitative) theories of verbal learnin$
are those of Hull, et.al. [1], Gibson [2J,
and Atkinson [3J. - -

122
3.2

THE BASIC EXPERIMENT
Early in the history of psychology,
the psychologist invented an experiment
to simplify the study of human verbal
learning. This "simple" experiment is
the rote memorization of nonsense
syllables in associate-pairs or serial
lists.
The items to be memorized are generally three-letter words having consonant letters on each end and a vowel
in the middle. Nonsense syllables are
chosen in such a way that the threeletter combinations have no ordinary
English meaning. For example, CAT is
not a nonsense syllable, but XUM is.*
In one basic variation, the rote
memory experiment is performed as follows:
a.

A set of nonsense syllables is
chosen and the syllables are
paired, making, let us say, 12
pairs.

b.

A subject
a viewing
syllables
pair at a

c.

First, the left-hand member of
the pair (stimulus item) is
shown. The subject tries to say
the second member of the pair
(response item).

d.

After a short interval, the
. response item is exposed so that
both stimulus and response items
are simultaneously in view.

e.

After a few seconds, the cycle
repeats itself with a new pair
of syllables. This continues
until all pairs have been
presented (a trial).

f.

is seated in front of
apparatus and the
are shown to him, one
time.

Trials ar~ repeated, usually until
the subject is able to give
the correct response to each
stimulus. There is a relatively
short time interval between
trials.

*People will defy an experimenter's
most rigorous attempt to keep the nonsense syllables association-free. Lists
of nonsense syllables have been prepared,
ordering syllables on the basis of their
so-called '-association value," in order
to perm~.t the experimenter to control
"meaningfulness .
P

g.

For successive trials the
syllables are reordered randomly.
This style of carryirig out the
experiment is called pairedassociates presentation.

The other basic variant of the
experiment is called serial-anticipation
presentation. The nonsense syllables
(say, 10 or 12 items) are arranged in
a serial list, the order of which is
not changed on successive trials. When
he is shown the nth syllable, the subject
is to respond with the (n+l)st syllable.
A few seconds later, the (n+l)st syllable
is shown and the subject is to respond
with the (n+2)nd syllable, and so on.
The experiment terminates when the subject
is able to correctly anticipate all of
the syllables.
Numerous variations on this experimental theme have been performed.*
The phenomena of rote learning are well
studied, stable, and reproducible.
For example, in the typical behavioral
output of a subject, one finds:
a.

Failures to respond to a stimulus
are more numerous than overt
errors.

b.

Overt errors are generally
attributable to confusion by
the subject between similar
stimuli or similar responses.

c.

Associations which are given
correctly over a number of
trials sometimes are then
forgotten, only to reappear
and later dissappear again.
This phenomenon has been called
oscillation.**

d.

If a list x of syllables or
syllable pairs is learned to
the criterion; then a list y
is similarly learned; and
finally retention of list x is
tested; the subject's ability
to give the correct x responses
is degraded by the interpolated
learning. The degradation is
called retroactive inhibition.
The overt errors made in the

*For an extended treatment of this
subject, see Hovland, C. I., llHuman
Learning and Retention. 11 [4J
**By Hull [5J. Actually he called it
"oscillation at the threshold of recall,"
reflecting his theoretical point of view.

123
3.2

retest trial are generally
intrusions from the list y. The
phenomenon disappears rapidly.
Usually after the first retest
trial, list x has been relearned
back to criterion.
e.

As one makes the stimulus
syllables more and more similar,
learning takes more trials.

The Information Processing Model
This section describes the processes
and structures of EPAM.
EPfu~ is not a model for a particular
subject. In this respect it is to be
contrasted with the binary choice models
of particular subjects which Mr. Feldman
is presenting in this session. The fact
is that individual differences play only
a small part in the results of the basic
experiment described above.

It is asserted that there are certain
elementary information processes which an
individual must perform if he is to
discriminate, memorize and associate
verbal items, and that these information processes participate in all the
cognitive activity of all individuals.*
It is clear that EPAM does not yet
embody a complete set of such processes.
It is equally clear that the processes
EPfu~ has now are essential and basic.
*Some information processing models
are conceived as models of the mental
function of particular subjects; e.g.,
Feldman's Binary Choice Model [6J. Others
treat the general subject as EPM~ does.
Still others are mixed in conception,
asserting that certain of the processes of
the model are common for all subjects while
other processes may vary from subject to
subject; e.g., the General Problem Solver
of Newell, Shaw and Simon [7J. Alternatively information processing models
may al~o be categorized.ac~ord~ng tOnhow
much of the processing lS harn core
(i.e., necessary and invariant) as opposed
to s trategic" (i.e, the result of
strategy choice by control processes). I
suggest the obvlous: that models ?f
strategies for information processlng
will tend to be mod31s of the general
subject. As exemplars, Lindsay's Reading
fvlachine [8J, a t1hard core" model, treats
the general subject; \I/ickelgren r s model
of the conservative Focusing strategJ"IT
in concept attainment (Hickelgren [9 ;
Br~~er, Goodnow, and Austin [10]), a pure
strategy model, can predict only the
behavior of particular subjects.
II

Overview: Performance and Learning
Conceptually, EPN~ can be broken
down into two subsystems, a performance
system and a learning system. In the
performa.nce mode, EPAM produces responses
to stimulus items. In the learning mode,
EPAM learns to discriminate and associate
items.
The performance system is the
simpler of the two. It is sketched in
Fig. 1. When a stimulus is noticed, a
perceptual process encodes it, producing
an internal representation (an input
code). A discriminator sorts the inpu~
code in a discrimination net (a tree 01
tests and branches) to find a stored
image of the stimulus. A response cue
associated with the image is found, and
fed to the discriminator. The discriminator sorts the cue in the net and finds
the response image, the stored form of
the response. The response image is then
decoded by a response generator le~~er by
letter in another discrimination net
into a form suitable for output. The
response is then produced as output.
The processes of the learning system
are more complex. The discrimination
learning process builds discriminations
by growing the net of tests and branches.
The association process builds associations between images by storing response
cues with stimulus images. These
processes will be described fully in due
course.
The succeeding sections on the
information processing model give a
detailed description of the processes and
structures of both systems.
Input to EPM~: Internal Representations
Of""External Data
The following are the assumptions
about the symbolic input process when a
nonsense syllable is presented to the
learner. A perceptual system receives
the raw external information and codes
it into internal symbols. These internal
symbols contain descriptive information
about features of external stimulus.
For unfamiliar 3-letter nonsense symbols,
it is assumed that the coding is done in
terms of the individual letters, for
these letters are familiar and are welllearned units for the adult subject.*
'\
*The basic perception mechanism I
have in mind is much the same as that of
Selfridge [llJ and Dinneen, whose computer
program scanned letters and perceived
simple topological features of these
letters.

124
3.2

The end result of the perception process
is an internal representation of the nonsense syllable--a list of internal symbols
(i.e., a list of lists of bits) containing descriptive information about
the letters of the nonsense syllable.
Using Minsky's terminology [12J, this is
the !!character" of the nonsense syllable.

called tests, which examine characteristics of an input cede and signal branchleft or branch-right. On each image
list will be found a list of symbols
called the image. An image is a partial
or total copy of an input code. I shall
use these names in the following
description of net processes.

I have not actually programmed this
perception process. For purposes of
this simulation, I have assigned coded
representations for the various letters
of the alphabet based on 15 different
geometrical features of letters. For
purposes of exploring and testing the
model, at present all that is really
needed of the input codes is:

Net Interpreter. The discrimination
net is examined and altered by a number
of processes, most important of which is
the net interpreter. The net interpreter
sorts an input code in the net and
produces the image list associated with
that input code. This retrieval process
is the essence of a purely associative
memory:-----tFle stimulus1nf'ormation i tse If
leads t6~--retrreva:l---or-the infC)rmation
associated with that stimulus~--T~l1e~
interpreter is a very simple process.
It finds the test in the topmost node
of the tree and executes this program.
The resulting signal tells it to branch
left or branch right to find the
succeeding test. It executes this,
tests its branches again, and repeats
the cycle until a terminal is found.
The name of the image list is produced,
and the process terminates. This is the
discriminator of the performance system
which sorts items in a static net.

a.

that the dimensions of a letter
code be related in some reasonable way to features of real
letters.

b.

that the letter codes be highly
redundant, that is, include
many more dimensions than is
necessary to discriminate the
letters of the alphabet.

To summarize, the internal representation of a nonsense syllable is a list
of lists of bits, each sublist of bits
being a highly redundant code for a
letter of the syllable.
Given a sequence of such inputs,
the essence of the learner's problem is
twofold: first, to discriminate each
code from the others already learned,
so that differential response can be
made; second, to assoclate information
about a "response!! syllable with the
information about a "stimulus ll syllable
so that the response can be retrieved
if the stimulus is presented.
Discriminating and Memorizing:
Trees of Images

Growing

I shall deal vtith structure first
and reserve my discussion of process
for a moment.
Discrimination net. The primary
information structure in EPAM is the
discrimination net. It embodies in its
structure at any moment all of the discrimination learning that has taken
place up to a given time. As an information struct').re it is no more than a
familiar friend: a sorting tree or
decoding network. Fig. 2 shows a small
net. At the terminals of the net are
lists called image lists, in which
symbolic information can be stored. At
the nodes of the net are stored programs,

Discrimination Learning. The discrimination learning process of the
learning system grows the net. Initial~
we give the learning system no discrimination net but only a set of simple
processes for growing nets and storing
new images at the terminals.
To understand how the discrimination
and memorization processes work, let us
examine in detail a concrete example
from the learning of nonsense syllables.
Suppose that the first stimulus-response
associate-pair on a list has been learned.
(Ignore for the moment the question of
how the association link is actually
formed.) Suppose that the first syllable
pair was DAX-JIR. The dis9rimination
net at this point has the simple twobranch structure shown in Fig. 3.
Because the syllables differ in their
first letter, Test 1 will probably be a
test of some characteristic on which the
letters D and J differ. No more tests
are necessary at this pOint.
Notice that the image of JIR which
is stored is a full image. Full response
images must be stored--to provide the
information for producing the response;
but only partial stimulus images need
be stored--to provide the information
for recognizing the stimulus. How much
stimulus image information is required

125
3.2

the learning system determines for itself
as it grows its discrimination net, and
makes errors vlhich it diagnoses as inadequate discrimination.

The net as it now stands is shown in
Fig. 4. Test 2 is seen to discriminate
on some difference between the letters
P and D.

To pursue our simple eyample, suppose
that the next syllable pair to be learned
is PIB-JUK. There are no storage
terminals in the net, as it stands, for
the two new it~ms. In other words, the
net does not have the discriminative
capability to contain more than two items.
The input code for PIB is sorted by the
net interpreter. Assume that Test 1
sorts it "doVln the plus branch of Fig. 3.
As there are differences between the
incumbent image (with first-letter D)
and the nevv code (with firs t -let ter p) an
attempt to store an image of PIB at this
terminal vJOuld destroy the information
previously stored there.

The input code for JUK is now sorted
by the net interpreter. Since Test 1
cannot detect the difference between the
input codes for JUK and JIR (under our
previous assumption), JUK is sorted to
th~ terminal containing the image of
JIR. The match for differences takes
place. Of course, there are no firstletter differences. But there are dif-·
ferences between the incumbent image and
the neyV code in the second and third
letters.

Clearly "tvhat is needed is the ability
to discriminate further. A match for
differences between the incumbent image
and the challenging code is performed.
i,'lhen a difference is found, a ne""J test is
created to discriminate upon this difference. The ne~'J test is placed in the net
at the point of failure to discriminate,
an image of the new item is created, and
both images--incUlnbent and new--arc
stored in terminals along their appropriate
branches of the new test, and the conflict
is resolved.*
*VIith the processes just described,
the discrimination net would be grown
each time a new item was to be added to
the memory. But from an information processing standpoint, the matching and netgrowins processes are the most timeconsuming in the system. In general, with
little additional effort, more than one
difference can be detected, and more than
one discriminating test can be added to
the net. Each redundant test placed in
the net gives one emptyli image list.
At some fut'cl.re time, if an item is sorted
to this empty image list, an image can be
stored "\Tithout grouing the net. There is
a happy mediurn bet\'leen small nets vJhich
must be grown all the time and large nets
replete with redundant tests and a wasteful surplus of empty image lists. Experimentation vii th this II structural parameter!!
has been done and it has been found
that for this study one or two redundant
tests per growth represents the happy
medium. Hovlever, I wov.ld not care to
speak of the generality of this particular result.
l'

Noticing Order. In which letter
should the matching process next scan
for differences? In a serial machine
like EPAN, this scanning must take place
in some order. This order need not be
arbitrarily determined and fixed. It
can be made variable and adaptive. To
this pnd EPMl has a noticing order for
letters of syllables, vlhich prescribes
at -- any--ii1ornent a letter-scanning sequence
for the matching process. Because it is
observed that subjects generally consider
end-letters before middle-letters, the
notiCing order is initialized as follows:
first-letter, third-letter, secondletter. lifhen a particular letter being
scanned yields a difference, this letter
is promoted up one position on the
noticing order. Hence, letter positions
relatively rich in differences quickly
get priority in the scanning. In our
example, because no first-letter differences '11ere found between the image, of
JIR and code for JUK, the third letters
are scanned and a difference is found
(between Rand K). A test is created
to capitalize on this third-letter
difference and the net is grown as
before. The result is shown in Fig. 5.
The noticing order is updated; thirdletter, promoted up one, is at the head.
LearninG of subsequent items
proceeds in the same way, and "de shall
not pursue the example further~

Asso~J.5t~n~~~ITl.~~~s:

Ret2'i~y..?-_~{_~ing; Cues

The discrimination net and its
interpreter associa~e codes of external
objects with internal image lists and
images. But the basic rote learning
experiment requires that stimulus
information someho~'J lead ~o responi3e

126
3.2

information and a response. The discrimination net concept can be used for the
association of internal images with each
other (i.e., response with stimulus) with
very little addition to the basic
mechanism.
An association between a stimulus
image and a response image is accomplished
by storing with the stimulus image some of
the coded information about the response.
This information is called the cue. A cue
is of the same form as an input code, but
generally contains far less information
than an input code. A cue to an associated image can be stored in the discrimination net by the net interpreter to retrie'le the associated image. If, for example, in the net of Fig. 3 we had stored
with the stimulus image the letter J as
a cue to the response JIR, then sorting
this cue would have correctly retrieved
the response image. An EPAM internal
association is built oy storing with the
stimulus image information sufficient to
retrieve the response image from the net
at the moment of association.
The association process determines
how much information is sufficient by
trial and error. The noticing order for
letters is consulted, and the firstpriority letter is added to the cue. The
cue is then sorted by the net interpreter
and a response image is produced. It
might be the wrong response image; for if
a test seeks information which the cue does
not contain, the interpreter branches left
or right randomly (with equal probabilities) at this test.* During association,
the selection of the wrong response is
immediately detectable (by a matching
process) because the response input code
is available. The next-priority letter is
added to the cue and the process repeats
until the correct response image is retrieved. The association is then considered complete.
Note two important possibilities.
First, by the process just described, a
cue which is really not adequate to
guarantee retrieval of the response image
may by happenstance give the correct
response image selection during association. This "luck" usually gives rise to
response errors at a later time.
*This is the only use of a random
variable in EPAM. We do not like it. We
use it only because we have not yet discovered a plausible and satisfying adaptive mechanism for making the decision.
The random mechanism does, however, give
better results than the go-one-way-allthe-time mechanism which has also been
used.

Second, suppose that the association
building process does its job thoroughly.
The cue which it builds is sufficient to
retrieve the response image at one particular time, the time at which the two ite
items were associated.' If, at some
future time, the net is grown to encompass
new images being added to the memory,
then a cue which previously was sufficient
to correctly retrieve a response image
may no longer be sufficient to retrieve
that response image. In EPAM, association
links are 11 dated, II and ever vulnerable
to interruption by further learning.
Responses may be lIunlearned" or "forgotten "
temporarily, not because the response information ha& been destroyed in the memory,
but because the information has been temporarily lost in a growing network. If an
'association failure of this type can be
detected through feedback from the
environmental or experimental situation,
then the trouble is easily remedied by
adding additional response information
to the cue. If not, then the response
may be more or less permanently lost in
the net. The significance of this
phenomenon will perhaps be more easily
appreCiated in the discussion of results
of the EPAM simulation.
Responding:

Internal and External

A conceptual distinction is made
between the process by which EPAIvl selects
an internal response image and the process by which it converts this image
into an output to the environment.
Response retrieval. A stimulus item
is presented. This stimulus input code
is sorted in the discrimination net to
retrieve the image list, in which the cue
is found. The cue is sorted in the net
to retrieve another image list containing
the proposed response image. If there is
no cue, or if on either sorting pass an
empty image list is selected, no response
is made.
Response generation. For purposes
of response generation, there is a fixed
discrimination net (decoding net),
assumed already learned, which, transforms
letter codes of internal images into
output form. The response image is decoded letter by letter by the net
interpreter in the decoding net for
letters.
The Organization of the Learning Task
The learning of nonsense symbols by
the processes heretofore described takes
time. EPAM is a serial machine. There-

127
3.2

fore, the individual items must be dealt
with in some sequence. This sequence
is not arbitrarily prescribed. It is the
result of higher order executive processes whose function is to control EPAM's
focus of attention. These macroprocesses,
as they are called, will not be described
or discussed here. A full exposition of
them is available in a paper by
Feigenbaum and Simon. [13]
Stating the f-1odel Precisely:
Computer Program for EPAM
The EPAM model has been realized
as a program in Information Processing
Language V [14] and is currently being
run both on the Berkeley 704 and the RAND
7090. Descriptive information on the
computer realization, and also the
complete IPL-V program and data structures
for EPAM (as it stood in October, 1959)
are given in an earlier work by the
author [15].

can also be printed out.
A number of simulations of the
basic paired-associate and serialanticipation experiments have been run.
Simulations of other classical experiments in the rote learning of nonsense
syllables have also been run. The
complete results of these simulation
experiments and a comparison between
EPAM's behavior and the reported behavior of human subjects will be the subject of a later report. However, some
brief examples here will give an indication of results expected and met.
a.

Stimulus and response generalization. These are psychological terms used to describe
the following phenomenon. If
X and X' are similar stimuli,
and Y is the correct response
to the presentation of X; then
if Y is given in response to
the presentation of X', this is
called stimulus generalization.
Likewise, if Y and Y' are similar responses, and Y' is given
in response to the presentation
of X, this is called response
generalization. Generalization
is common to the behavior of
all subjects, and is found in
the behavior of EPAM. It is
a consequence of the responding
process and the structure of the
discrimination net. For those
lIstimulil! are similar in the
EPAM memory whose input codes
are sorted to the same terminal;
and one 11responseft is f?imilar
to another if the one is stored
in the same local area of the
net as the other (and hence
response error may occur when
response cue information is
insufficient).

b.

Oscillation and Retroactive
Inhibition. We have described
these phenomena in an earlier
section.

IPL-V, a list processing language,
was well suited as a language for the
EPAM model for these key reasons:
a. The IPL-V basic processes deal
explicitly and directly with list structures. The various information structures
in EPAM (e.g., discrimination net, image
list) are handled most easily as list
structures. Indeed, the discrimination
is, virtually by definition, a list
structure of a simple type.
b. It is useful in some places, and
necessary in others, to store with some
symbols information descriptive of these
symbols. IPL-V's description list and
description list processes are a good
answer to this need.
c. The facility with-which hierarchies of subroutine control can be
written in IPL-V makes easy and uncomplicated the programming of the kind
of complex control sequence which EPAM
uses.
Empirical Explorations with EPAM
The procedure for exploring the
behavior of EP~l is straightforward. We
have written an I!Experimenter program
and we give to this program the particular conditions of that experiment as
input at the beginning of an experiment.
The Experimenter routine then puts EPAM
qua subject through its paces in that
particular experiment. The complete
record of stimuli presented and responses
made is printed out, as in the final net.
Any other information about the processing or the state of the EPAM memory
l1

Oscillation and retroactive
inhibition appear in EPAM's
behavior as consequences of
simple mechanisms for discrimination, discrimination learning,
and association. They were in
no sense "designed into" the
behavior. The appearance of
rather complex phenomena such
as these gives one a little
more confidence in the credibility of the basic assumptions
of the model.

128
3.2

These two phenomena are discussed
together here because in EPAM
they have the same origin. As
items are learned over time,
the discrimination net grows to
encompass the new alternatives.
Growing the net means adding
new tests, which in turn means
that more information will be
examined in all objects being
sorted. An important class of
sorted objects is the set of
cues. Cue information sufficient
at one moment for a firm association may be insufficient at a
later moment. As described
above, this may lead to response
failure. The failure is caused
entirely by the ordinary process
of learning new items. In the
case of oscillation, the new
items are items within a single
list being learned. In the case
of retroactive inhibition, the
new items are items of the second
list being learned in the same
discrimination net. In both
cases the reason for the response
failure is the same. According
to this explanation, the phenomena are first cousins (an hypothesis which has not been widely
considered by psychologists).
In the EPAM model, the term
interference is no longer merely
descriptive--it has a precise
and operational meaning. The
process by which later learning
interferes with earlier learning
is completely specified.
c.

Forgetting. The usual explanations of forgetting use in one
way or another the simple and
appealing idea that stored information is physically destroyed
in the brain over time (e.g., the
decay of a Itmemory trace," or
the over~riting of old information by new information, as
in a computer memory). Such
explanations have never dealt
adequately with the commonplace
observation that all of us can
remember, under certain conditions, detailed and seemingly
unimportant information after
very long time periods have
elapsed. An alternative explanation, not so easily visualized,
is that forgetting occurs not
because of information destruction but because learned
material gets lost and inaccessible in a large and growing
association network.

EPAM forgets seemingly welllearned responses. This forgetting occurs as a direct
consequence of later learning
by the learning processes.
Furthermore, forgetting is only
temporary: lost associations
can be reconstructed by storing
more cue information. EPAM
provides a mechanism for explaining the forgetting phenomenon
in the absence of any information loss. As far as we
know, it is the first concrete
demonstration of this type of
forgetting in a learning machine.
Conclusion:

A Look Ahead

Verification of an information
proceSSing theory is obtained by simulating many different experiments and
by comparing in detail specific
qualitative and quantitative features of
real behavior with the behavior of the
simulation. To date, Mr. Simon and I
have run a number of simulated experiments.
As we explore verbal learning further,
more of these will be necessary.
We have been experimenting with a
variety of "sense modes" for EPAM,
corresponding to "visual input and
"written fl output, ttauditorylt input and
Ilorall! output, Itmuscular" 1nputs and
outputs. To each mode corresponds a
perceptual input coding scheme, and a
discrimination net. Associations-acrossnets, as well as the familiar associat10nsWithin-nets, are now possible. Internal
transformations between representations
in different modes are possible. Thus,
EPAM can II sound" in the flmind 1 sear"
what it lIsees tt in the "mind's eye," just
as all of us do so easily. We have been
teaching EPAM to read-by-association,
much as one teaches a small child beginning
reading. We have only begun to explore
this new addition.
lt

The EPAM model has pointed up a
failure shared by all existing theories
of rote learning (including the present
EPAM). It is the problem of whether
association takes place between symbols
or between tokens of these symbols.
For example, EP M'l cannot learn a serial
list in which the same item occurs
twice. It cannot distinguish between
the first and second occurrence of the
the item. To resolve the problem we
have formulated (and are testing)
processes for building, storing, and
responding from chains of token associations.

129
3.2

References
1.

2.

Hull, C. L., C. I. Hovland, R. T.
Ross, M. Hall, D. T. Perkins
and F. B. Fitch, Mathematicodeductive Theory of Rote Learning,
New Haven, Connecticut, Yale
University Press, 1940.
Gibson, E. J., "A Systematic Application of the Concepts of
Generalization and Differentiation
to Verbal Learning," Psychol. Rev.,
Vol. 47, 1940, pp. 196-229.

3.

Atkinson, R. C., "An Analysis of
Rote Serial Learning in Terms
of a Statistical Model,l1 Indiana
University, Doctoral Dissertation,
1954.

4.

Stevens, S. S., ed., Handbook of
Experimental Psychology, New York,
Wiley, 1951.

5.

Hull, C. L., llThe Influence of
Caffeine and Other Factors on
Certain Phenomena of Rote Learning,
J. Gen. Psychol., Vol. 13, 1935,
pp. 249-273.

6.

Feldman, J., ,1 An Analysis of Predictive Behavior in a Two Choice
Si tuation, Carnegie Institute of
Technology, Doctoral Dissertation,
1959.
II

7.

Newell, A., J. C. Shaw, and H. A.
Simon, IIReport on a General Problem
Solving Program," Information
Processing: ProceedIng:S of the
International Conference on
Information Processing~SCO,
Paris, June, 1959, pp. 256-264.

8.

Lindsay, R., "The Reading f.1achine
Problem,1I CIP Working Paper #33,
Carnegie Institute of Technology
(Graduate School of Industrial
Administration), Pittsburgh, 1960.

9.

Wickelgren, W., "A Simulation
Program for Conservative Focusing,1I
unpublished manuscript, University
of California, Berkeley, January,
1961.

10.

Bruner, J. S., J. J. Goodnow, and
G. A. Austin, A Study of Thinking,
New York, Wiley, 1956.

11.

Selfridge, O. G., IIPattern Recognition and Modern Computers,"
Proceedings of the 1955 Western
Joint Computer Conference, IRE,
March, 1955.

II

12.

lVl~nsky,

13.

Feigenbaum, E. and H. A. Simon,
"A Theory of the Serial Position
Effect,1\ CIP Working Paper 1114,
Carnegie Institute of Technology
(Graduate School of Industrial
Administration), Pittsburgh, 1959.

14.

Newell, A., F. M. Tonge, E. A.
Feig;enbaum" G. H. Mealy, N. Saber,
B. F. Green, and A. K. Wolf,
Information Processing Language V
ManUaTlSectionSIanall) ,
Santa Monica, California, The RAND
Corporation, p-1897 and P-19l8.

15.

Feigenbaum, E., An Information
Processing Theory or-VeroarLearning,--Santa l\lonica, California,
The RAND Corporation, p-1817,
October, 1959.

M., "Steps Toward Artificial
Intelligence,1I Proceedings of
the IRE, Vol. 49, No.1, January,
1961, pp. 8-30.

130
3.2

RAW STIMULUS

PERCEIVE FEATURES
OF STIMULUS
EPAM STIMULUS
INPUT CODE

DISCRIMINATE
STIMULUS TO
FIND STIMULUS IMAGE
IMAGE
FIND ASSOCIATED CUE
CUE

DISCRIMINATE
CUE TO FIND
RESPONSE I MAGE
RESPONSE IMAGE
GENERATE RESPONSE
TO ENVIRONMENT
USING DECODING NET

RESPONSE

OUTPUT

Fig. 1- EPA M performance process
for producing the response
associated with a stimulus

131

3.2

@
ITJ
Irtci

= Image and cue at a terminal

0

= Empty terminal

= Discriminating test at a node
= Image at a terminal

Fig. 2 - A typical EPA M discrimination net

STIMULUS

RESPONSE

DAX

JIR

Fig. 3 - Discrimination net after the learning
of the first two items. Cues are not shown.
Condition: no redundant tests added.
Test 1 is a first-letter test.

132
3.2

STIMULUS

RESPONSE

PIB

JUK

Fig. 4 - Discrimination net of Fig.3 after the
learning of stimulus item,PIB.
Tes t 2 is a first letter test

STIMULUS

RESPONSE

PIB

JUK

Fig. 5-Discrimination net of Fig.4 after
the learning of the response item, J UK.
Test 3 is a third-letter test

133
3.3
SIMULATION OF BEHAVIOR
IN nlE BINARY CHOICE EXPERIMENT
Julian Feldman
University of California, Berkeley
Summary
A modern, high-speed digital computer has
been used to simulate the behavior of individual
human subjects in a classical psychological experiment where the subject is asked to predict a
series of binary events. The representation of
models of human behavior in the form of computer
programs has permitted the construction and study
of more realistic hypothesis-testing models of
behavior in this experiment rather than the oversimplified conditioning models previously proposed.
~ model for one subject is described in detail,
and the problem of comparing the behavior of the
model to the behavior of the subject is also discussed.
Introduction
Modern, high-speed digital computers have
been used to simulate large, complex systems in
order to facilitate the study of these systems.
One of these systems that has been studied with
the aid of computer simUlation is man. The present paper describes another addition to the growing list of efforts to study human thinking processes by simulating these processes on a computer. The research summarized here has been
concerned with simulating the behavior of individual subjects in the binary choice experiment. 3
The first section of this paper contains a description of the experiment. An overview of the
model is given in the second section. The model
for a particular subject is described in SOme
detail in the third section.
The Binary Choice Experiment
In the binary choice experiment, the subject
is asked to predict which of two events, El or E2,
will occur on each of a series of trials. After
the subject makes a prediction, he is told ~ich
event actually occurred. The sequence of events
is usually determined by Some random mechanism,
e.g., a table of random numbers. One and only one
event occurs on each trial. The events may be
flashes of light or symbols on a deck of cards.
The subject is usually asked to make as many
correct predictions as he can.
In the research reported here, the experiment
described in the preceding paragraph was modified
by asking the subject to "think a10ud"--to give
his reasons for making a prediction as well as the
prediction itself. The subject's remarks were
recorded. The subject was instructed to "think
aloud" in order to obtain more in forma tion on the
processing that the subject was doing. This technique has been used in some of the classical investigations of problem-solving behavior 2 ,S and

in other computer simulation studies of
thinking. l ,7 A comparison of the behavior of subjects in the binary choice experiment who did
"think aloud" with the behavior of subjects who
did not "think aloud" did not reveal any major
differences. 3 The events in the present experiment were the symbols "plus" and "check." "Check"
occurred on 142 of 200 trials and "plus" on the
remaining S8 trials. The symbols were recorded on
a numbered deck of 3 inch x S inch cards. After
the subject made his prediction for trial t, he
was shown card t which contained a "plus" or
"cheCk." While the subject was predicting the
event of trial t, he could only see the event of
trial t-l. A transcription of the tape recording
of the remarks of subject DH and the experimenter,
the author, in an hour-long binary choice experiment is pre~ented in the Appendix. In the Appendix and the rest of this paper, the symbols "plus"
and "check" are represented by uP" and "C" respectively. The transcription will be referred to as
a protocol.
The Basic Model
To simulate the behavior of an individual
subject in the binary choice experiment, a model
of the subject's behavior must be formulated as a
computer program. If the program is then allowed
to predict the same event series as the subject
has predicted, the behavior of the program--the
predictions and the reasons--can be compared to
the behavior of the subject. If the program's
behavior is a reasonable facsimile of the subject's behavior, the program is at least a sufficient explanation of the subject's behavior. The
level of explanation is really determined by the
subject's statements. No attempt is made to go
beyond theae to more basic processes, e.g., neurological or chemical, of human behavior. Thus, the
model is an attempt to specify the relationship
between the reasons or hypotheses that the subject
offers for his predictions and the preceding hypotheses, predictions, and events. The subject is
depicted as actively proposing hypotheses about
the structure of the event series. These hypotheses are tested by using them to predict events.
If the prediction of the event is correct, the
hypothesis is usually retained. If the prediction
of the event is wrong, a new hypothesis is generally proposed.
The Model for 00
The model for each subject is based on a
detailed examination of the protocol and some conjectures about hUman behavior. Perhaps the best
thing to do at this point is to describe in Some
detail a model for the subject. DH, whose protocol
appears in the Appendix.

134
3.3
The Hypotheses

The Basic Cycle

This model proposes two types of hypotheses
about the event series. The first type of hypothesis is a pattern of events. The model has a
repertoire of nine patterns:

The basic cycle of the model is as follows:
The model uses an hypothesis to predict the event
of trial t. The event is then presented. The
model in Phase One "explains" the event of trial t
with an explanation-hypothesis. In Phase Two a
prediction-hypothesis for trial t+l is formed.
The model uses this prediction-hypothesis to predict trial t+l. The event of trial t+l is presented, and the cycle continues.

progression of C's
progression of P's
single alternation
2 C's and I P
1 C and 2 P's
2 P's and 2 C's
3 P's and 3 C's
4 p's and 4 C's
4 P's and 3 C's
The model can propose that the event series
is behaving according to one of these patterns and
use the pattern-hypothesis to predict the event of
a given trial, t. The predictions of the first two
patterns--progression of C's and progression of
P's--for trial t are independent of the events
preceding trial t. The predictions of the other
patterns (the alternation patterns) are dependent
on these preceding events. Thus, if the subject
proposes the pattern "single alternation" for
trial t and the event of trial t-l was a C, the
prediction for trial t is a P. In order to facilitate the determination of the prediction of an
alternation pattern for trial t, the patterns are
coded as sorting nets. Por example, the pattern
"2 C's and 1 P" is represented in the following
fashion:
Is event t-l a C?
No--Predict C for trial t.
Yes-Is event t-2 a C?
No--Predict C for trial t.
Yes-Predict P for trial t.
The second type of hypothesis that the model
can propose is an anti-pattern or guess-opposite
hypothesis. Por example, the model can propose
that the event of trial t will be the opposite of
that predicted by a given pattern. This type of
hypothesis is the model's representation of the
notion of "gambler's fallacy"--the reason people
predict "tails" after a coin falls "heads" seven
times in a row.
The most general form of hypothesis has two
components: a pattern component and a guess-opposite component. The prediction of the hypothesis is obtained by finding the prediction of
the pattern component. If the hypothesis has a
guess-opposite component, then the prediction of
the hypothesis is the opposite of the pattern prediction. If the hypothesis does not have a guessopposite component, then the prediction of the
hypothesis is the prediction of the pattern component. Thus, while the prediction of the
pattern-hypothesis "progression of C's" is always
a C, the prediction of the hypothesis "guess-opposite-progression-of-C's" is always a P.

Phase One
The basic motivation for this phase of the
model is that the model must "explain" each event.
An acceptable explanation is an hypothesis that
could have predicted the event. The processing of
Phase One is represented in the flow chart of
Pig. 1.
The processing to determine the explanationhypothesis for trial t begins by testing whether
the pattern component of the prediction-hypothesis
for trial t could have predicted the event of
trial t correctly. If the pattern component could
have predicted correctly, the pattern component is
the explanation-hypothesis. If the pattern component could not have predicted correctly, the
pattern-change mechanism is evoked. Thus if the
prediction-hypothesis for trial t contained only a
pattern component and the hypothesis predicted
correctly, the explanation-hypothesis for trial t
is the prediction-hypothesis for trial t. If the
prediction-hypothesis for trial t contained a
guess-opposite component and the hypothesis predicted correctly, the pattern-change mechanism is
evoked because the pattern component could not
have predicted the event correctly by itself. If
the prediction-hypothesis for trial t was a guessopposite-hypothesis and it predicted incorrectly,
the pattern component of the prediction-hypothesis
becomes the explanation-hypothesis for trial. t.
The motivation here is really quite simple
although the explanation may sound involved.
If, in this binary situation, the hypothesis that
a pattern will change leads to an incorrect prediction, the pattern must have persisted; and the
pattern is an acceptable explanation of the event.
If the hypothesis that a pattern will change leads
to a correct prediction, the pattern obviously did
not persist; and the possibility of a new pattern
is considered.
The pattern-change mechanism is evoked on
trial t if the pattern component of the prediction-hypothesis for trial t is unable to predict
the event of trial t. The pattern-change mechanism consists of two parts. The first part evokes
a subset of the nine patterns listed above. The
second part of the pattern-change mechanism
selects a single pattern out of the evoked set.
A pattern is evoked, i.e., considered as a possible explanation of the event of trial t, if the
pattern can predict the events of trials t and
t-l. The pattern of the prediction-hypothesis for
trial t, i.e., the pattern that cannot predict
event t, is included in the evoked set if it can

135
3.3
predict events t-l, t-2, and t-3. Of the patterns
that are evoked, the pattern that has been
selected most often on prior trials is selected as
the pattern component of the explanation-hypothesis. If the pattern component of the prediction-hypothesis for trial t is selected, then the
explanation-hypothesis is an anti-pattern hypothesis which is the model's interpretation of the
subject's hypothesis "you have thrown me off the
pattern" (cf. trial 9 of the protocol in the
Appendix). The model interprets event t as an
attempt to "throw me off" when the following three
conditions are met: (1) the pattern is unable to
predict the event of trial t; (2) the pattern is
able to predict at least the three consecutive
events of trials t-l, t-2, and t-3; and (3) the
pattern is also the most frequently selected of
those patterns that are evoked.
Phase Two
While Phase One is concerned mainly with the
processing of the pattern-component of the hypothesis, Phase Two is concerned with the processing
of the guess-opposite component. Phase Two is
represented in the flow chart of Fig. 2.
If the prediction-hypothesis for trial t contained a guess-opposite component, the guessopposite component is processed in a fashion quite
analogous to the processing of the pattern component in Phase One. If the anti-pattern prediction-hypothesis for trial t predicted the event of
trial t correctly, the guess-opposite component is
retained, and the prediction-hypothesis for trial
t+l is guess-opposite-the-pattern-of-the-explanation-hypothesis. If the anti-pattern predictionhypothesis for trial t predicted the event of
trial t incorrectly, the guess-opposite component
is considered for retention in a fashion analogous
to the ttthrow-me-off" consideration for patterns.
If the prediction-hypotheses for trials t-l and
t-2 had guess-opposite components and these hypotheses predicted correctly, then the guess-opposite component is retained for the prediction-hypothesis of trial t+l. If these conditions are
not fulfilled, the guess-opposite component is
dropped; and the prediction-hypothesis for trial
t+l is the explanation-hypothesis for trial t.
If the prediction-hypothesis for trial t did
not contain a guess-opposite component, the model
considers whether or not the guess-opposite component should be introduced on trial t+1. The
model makes this decision on the basis of its
past experience. It determines the.number of consecutive events including and preceding the event
of trial t that can be predicted by the explanation-hypothesis for trial t. This number will be
called Nl. Then the model searches its memory
backwards from the last trial included in Nl to
find a trial for which the explanation-hypothesis
was the same as the explanation-hypothesis for
trial t. Then the model determines the number of
contiguous events including, preceding, and following this prior occurrence of the explanationhypothesis of trial t that can be predicted by
this hypothesis. This number will be called N2.

If N2=Nl, the model decides that the explanationhypothesis for trial t will not be the predictionhypothesis for trial t+l. The prediction-hypothesis for trial t+l becomes guess-opposite-theexplanation-hypothesis for trial t. If N2::> Nl,
the model decides that the explanation-hypothesis
for trial t will be the prediction-hypothesis for
trial t+l. If NI> N2, the model decides that this
prior occurrence of the explanation-hypothesis for
trial t is really not pertinent and continues to
search its memory for an occurrence of the explanation-hypothesis where N2~Nl. If no such
occurrence can be found, the prediction-hypothesis
for trial t+l is the explanation-hypothesis for •
trial t.
Predicting with the Models
Models of individual behavior like the one
described for DH can be used to predict the same
series of binary events that the subject was
asked to predict. The predictions and hypotheses
of the model--the model's protocol--can then be
compared to the subject's protocol. The model
does not speak idiomatic English, and so the comparison is made between the machine's protocol
and a suitably coded version of the subject's
protocol.
The model's protocol can be generated by
presenting the model with the events in the same
way the subject was presented with the events in
the binary choice experiment; or the computer can
take the experimenter's role, too, if suitable
precautions are taken to prevent the model from
peeking. However, this straightforward method of
simulating the subject's behavior raises difficulties. These difficulties are identical to those
of getting a chess or checker program to play a
book game. 6 ,8 Because the decision of the chess
or checker program at move m depends on its decisions at the preceding moves, m-l, m-2, ••• , such
a program, when it is playing a book game, must be
"set back on the track" if its move deviates from
the book move. Th~ program and the book must have
the same history if the program is to have a fair
chance to make the same decision as that made in
the book game. This ttsetting-back-on-the-track"
may involve resetting a large number of parameters
as well as changing the move itself. Elsewhere, I
have called this "setting-back-on-the-track" technique conditional prediction. 4 The prediction of
the model is conditional on the preceding decisions of the model being the same as those of the
subject it is trying to predict.
The application of the conditional prediction
technique to binary choice models such as the one
described above for subject DH involves (1) COmparing the program's behavior and the subject's
behavior at every possible point, (2) recording
the differences between the behaviors, and (3)
imposing the subject's decision on the model
where necessary. A type of monitor system is
imposed on the program to perform these functions.
The model for DH with the conditional prediction
system controls is represented in Figs. 3 and 4.
An example will help clarify these figures. In

136
3.3
Fig. 3, after each decision by the model to keep
the pattern of the prediction-hypothesis for
trial t for the explanation-hypothesis for trial
t (8), this decision is compared to the subject's
decision (1). If the model's decision was different from that of the subject, control is transferred to the pattern-change mechanism (3 trials).
If the model's decision was the same as that of
the subject, control is transferred to another
part of the monitor (117 trials). Figs. 3 and 4
only contain the results for 195 trials because
the model began at trial 6.

Some evidence also exists that when suitably
motivated by money, Some subjects in a binary
choice experiment will predict the most frequent
event on each trial. Models for these subjects
require statements of the conditions under which
subjects abandon testing other hypotheses or at
least abandon testing hypotheses by using them to
predict events. Hypotheses could still be considered and tested without using them to predict
events.

Conclusions

The consequences of computer simulation for
the study of hUman behavior have been discussed at
some length in several places, and I have made a
limited statement of my views on this matter in
~nother place. 4 It will suffice then to discuss
some of the implications of the work reported here
for our understanding of behavior. The computer
models of binary choice behavior are relatively
simple computer programs; however, they are relatively complex psychological models. A widely
accepted view of binary ch~ice behavior has been
the idea of verbal conditioning embodied in the
stochastic learning model. In its simplest form,
this model says the subject's probability of
predicting El or E2 in the binary choice experiment is an exponentially-weighted moving average
over preceding events. The verbal conditioning
model is hardly consistent with the hypothesis~
testing behavior exhibited by DH and a dozen other
subjects for whom I have protocols. Protocols of
group behavior in the binary choice experiment
made available to me by David G. Hays are also
consistent with the general idea of hypothesistesting. Other inadequacies of the verbal conditioning model and evidence for hypothesis-testing
models have been discussed elsewhere. 3

Deficiencies of the Models
The model for DH and the similar models that
have been constructed to simulate the behavior of
two other subjects in the binary choice experiment 3 are deficient in several respects. First of
all, the comparison of the behavior of the model
to that behavior of the subject from which the
model was developed is, of course, not a very good
test of the model. This type of comparison only
yields some indication of the adequacy of the
model and its components. Comparison of the behavior of the model to sequences of behavior of the
subject not used in constructing the model awaits
correction of some of the deficiencies mentioned
below.
The segment of the model which has the
highest number of errors relative to the number of
times it is used is the guess-opposite segment
(see Fig. 4). The subject certainly exhibits this
type of behavior, but the model does not very
often predict "guess opposite" when the subject
does.

Contributions of the Models

The pattern-change segment has a better error
record, but it raises another issue. This segment
is actually a selection device. A pattern is
selected from the list of patterns that the subject uses. A more elegant pattern-change meChanism would generate a pattern out of the preceding
sequence of events and some basic concepts. One
of these concepts might be that patterns with
equal numbers of pes and C's are preferred to
alternation patterns with unequal numbers of
pIS and C's, all other things being equal.

The computer has provided the exponents of
hypothesis-testing models of behavior with the
means for studying and testing these complex
models. Oversimplified explanations of human
behavior can no longer be justified on the grounds
that the means for studying complex models do not
exist. Hopefully, the use of computers to simulate human behavior can extend man's intellect by
helping him study his own behavior.

The models have no mechanisms for making
perceptual errors--"seeing" one symbol when
another has occurred. Examination of the protocol
of DH (Appendix) indicates that he does sometimes
think that a C is a P (e.g., trial 196).

The author is indebted to Allen Newell for
advice and suggestions made during the course of
the research summarized in this paper.

The models do not have a sufficiently rich
repertoire of hypotheses. Subjects entertain
more types of hypotheses about the event series
than the two types, pattern and anti-pattern used
in the model for DH. Some subjects entertain more
sophisticated hypotheses. For example, one subject was able to detect the fact that a series of
events was randomized in blocks of ten trials,
i.e., the series had 7 P's and 3 C's in each block
of ten trials.

1.

Clarkson, G.P.E., and Simon, H.A. Simulation
of individual and group behavior. American
Economic Review, 1960, 50, 920-932.--------

2.

Duncker, K. On problem solving. Psychological Monograph!_ 1945, 58, No. 270.

3.

Feldman, J. An analysis of predictive beh!vior in a two-choice situation. Unpublished
doctoral dissertation, Carnegie Institute of

Acknowledgment

References

~

137
3.3
Technology, 1959.
4.

Computer simulation of cognitive
processes. in H. Borko (ed.), Computer
applications in the behavioral sciences,
Englewood Cliffs, Prentice-Hall, forthcoming.

(What do you say for the 4th symbol?) I'll say C
again. (Why?) This time I feel it'll be a C.
(The 4th symbol is a C. When you give your
answer, if you say, "1 think the 5th one will be
something," it'll be easier to check the tape
against the answer sheet.)

5.

Heidbreder, E. An experimental study of
thinking. Archives of Psychology, 1924, 11,
No. 73.

(What do you think the 5th one will be?) The 5th
one will be a P. (Why is that?) I feel it'll be
a P, that's all. (The 5th one is a C.)

6.

Newell, A., Shaw, J.C., and Simon, H.A.
Report on the play of Chess Player 1-5 of a
book game of Morphy vs. Duke Karl of
Brunswick and Count Isouard. CIP Working
Paper No. 21, Graduate School of Industrial
Administration, Carnegie Institute of Technology, 1959.

(What do you think the 6th one will be?) The 6th
one will be a C because you've been giving me C's
all along, and I don't think this progression will
end. (The 6th one is a C.)
,

Newell, A., and Simon, H.A. The simulation of
hUman thought. RAND Corporation Paper
P-1734, 1959.

C.)

Samuel, A.L. Some studies in machine learning, using the game of checkers. IBM
Journal of Research and Developmen~1959,
l, 210-230.

a P.)

7.

8.

Appendix:

Protocol of Subject DH*

(All right, now I'll read the instructions to you.
I'm going to show you a series of symbols. They
will either be a P symbol or a C symbol. Before
each word I'll give the signal N~~. When you hear
the signal NOW, tell me what symbol you expect
will occur on the next trial and why you selected
that symbol. That's the purpose of the tape
recorder. Take your time. After you have given
me your guess, I will show you the correct symbol.
Your goal is to anticipate each word as accurately
as you can. Please ••• Well, do you have any
questions?) Primarily, I just guess whether it'll
be a P or a C. (That's it.) But this explaining
why I think so. It can be little more than--I
think it'll be this, I guess, I have a feeling.
How more involved can it be than that? (Well,
whatever reasons you have. If those are the only
reasons that occur to you as you go thru this
those will be the only reasons. Maybe they won't.
OK, we'll try a few and then if you have any
questions ••• )

(What do you think the 7th one will be?) The 7th
one will be a C because I don't think the progression will be broken. (OK, the 7th one was a
The 8th one will be a C for the same reason. You
won't break the progression. (OK, the 8th one is
(What do you think the 9th one will be?) The 9th
one will be a C. (Why is that?) I think that you
just gave me the P to throw me off and you'll continue the progression. (The 9th one is a C. Oh,
one thing, can you see these cards?) Yes. (Can
you see me writing?) No, I can't. (OK.) I'm not
looking. (Well, you can look at these cards. I
want you to see I'1Ti not picking these out of my
head. This set has been predetermined.)
All right. This one will be a P. The 10th one
will be a P. (Why is that?) I feel that the progression will start to mix up now. (The 10th one
is a C.)
(What do you think the 11th one will be?) The
11th one will be a C. You're continuing the progression. (The 11th one is a C.)
(What do you think the 12th one will be?) The
12th one will be a C because you're continuing the
progression. (The 12th one is a P.)
The 13th one will be a C. The 12th one was a P.
You were trying to throw me off. The progression
will continue. (The 13th one is a P.)

(Now what do you expect the first symbol will be?)
P. (OK, the 1st symbol is a C.)

The 14th one will be a P. You're beginning a new
progression ~ith P' s. (The 14th one is a P.)

(OK, now what do you expect the 2d symbol will
be?) It'll be a P. (Why?) It's pictured in my
mind. (OK, the 2d symbol is a C.)

The 15th one will be a P. You're still continuing the progression. (The 15th one is a P.)

I'll say a C. (Why?) Primarily this time because I'm trying to outguess you. (OK, the 3d
symbol is a C.)

(What about the 16th one?) The 16th one will be
a C. • •• to throw me off now. (The 16th one is a
C.)

The 17th one will be a C. You're going to see if
Itll revert to the progression of P's. (The 17th
one is a C.)
*The statements in parentheses are those of the
experimenter.

The 18th one will be a P. You're going to break
this progression of C's. (The 18th one is a C.)

138
3.3
The 19th one will be a P. You're going to get off
this progression of C's. (The 19th one is a P.)

The 39th is a C. You'll continue the progression.
(The 39th is a C.)

The 20th one will be a P. You're going to try to
throw me off trying to make me think that all-think you're going back to the other progression
which I'm confused about now. I don't remember
wbat the last one was--C, I believe. (The 20th
one is a P.)

The 40th is a C. You'll continue the progression.
(The 40th is a C.)

The 21st one will be a C. You won't continue with
the progression of P's. (The 21st one is a P.)
lbe 22d one is a C. You're doing this so that I
might think the P progressi.on will continue. (The
22d one is a C.)
The 23d one will be a C. You're trying to make me
think that the next one will be a P--going back to
the old progression. (The 23d one is a C.)
The 24th one will be a C. You're going to continue the progr~ssion of C's. (The 24th one was a

The 41st is a C. You'll continue the progression.
(The 41st is a C.)
The 42d is a C. You'll continue the progression.
(The 42d is a C.)
The 43d is a C. You'll still continue the progression. (The 43d is a C.)
The 44th is a C. You'll still continue the progression. (The 44th is a C.)
The 45th is a C. You'll still continue the progression. (The 45th is a C.)
The 46th is a C.

greSSion.

You'll still continue the pro(The 46th is a C.)

C.)

The 25th one is a C. You're still going to continue the progression of C's. (The 25th one is

The 48th will be a C. You'll go back to the old
progression. (The 48th is a C.)

a C.)

The 26th one is still a C. You'll continue the
progression. (The 26th one is a C.)
The 27th one is a P. You'll break the progression
now. (The 27th one is a C.)
The 28th one will be a P. You're going to break
the progression now. (The 28th one is a C.)
ThE! 29th one is a C. You're ~ontinuing the progression. (The 29th one is a C.)
The 30th is a C. You'll still continue the progression. (Tht! 30th is a C.)
The 31st is a C. You'll continue the progression.
(The 31st is a C.)
The 32d is a C. You'll still continue the progression. (The 32d is a P.)
The 33d is a C. Yoti gave me a P last time to
thro~ me off.
(The 33d is a C.)
The 34th is a C. You'll continue the progression.
(The 34th is a C.)
The 35th is a P. You're going to throw me off the
progr~ssion.
(The 35th is a C.)
The 36th is a C. You'll
(The 36th is a C.)

The 47th will be a P. You'll now break the progression. (The 47th is a C.)

con~inue

the progression.

The 49th is a C. You'll continue the progression.
(The 49th is a C.)
The 50th is a C. You'll continue the progression.
(The 50th is a P.)
The 51st will be a C. You gave me the P to throw
me off. (The 51st is a P.)
The 52d is a p. You've begun a progression of
P's. (The 52d is a C.)
The 53d is a P. You gave me a C to throw me off.
(The 53d is a C.)
The 54th is a C. You'll continue the progression
of C's. (The 54th is a C.)
The 55th is a C. You'll still continue the progression. (The 55th is a C.)
The 56th is a C. You'll continue the p"rogression.
(The 56th is a P.)
57 is a P. The P will throw me off the progression thinking you had tried to throw me off the C
progression with your last P. (57 you said was a
P?)
P. (57 was a C.)
58 is a C.

You began a progression of C's.

(58

is a p.)

The 37th is a C. You'll continue the progression.
(The 37th is a C.)

59 is'a C. You're still trying to throw me off
with the C's. (59 is a P.)

The 38th is a C. You'll continue the progression.
(The 38th is a C.)

60 will be a P. You're beginning a progression
of P's. (60 is a C.)

139
3.3
61 is a P. You're zigzagging between pt s and
(61 is a P.)

C~s.

84 will be a C. The P's were given to throw me
off. (84 is a P.)

62 is a C.
is a C.)

(62

85 will be a P. You've begun a new alternating
sequence. (85 is a P.)

You'll continue the oscillation.

63 is a C--rather 63 is a P because of the oscillation pattern. (63 is a P.)

86 will .be a C. You're following with a C and 2
P's. Another C will come. (86 is a C.)

64 is a C becuase of the oscillation pattern.
is a C.)

(64

87 will be a P.
(87 is a C.)

65 is a P because of the oscillation pattern.
is a C.)

(65

88 will be a P. You've begun a sequence of 2 Cts
and a P. (88 is a C.)

66 is a C¥ You've begun a progression of C's.
(66 is a P.)
67 will be a C.

(67 is

89 is a C. You've begun a new progression of C's.
(89 is a C.)
90 is a C.
is a C.)

You'll continue the progression.

a C.)

68 is a C. You're having a different type of oscil1ation--2 C's between a P. (68 is a P.)

91 is a C.

The progression continues.

(91 is a

C.)

69 is a C. You're oscillating with C's and P's.
(69 is a C.)

The progression continues.

(92 is a

P .)

70 will be a P.
is a P.)

You're oscillating again.

You'll fOllow the same sequence.

It's the alternate symbol.

(70

71 will be a C because of the oscillation
sequence. (71 is a C.)

92 is a C.

(90

93 is a P. The P's given to me previously to make
me think that the progression was being broken and
that you would revert to it after the P. The next
one will be a P. (93 is a C.)

72 will be a P because of the oscillation
sequence. (72 is a C.)

94 will be a C. You've gone back to the C progression. (94 you say now is a C.) 94 is a C.
(OK, 94 is a C.)

73 will be a C. You've begun a new progression of
C's. (73 is a C.)

95 is a C. You've begun a progression of C's.
(95 is a P.}

74 is a C. You're continuing the progression.
(74 is a C.)

96 will be a C. You're alternating now with C's
and P' s • (96 is a P.)

75 is a C. You're still continuing with the progression. (75 is a C.)

97 is a C. You've begun a progression of a C and
2 P' s • (97 is a P.)

76 is still a C. You're continuing with the progression. (76 is a C.)

9b is a P. You've begun a progression of P's.
(98 is a C.)

77 is a C. You're still continuing with the progression. (77 is a C.)

99 is
and 3
(sic)
a C.

78 is a C.
is a P.)

The progression is continuing.

79 is a C. The P is to throw me off.
gression continues. (79 is a C.)

(78

The pro-

80 is a C.
is a C.)

The progression will continue.

81 is a C.

The progression continues.

(80

(81 is a

a C. You've begun a progression of 3 P's
C's. You've already had the 3 pts. 98
will be a C. (That was ••• 99 is going to be
You said. 99 is a C.)

(What's 1001) 100 will be a C.
progression. (100 is a C.)

It follows the

101 will still be a C. Continue the progression
of 3 P's and 3 C's. (101 is a C.)
102 will be a C. You've begun a progression of
C's. (102 is a C.)

P .)

82 will be a C. You're alternating now with C's
and P's. (82 is a P.)
83 is a P. You'Ve begun a progression of P's.
(83 is a C.)

103 is a C. You'll continue the progressi.on of
C's. (103 is a C.)
104 is a C. You'll continue with the progression. (104 is a C.)

140

3.3
105 will be a C. You'll continue the progression. (lOS is a C.)

127 will be a C. You gave me the P to throw me
off. (127 is a P.)

106 will be a P. You'll break the progression
now. (106 was a C.)

128 will be a P. You've begun a progression of
P's. (128 is a C.)

107 will be a C. You'll continue the progression. (107 was a P.)

129 will be a C. You've begun a progression of 2
P's and 2 CiS. (129 was a C.)

108 will be a C.

130 will be a C.

off.

C' s.

You gave me the P to throw me
The progression will continue. (108 is a

You've begun a progression of
(130 is a P.)

C.)

109 will be a C.

sion.

You'll continue the progres(109 was a P.)

110 will be a C. You"re alternating with C's
and P's. (110 is a C.)

131 will be a P. You're continuing the progression of 2 P's and 2 C's. (131 is a C.)
132 will be a P. You're alternating the signs
now. (132 is a C.)

You'll continue the alternation.

133 will be a C.
(133 is a C.)

You've begun a sequence of C's.

111 will be a P.
(111 was a P.)

You'll continue the alternation.

134 will be a C.
(134 is a C.)

You're continuing the sequence.

112 will be a C.
(112 was a P.)

113 will be a C. You've begun a progression of a
C and 2 p, s • (113 is a P.)

135 is a C. You're continuing with the progression. (135 is a P.)

114 will be a P. You've begun a progression of
p, s • (114 is a P.)

136 will be a P.

You've begun ••• you're trying to
throw me off now with a 2d P. Think there would
be only one P. (136 is a C.)

115 will be a P.
(115 is a C.)

You'll continue the progression.

137 is a C. You're going to continue with the
progression of C's. (137 is a C.)

116 will be a p.
(116 is a C.)

The C was given to throw me off.

138 is a C. You'll continue the progression.
(138 is a C.)

117 is a C.

and

4

C's.

You've begun

a

progression of 4 P's

(117 is a P.)

118 will be a P. The progression has changed
from 4 P's and 4 C's to 4 pes and "3 C's. (118 is

139 is a C. You'll continue the progression.
(139 is a P.)
140 is a C. The P was given to throw me off.
(140 is a P.)

a C.)

119 will be a P. You're alternating with C's and
P' s • (119 is a C.)

141 is a C. You gave me the 2 C's (sic) for the
same reason as the previous time you "had given me
the 2 C's 'er 2 P's... (141 is a C.)

120 will be a C. You're continuing the progression. (120 is a P.)

142 is a C. You'll continue with the progression.
(142 is a C.)

121 will be a P. You have a progression'of 2 C's
and 2 P's. (121 is a P.)

143 is a C. You'll continue with the progression.
(143 is a C.)

122 will be a C. You'll continue this progression
of 2 and 2. (122 is a C.)

144 is a C. You'll continue with the progression.
(144 is a C.)

123 will be a C. You're continuing the progression. (Of what?) Of 2 C's and 2 P's. (123 is a

145 is a P.
is a C.)

You'll break the progression.

(145

C.)
124 will be a C.

C's.

You've begun a progression of
(124 is a C.)

124 (sic) will be a C. You're continuing the progression. (125 is a C.)
126 will be a C. You're continuing the progression. (126 is a P.)

146 is a C. You'll continue the progression.
(146 is a C.)

147 is a C. You'll continue the progression.
(147 is a C.)
148 is a C. You'll continue the progression.
(148 is a C.)

141

3.3
149 is a C. You'll continue the progression.
(149 is a C.)

171 will be a P.
(171 is a P.)

You'll begin a sequence of P's.

150 is a C. You'll still continue the progression. (lSO is a C.)

172 will be a C.
is a C.)

You'll revert to the C's.

151 is a C. You'll still continue the progression. (lSI is a C.)

173 will be a C. You're alternating with 2 P's
and 2 C's. (173 is a P.)

152 will be a P.
(l52 is a C.)

174 will be a C.
(174 is a C.)

You'll break the progression.

(172

The alternation is a C and a P.

153 is a C. You'll continue the progression.
(153 is a P.)

175 will be a P. You'll continue this alternation. (175 is a C.)

154 is a C. You've broken the progr~ssion and
you'll revert to it now. (154 is a C.)

176 will be a C.
(176 is a P.)

You've begun a sequence of C's.

155 is a C. You'll continue the progression.
(155 is a P.)

177 will be a C.
gression of C's.

You'll continue with the pro(177 is a P.)

156 is a C. You're alternating with P's and C's.
(156 is a C.)

178 will be a C. You've begun a progression of 2
P's and 2 C's. (What did you say 178 was?) A C.
(178 is a C.)

157 is a C.

The alternation of P's and C's was to
throw me off the progression of C's. The C progression will continue. (157 is a P.)
158 is a C. You're still going back to C sequence. (158 is a C.)

159 is a C. You're still going to continue this
sequence. (159 is a P.)
160 is a C. You have an alternating sequence of
P's and C's. (160 is a C.)

161 will be a P.
(161 is a P.)

You'll continue with another C
to complete the sequence of 2 P's and 2 C's.
(179 is a C.)

180 will be a P.
(180 is a C.)

You'll continue this sequence.

181 is a C. You've begun a sequence of C's.
(181 is a C.)
182 is a C.
is a C.)

You'll continue the sequence.

(182

183 is a C.
is a P.)

You'll continue the sequence.

(183

You'll continue to alternate.

162 will be a C. You'll continue this oscillation. (162 is a P.)

163 will be a C.
(163 is a C.)

You'll continue the alternation.

164 will be a P.
(164 is a C.)

You'll continue the alternation.

184 will be a C.
(184 is a C.)

The P was given to throw me off.

185 is a C. You'll continue the sequence of C's.(185 is a C.)

165 will be a P. You'll go back to the alternation. (165 is a C.)
166 will be a C.
(166 is a C.)

You've begun a sequence of C's.

167 will be a C.
(l67 is a P.)

You've begun a sequence of C's.

168 will be a P. You've begun a sequence of 2
C's and 2 P's. (168 is a C.)
169 is a C. The previous P's were given to throw
me off. You'll continue the sequence of C·s.
(169 is a C.)
170 will be a C.
(170 is a P.)

179 will be a C.

You'll continue the sequence.

186 will be a C.
(186 is a c.)

You'll continue the sequence.

187 will be a C.
(187 is a C.)

You'll continue the sequence.

188 is a C.
is a C.)

You'll continue the sequence.

(188

189 is a C.
is a P.)

You'll continue the sequence.

(189

190 will be a C.
(190 is a C.)

The P was given to throw me off.

191 will be a C.

The double P (sic) was given to
throw me off a little more. (lql is a f"' . )

You've ••• been giving me a sequence
of 2 pes and 2 (;'s. (192 is a C.)

192 is a C.

142

3.3
192 (sic) is a P. You're continuing the sequence
of 2 pes and 2 C's. (193 is a C.)
194 is a C. You've begun a sequence of C's.
(194 is a C.)
195 is a C.
is a P.)

You'll continue the sequence.

(195

196 will be a P. You have a sequence here of inserting 2 pes. (196 is a C.)
197 is a C. The P was given to throw me off.
(197 is a C.)

19.8 will be a C.
(198 is a C.)
199 is a C.

You'll continue the sequence.

You'll continue the sequence.

(199

is a C.)
200 will be a C.
(200 is a C.)

You'll continue the sequence.

A. COULD THE PATT~~N CUMPUN~NT UF THE ~~tOlCTIUN-HYPOTHESIS
FOR TRIAL T HAV~ PRtuICTtu Tht EVtNT OF TRIAL T CO~RECTLY.
b. YES-EXPLANATIU~-HY~0THESIS FU~ TKIAL T IS THE PATTE~N
COMPON~iH OF THl:: Pl-ID The.. P .... t.l)l<..TI().~

hY~UTrH:.5I:.S

FOt<

TkIALS T-1 AND

T-2

WEkE THE
CORRECT.
1... Yt.S-PKt.ulCTIvi~-r1Yr-uTdtSlS l-vl< TKIAL 1+1 IS GUESSUl-itJOSITc THe. c.X~LANATIO~-HYPUTHe.~I& fO~ TRIAL T.
M. NU--PKEulCT10i\-HYPUTH~Sl~ l-OU TRIAL T+1 IS THE
EXPLANATION-hYPOTHESIS l-OR T~IAL T.
~i I LL THe. f.:JCtJLANA T I Oi-i-Hypr) THES I S FO~ TR 1 AL. 1 CONT I NUE..
(SlE TEXT FOR AN EX~LANATIOh UF THIS TEST)
O. YES-PREDI~TION-HYPOThl~l& FOR TRIAL T+1 IS THE
CO~TAIN GUES~-0PPUSITE COMPO~£NTS A~D
j-IRtDltTIUN~ Uf Th~ /.:;.Vc..NTS Uf THt:SE T~IALS

~h

f:.XPLANATION-HYPOTI1t..SIS FOR TRIAL T.
p. NO--PK(DICTION-HYP0THESrS FOR TRIAL T+l IS GUESSvPPOSI Tf TH~ i;:XPLAf"AT I(.Ii,-hYPUTHESIS FOK Ho/'ho attempted to solve a "concept learning"
problem which had three logically correct
answers; a disjunction, a conjunction, and
a relation. (This problem has been described previously9 and same data on its
difficulty ..las available.) All three conditions of presentation 'Ivere given to each
sUbject. The model 'to/'e have just presented
gave the best overall "postdiction" of
responses of any model 'YTe could devise.
In fitting it 'Vle altered the order and
identity of symbols on reference lists, but
othervlise kept the model constant. Since
each subject solved three problems, we were
able to make some tests of our transfer procedures and thus do not rely too heavily upon
pre-specified orders. The results of our
match were generally encouraging. However,
they cannot be taken as validating evidence
since the protocols were used to develop
the program.
Some more encouraging evidence came
'.·,hen the artificial subject attempted a
series of problems used by Shepard, Hovla.~d
and Jenkins21. This 'Ylas a completely
separate study. Human subjects 'Ylere asked
to find categorizing rules for each of the
six possible types of splits of eight instances, each describable by one of two
values on tr.ree dimensions, into t'l1O sets
of four each. Huoan subjects could solve,

A somewhat similar, unpublished,
experiment was performed by Hunt and H. H.
vlells. Here the five commonly used logical
connective betvleen two elements provided
the anS~·ler. A 11 truth table" was constructed
and presented to subjects in geometriC form.
For example, the connective II p and q" might
be represented by ured and starT The five
problems ..lere presented in five orders,
each subject solving all five problems in
one of the orders. Simulation and analysis
of this experiment has not been completed
at the time of this 'Yr.citing, however, vie have
same preliminary results. There is good
general agreement bet'lo/'een our simulation
routines and some protocols. Both the
computer model and the subjects are sensitive
to the order in 'I'7hich problems are presented,
but their reactions are not as similar as
'-Ie vlould like. A neyl transfer procedure is
needed. In an experiment which is not directly
related to simulation, ~'lells is studying the
manner in vlhich human subjects learn methods
of solution for disjunctive problems. We
hope that his experiments .vill provide some
clues about the nature of the transfer
procedures Y/e should include in our model.
;"je do not claim to have presented a
complete explanation of qoncept learning!
Certainly others Ifill agree 'Ylith us. In
programming the model ife made many decisions
ivi th little theoretical or empirical justification. Some of these are certain to be
wrong. But ~'lhich ones'?

He sha.ll probably ha.ve to change our

151

3.4
routines for memory and recognition. Same
of the known phenomenon of memory cannot
be reproduced by a simple occupancy model.
For instance, the effect of stimulus similarity upon memory cannot be represented. Our
model bas an all or none aspect in its
interference features. An intervening
instance either completely eliminates the
record of a previous instance or does not
affect it at all. This does not seem to b~
the final answer to the problem of memory in
concept learning.
Two alternative memory systems have been
considered. One system retains and extends
the limited occupancy model. Instead of
storing one IIcodeword" (actually, a list
structure), representing all knovTn information about an instance, on a single occupancy
list, several code vlords would be stored in
several occupancy lists. Each of these code
words would represent a particular type of
information about some part of the instance
in question. Storage of each code\vord would
be independent on each occupancy list. Code..vords referring to the same instance 1-lould
reference each other's locations. rlhen information fram memory '\-Tas required a picture
of each instance would be reconstructed from
the cross referencing system. HO\'Tever, since
intervening instances ''i'ould be storing codewords independently on each occupancy list,
some of the codevTords might be replaced.
The extent of this replacement would depend
upon the similari ty bet~'i'een the instance to
be recalled and the stimuli which follovled
its presentation. This system would be
sensitive to stimulus similarity effects.
Alternately, He could use an associationist memory system. Instead of trying to
remember units of information directly we
would build II associations II bet"i.Teen names and
sti.rnulus feature:::;. This is the logic of the
technique used by many learning theorists
in psychology. Machinery to realize such
a memory has be enS extensively investigated
by Rosenblatt17,l. There is also some
similari ty bet"leen this approach and the
classification mechanisms based upon Selfridge~
"Pandemonium" scheme19 • To adopt such a
memory system vTould require changing the
entire logic of our model. Association
schemes generally contain, in themselves,
a mechanism for concept learning. It also
seems that they require some sort of gradient
of generalization. Recent experiments 20 ,21
indicate that, in concept learning, the
tendency to code stimuli symbolically plays
a greater role than generalization based
upon stimulus Similarity. For these reasons

we ~ve, tentatively, rejected an associationist memory mechanism.
In the present model '-Ie subject the formal
description of an instance to t'·l0 transformations. \wen an instance is presented the
dimensions of the formal description are
sampled to determine what information is to
be placed in memory. At some later time,
that part of the formal description which is
in memory is re-transformed to provide a·
''i'orlting description. The two procedures
could be combined if the description routine
currently at the head of the description
routine reference list viere to be applied
directly to an instance before it entered
memory.
Such a procedure would have advantages
in saving storage space. Instead of having
to have t-';-lQ separate locations, for working
and permanent description, in the internal
memory, o~ one description need be stored.
But . .·re pay for saving this space by losing
information. By definition, any working
description can be derived from the formal
description. All vlOrking descriptions cannot be derived from each other. For instance,
if vIe know that an instance contained t,'i'O
figures of the same color, ·we do not know
what that color is. As a result, our artificial subject's ability to utilize a
particular description routine at time t
would depend very much upon the description
routines used previously.
The role of IIset ll at time of presentation
as a determinant of later memory characteristics needs more extensive investigation. Same
experiments12 ,13 suggest that 11 set" is a
function of how memory is searched rather
than hOl-l i terns enter into memory. Also,
there exists a rather contradictory literature
on lllatent learning", a term used to describe
experiments in which animals, trained to
respond to cue A in the presence of cue B,
which is irrelevant to the animal's current
need, learn more rapidly a later response
to cue B. From present experimental results
it is not obvious hOI'; stimulus recognition
and ans'l-ler development procedures should be
connected in a concept learning simulation.
Procedures for representing transfer
may not be represented adequately in the
present model. Transfer is defined as the
effect of previous problem solving experience
upon solution of the problem with which the
subject is faced at the time of the test.
We decided to work first with a simple
method of representing transfer, in which

152
3.4

the subject tries \.,hatever vlorked last time.
A principal result of the simulation of the
Hunt and Hells work on logical connectives
has been a demonstration that a ne\{ transfer
procedure is needed.
In the tradition of classical learning
theory, we could attach a modifiable numerical
index to each routine on a reference list.
This index could be used to determine the
probability that a routine Hould be selected.
This method of representing learning is
probably the most connnon. The prinCipal
objection to it is that it implies the
existence of II counters in the head" and,
essentially, makes our program a digital
simulation of an analog system.
The alternative to association indices
is a new method of ordinal rearrangement
of routines on a reference list. The problem
with ordinal re-ar~angements is that they
do not permit us to specify a variable distance bet"leen routines on a list. Suppose
vle consider each concept learning problem
as a contest betw'een routines on the same
reference list. The one that finds a place
on a successful execution list is victorious.
How many times must the routine in position
£ "ivin" before it gains the next hiGhe st
position? Should it jump pOSitions? As
vTe have indicated, some research relevant
to this topic is being conducted.
Conceivably, Ive may have to change our
entire method of transfer. At present our
model records ansvlers, 1..;1th associated information about useful routines. T,;e could
attach to routines information about problems
on ,\-Thich they had been useful. 'de Ilould then
have to develop some i'lay for the artificial
subject to extract, rapidly, key features
of a problem "i-lhile the anmrer is being
developed~ Houtines vlOuld be examined to
see "That, in the light of past experience,
'\-las their probable usefulness on this sort
of problem.
Closely related to the problem of transfer
is the problem of response selection d~~ir~
learnill3. Our present model rearranges its
order of response selection after a problem
is solved. During a problem, response selection
is controlleli by time parameters which are independent of prog:'"'am contl'"'ol. No use is made
of intermediate computations in selecting
the next item to be placed on an execution
list. In an alternate model this miGht be the
controlling factor. The means-end analysis
of the Loeic Theorist15 uses intermediate

calculations heavily. Amarell has proposed
a computer model for an area very similar
to ours in ivhich intermediate computations
exert control on ans'\'ler development.
Our simulation '\'lork, and analysis of
experimental data, has convinced us that
some method of making the choice of one
item on an execution list dependent upon
the product of execution of previously
selected routines is desirable. ~~t is
not clear is the optimum amount of dependency.
Bartlett 2 has presented an analogue, in an
entirely different context, vlhich may clarif'y
the problem. He compared problem solving
and thinl~ing to motor skills responses, such
as serving in tennis. There are certain
pOints at vlhich a chain of responses can
be altered, in betT,Teen these points a series
of acts vTill be executed uithout interruption.
Our problem, experimentally, is to identify
the responses and choice pOints.
He feel that the principal use of our
model, so far, has not been in the generating
of an eA~lanation of concept learning so much
as it has been in indicating the type of new
experimental data needed. J,:e have had to be
very specific in our thoughts as we programmed
this model. Ac a result, we have reached
some conclusions about the kind of experiments
that need to be done. It may well be that
the typical concept learning experiment
confuses three processes; memory, recognition,
and symbolic problem solving. It is not
clear whether or not these should be treated
as part of' a unitary "concept lea.rningll act.
They can be programmed separately. In addition,
He have become concerned vlith questions of
transfer, the effect of the subject's current
hypothesis upon his later retention of information, and the effect of time pressure
upon information processing. A real ailareness of these problems has been a major outcome of programming a concept learning mouel.
Comparisons with Related Work
VieT,-led formally, our problem is closely
related to models of pattern recognition.
ProgramminG either a pattern recognizer or
a concept learner involves the development
of a mechanism which operates on a specified
stimulus universe to map stimuli from predetermined subsets into particular responses.
Because of this mathematical identity, at
least one critic lO has suggested that problems
of this sort should be treated together,
without "psychologizing" or lIneurologizing."
\ihile this may be useful in developing

153
3.4

theorems about a canonical form of categorization, it may not be an appropriate strategy
for simulation studies. In particular, our
approach is quite different from that of the
pattern recognition studies i-lith "lhich we are
familiar.
The most striking difference is in the
manner in \'lhich '.le pre-code the stimuli.
Pattern recognizers usually accept stimuli
coded into projections on a grid. The result
is a string of bits, each bit representing
the presence or absence of illumination of'
some pa..l't of the grid. The Game representation could be used for a temporal pattern.
:2ach bit T,'lOuld stand for the presence or
absence of some stL~ulus feature.
He presuppose the existence of a dimension
and value coding 6 and deal i1ith perceptual
as:gects '.Thicn are readily verbalizable. A
pattern reco:,;nizer develops its mTn code.
.\ny coding scheme developed by a pattern recoe;nizer '.7ill be speci.?ic to the stimuli used
(visual VB. audi to:ry, etc.). Since ,Te are
interested in the manipulation of coded elements
"7e avoid this problem by fiat in our nro!."ramming and by explicit instructions to ~uro
subjects in our experimental 'tjork.
Our model is also different from most
pattern recognizers in the processes it uses.
Pattern recognizers, at least as developed
by self:ri~~ aga. his co--workers 19, and by
Rosehblatt 7,1 , are basically parallel
processing devices ,;hich utilize a large number
of redunclant, e:..~ror prone tests. Our proGram
is a serial processor i{hich tries to develop
a single, perhaps complex, error free
classification test. 'He do not see any incompatibility in the t',lQ approaches. Pattern
recognizers are inspired by the capability
of biological systems to amplify u:90n their
sensory inputs. Our program deals ''.'1ith
the simulation of a symbolic process. That
the tvo problems are formally similar does
not mean that they are realized in the same
way by problem solvers.
In principle, there ,vould be no objection
to utilizing a pattern recognizer to provide
the input to the concept learner. The combined system could develop its own dimensions
and values and then operate upon them. In
practice, such a scheme is undoubtedly premature. But it is a long ran3e goal.
The concept learning problem has been
attacked directly in two previously mentioned
studies by Kochenll and A.ma.rell • Kochen
restricted his program to solution of

concepts II based upon a conjunctive rule
involving stimuli specified by strings of
bits. His program consisted of executing
algorithms upon the information about the
universe of objects which was available
~ ~ ~~, in memory.
The program
also contained features for making random
guesses about the correct concept. These
guesses could be weighed for II cOnf'idencE;" ,
using an index ,vhich satisfied Polya' s16
postulates for plausible reasoning. One
of Kochen's findings, based on Monte Carlo
runs of his system, was that changes in
the value of the confidence index could be
used to estimate the probability that an
anS'lver "las correct before a proof of the
anSi-leI' 'fas available.
It

Amarell proposed a machine that could
generate routines to map arguments to values
in symbolic logic. The key feature of his
proposal, one \ole might '\-lell adopt, is his
use of intermediate results to IImonitorll
future anS\ver development.
Neither Kochen nor Amarel "lere directly
concerned "lith simulation of human performance. This difference in goals, and
features of~ programming, are the major differences bet'-leen our work and theirs.
Superficially, our program is similar
to the list processing programs \~itten by
the Carnegie Institute of TechnologyRM1) Corporation group headed lly Ne'-lell,
Shavl, and Simon, and McCarthy! and his
associates at M.I.W. In particular, the
,.;ark of Feigenbaum at Carnegie, is related to ours. He developed a program to
simulate paired-associateG learning. As
part of his program he included a routine
for selective recognition of stimulus
features. As more experience idth the
stimulus universe '(las provided, more
features '\-lere read into the system to
enable it to make finer discriminations.
The logic of Feigenbau~fs recognizing
systen, and in particular its capability
for dropping stimulus features 1lhich are
not useful in discrimination, could be incorporated into our program.

I,

Our present program, although running
nOvl, is in no sense complete. Almost every
ne\ol simulation has indicated ways in which
it could be improved. ~le intend to continue
to investigate concept learning by
use of an information processing model.
But we do wish to add a word of caution.

154

3.4
Neither our model, nor any other, has generated
a large number of new experiments. This is
a traditional test of the utility of a scientific model, and it is going to have to be met
by us and by others interested in this field.
We do not feel that the utility of computer
programming models in psychology has been
proven or disproven. The jury is still out.
We, of course, hope that a favorable verdict
will be returned.
References
1. Amarel, S. An approach to automatic
theory formation. Paper presented at
the Illinois Symposium on the Principles
of Self Organization, 1960.
2.
3.

Bartlett, F. C. Thinking.
Basic Books, 1958.

Church, A. Introduction to mathematical
logic, vol. I. Princeton, N.J.:
Princeton U. Press, 1956.
An information processing
theory of verbal learning. RAND Corp.
publication, p. 1817, 1959.

6.

7.

14. McCarthy, J. Recursive functions of
symbolic expressions and their computation by machine. Communications
2! ~ Association for Computing
Machinery, April, 19bo.
15. Newell, A. and Shaii, J. C. Programming
the logic theory machine. RAND Corp.
publication, p. 95 4, 19)7.
16. Polya, G. Mathematics and plausible
reaponing. Princeton: Princeton U.
Press, 1954.

New York:

4. Feigenbaum, E.

5.

13. Lawrence, D. H. and LaBerge, D. L. The
relationship betvleen recognition
accuracy and order of reporting
stimulus dimensions. ~.~. Psychol.,
1956, 51, 12-18.

17. Rosenblatt, F. The perceptron, a
probabilistic model for information
organization and storage in the
brain. Psychol. ~., 1958, 65,
368-408.
18. Rosenblatt, F. Perceptual generalization
over transformation groups. In Self
organizing systems, London, PergaiiiOii
Press, 1959.

Green, B. F. IPL-V, the Newell-Shaw-Simon
programming language. Behavioral Science,
1960, 5, #1.

19.

Selfridge, O. and Neisser, U. Pattern
recognition. Scientific American,
1960, 203, 60-79.

Hovland, C. I. A "communication analysis"
of concept learning. Psychol. Rev., 1952,
59, 461-472.
-

20.

Shepard, R. N. and Chang, J. J. Stimulus
generalization in the learning of
classifications. Bell Telephone Lab.
mimeographed report, 1961.

21.

Shepard, R. N., Hovland, C. I. and
Jenkins, H. 1,1. Learning and memorization of classifications. Psychol.
Monogr., 1961, in press.

Hovland, C. I. and Hunt, E. B. The computer
simulation of concept attainment.
Behavioral Science, 1960, 5, 265-267.

8.

Hunt, E. B. An experimental analysis and
computer simulation of the role of
memory in concept learning. Unpubl.
Ph.D. dissertation, Yale U., 1960.

9.

Hunt, E. B. and Hovland, C. I. Order of
consideration of different types of
concepts. ~.~. Psychol., 1960, 59,
220-225.

Acknowledgment

10. Keller, H. Finite automata, pattern recognition, and perceptrons. AEC Computing
Center and Applied Mathematics Center,
Report NYO-2884, 1960.
li.

Kochen, M. Experimental study of hypothesis
formulation by computer. IBM Research
report, RC-294. IB1 Research Center,
Yorktown Heights, New York, 1960.

12.

Lavrrence, D. H. and Coles, G. R. Accuracy
of recognition with alternatives before
and after the stimulus. ~.~. Psychol.,
1954, 47, 208-214.

The work reported in this paper was
supported by a grant for the study of concept
learning, from the Ford Foundation to Carl
I. Hovland. The computational work involved
was supported, for the most part, by the
Computation Center, l~ssachusetts Institute
of Technology. Dr. Bert F. Green, Jr., and
Alice Holf, of Lincoln Laboratory, M.I.T.
made available the 709 version of the IPL-V
interpreter and instructed the first author
in its use. The aid received is greatfulJ~
acknowledged.

155
3.4

Subject

Figure 1.

Program Control Chart.

Problem
Characteristics

Time

Checks

Internal
Memory

Figure 2.

Answer Developing Procedure.

157
4.1
PARALLELISM IN COMPUTER ORGANIZATION
RANDOM NUMBER GEnERATION
IN THE FIXED-PLUS-VARIABLE COMPUTER SYSTEM
M. Aoki*, G. Estrin*, and T. Tang**
Department of Engineering
University of California*
National Cash Register Co.**
Los Angeles, California
Summary
TV e Fixed-PIus-Variable Structure Computer
System utilizes an inventory of modules which can
be interconnected as special purpose configuratwns
operating simultaneously with other parts of the
system. Since the structure considered makes no
permanent committment of hardware to relatively
rarely used operations it permits reconsideration
of designs previously discarded as uneconomic.
Problem formulations utilizing random numbers generally require large nuobers of trials to achieve
high confidence results. Despite the fact that in
most problems random number generation does not r~
quire a large percentage of the total computing
time, it may be desirable to eliminate even that
time by use of special purpQse random number generators.
Special purpose circuits can be designed to
generate pseudo-random numbers in parallel with
other activities such that these numbers are available on demand in the same sense as any other operand stored for use in the computation.
This paper discusses a number of different
methods of generating pseudo·random numbers, the
time required in existing programs, the hardware
implications of different parallel and serial designs. The criteria for choice of one method over
another in the context of particular problems and
the Fixed-PIus-Variable Computer System are evaluated.
1.

Introduction

T~e Fixed-PIus-Variable Structure Computer
System is organized such that there is associated
with a general purpose computer an inventory of
modules which may be reorganized as needed to reduce the overall computation time required for solution of particular problems. The speci81 purpose
configurations established in the variable structure
part of the system ~~y carryon their operations
simultaneo~sly with those of the fixed part of the
system anQ may have their configurations altered
during the course of a given problem.

In general, of course, any of the operations
!he preparation of this paper was sponsored, in
pqrt, by the Office of Naval Research and the
Atomic Energy Commission. Reproduction in whole or
part is permitted for any purpose of the United
States Government.

which may be executed by the special purpose configuration may either be programmed in the fixed
structure general purpose part of the system or
may be added to the list of permanently available
commands. The latter course will not in general
be taken for relatively rarely used operations and
both methods imply sequential operation on all
commercially available systems. Since in basie
concept the (F + V) system makes no permanent
committment of its variable structure inventory,it
is reasonable to consider its application to many
problems It,here special purpose techniques would
previously have been discarded as uneconomic.
This paper considers the class of problems
which makes use of long sequences of random numbers for simulation ~!4complex processes using
Monte Carlo methods.
In most such problems the
operations which must be done on the random number
are more time consuming than the random generations
themselves and deserve most of the attention of speCial
purpose equipment. However i t will be found that the
large number of operations generally will make it
worth while to utilize some fraction of the variable structure inventory to essentially remove the
time for random number generation from the overall
computation time.
2.

Methods for Generating ana Testing
Random Numbers

2.1 Introduction
The successful ~~e of the t10nte Carlo method
in digital computers
depends on*a good supply of
long sequences of random numbers.
There are,
broadly speaking, four ways of supplying such
sequences.
* During the course of the calculations associated
wi th any Monte Carlo problem, it is also necessary to
produce random variables accordingtoa variety of
different probability distributions. A co~~onmethod
of doing this is to incorporate into the computing
machine a sequence of uniformly distributed random
numbers. The desired random variables ~,e then obtained by transfor~ations based on them' •
There are three methods which can be adapted to
serve this purpose. These are based on direct
functional transformations, the principle of compound probabilities, and the procedure of rejecting
part of the sampled ~a$ues according to an appropriate test or rule.' The first is called the
"direct method", the second the "composi tion method"
and -the third the "rejection method".

158

4.1
1. Tables of random numbers may be recorded
on paper tape, punched cards or magnetic tape. The
tapes or cards are fed into the comnuter at suitable stages of the calculation. This method has
found little favor in those problems which require
very long sequences of random numbers since the
time taken to read in the tables may soon become
excessive. For example the IBM 711 card reader
operates at the rate of 250 cards per minute or
7
6,000 words per minute.
If a problem requires one
million random numbers, it will take the card
reader 166.66 minutes to transfer the data into the
computer. If magnetic tape is used, over 100 seconds
would be required on the fastest available magnetic
tape sys tem.
2. Random numbers may be generated by physical processes such as radioactivity or discharges
in gases. The chief objection is a rather paradoxical one; the number sequences cannot be repeated
and so it is very difficult to check the calculation because it is not always possible to distinguish between variations in re~ults due to random
fluctuations and those due to changes in the program or even to the faulty running of the computer.
3. Pseudo-random numbers may be generated by
arithmetical processes on general purpose high
speed digital computers. The most common methods
are the mid-square process ana 9dditive and multiplicative conqruence methods. '
Whenever the time required for generation
of pseudo-random numbers in a problem is significant, a special purpo~e computer or wired-in instruction may be considered utilizing any of the
methods in (3). Such a special purpose configuration would essentially deliver a pseudo-random
number on demand and might operate in parallel with
any other processes going on.
There are three criteria which should be satisfied by any method of generation of pseudo-random
numbers.
1. The numbers generated should satisfy the
tests of randomness prescribed by the user.
2. The rate of generation of pseudo-random
numbers should compare favorably with the rate at
which they may be used in typical computations.
3. The recursion relation should produce a
sufficiently long sequence of pseudo-random numbers.
All sequences generated by arithmetic process are
either cyclic or enter a cycle if prolonged far
enough. A cyclic sequence will not be a satisfactory source of pseudo-random numbers unless the
period is so ereat that it is of no consequence in
practical computations.
2.2 The Testing of Sequences for Randomness
10

The tests of Kendall and Smith
have been
used frequently on sequences of decimal digits.
They may be adopted in various ways for testing
sequences of binary digits generated by pseudorandom number generators. These tests are the frequency test, the serial test, the poker tp.st and
runs test. In all these tests the deviations of the
observed counts from those expected from a perfectly random sequence are studied. Chi-squared

tests are usually used to give a measure1~f the
permissible deviations from expectations.
In practical applications of various tests of
Significance, the 5%, 1% and 0.1% levels of significance are often used. Some empirical results related to the fo~r tests below gan be found in
papers by Green and Rotenberg •
1. The frequency test. The primary requirement for randomness is that the numbers, considered
as fixed point, positive fractions, be uniformly
distributed over their range (0 ~ X ~ 1)
To test this, the binary numbers can be
separated into octal digits, and a tally of the
frequency of occurrence of each octal numeral can
be made and tested since a random process should
produce uniformity in every octal digit position.
2. Serial correlations. Another vital requirement is that the successive generated numbers
be independent of others.
The correlation coefficient ~ of two onedimensional random variables X and Y is defined by

f
'Vf20 }A02
where

j)v . .
1J

= E[(X-E (X)yi

If the two random variable are independent
we have '11=0 and thus f=o. Thus two independent
randQm varlables are always uncoffelated, although
the converse is not always true.
One can com~ute from the generated sequence
of pseudo-random numbers the f between the successive terms and between two numbers k positions
apart, k>1, by defining X and Yappropriately.
If the computed ~ is far from zero one rejects the hypothesis that the sequence of pseudorandom numbers are independent, although one has no
guarantee of their independence even if the computed Y is close to zero.
3. The poker test. The mutual independence
of the various digit positions can also be checked
by the poker test. For each generated number, a
designated five octal digits are treatedas a poker
hand, without regard to suit, and the hand is
tallied as five of a kind, four of a kind, full
house, three of a ki~d, two pairs, one pAir, or bust.
Table 1 shows the expected frequency of poker hands.
Table 1 12
Class

Symbol

Busts

abcde

EXEected

freguenc~

302.4

Pairs

aabcd

504.0

Two Pairs

aabbc

108.0

Threes

aaabc

72.0

Full house

aaabb

9.0

Fours

aaaab

4.5

aaaaa

0.1

Fives

1,000.0

159
4.1
Good pseudorandom numbers should not dethese expected frequencies.
4. Runs test.
Another aspect of serial dependency can be checked by finding the distribution
of runs above and below some constant and runs up
and down as described below. If the observations
are randomly drawn from the same population we do
not expect very long runs of either type, and the
occurrence of such runs will, therefore, usually be
taken as indicating non-randomness.
13 14
a. Runs above and below the median. '
Consider a random arrangement of n elements consisting of n a's and n b's, n = n + n , such as
the followink arrangeme~t of 11 a's1and214 blS:
viate

Table 2

significantlY1~rom

babaaaababbbaababbbbbabba
An uninterrupted sequence of elements
of the same kind is called a run, and the length of
a run is given by the number of elements defining
the run. The sequence above begins with a run of
one b, then follows a run of one a and so on.
Assuming that all possible arrangements occur with the same procability we may find
the distribution function of the number of runs of
a's of length i by means of combinatorial theory.
For small values of n the distribution function of the mean number of runs of a's of
length i, R ., and R , similarly defined for runs
2i
1
ofb's,
ma~ be taoul.ated by writing dOvln all
possib~e arrange~ents of a's and bls, as shown in
Table 2 for n=6 and n = n = 3. The numcer ofdif2
1
ferent arrangements is (6) = 20. The expected number of runs of both kind~ of ele~ents of length k
or more is approximately given by
(2.2)

from which we get the expected total number of runs,
R, as
_ n+2
R - 2
The theory above may be applied to
samples from a population with a continuous distribution function by cl~ssifying sample values larger
than the sample median as a's and values smaller
than the median as b's.
In some cases the population median
is known or a hypothetical value is tested. It
will be noted that in this case the number of a's
will be a random variahle, having a binomial distribution with n1/ = 0.5.
n
14
b. Runs up and down.
Consider a sequence of n different observations, X , X ••• ~ X
2
and the sequence of signs (+ or -) of 1 the (n-1; n
differences Xi 1- X.• A sequence of successive +
signs is called a ran up and a sequence of successive - signs is called a run down. The length ofa
run is given by the number of equal signs defining
the run. The total number of runs is denoted by ~
the number of runs of length i by r and the numi
ber of runs of length k or more by Rk , Rk = Z r i .
i

.lk.

No.
1
2

3
4
5
6
7
8

9
10
11

12
13
14
15
16
17
18
19
20

Arrangement
aaabbb
aabbba
abbbaa
aababb
aabbab
abaabb
abl)aab
ababba
abbaba
ababab
bababa
baabab
babaab
baabba
babbaa
bbaaba
bbabaa
baaabb
bbaaab
bbbaaa

R

R11

2

4
4
4
4

1
2
2
2
2
2
2

5
5

3
3

6
6
5

3
3

Total

80

40

Mean

4

2

3
3

1
1

1
0

o
o

1

1
1
1
2

o
o
o
o

102
102
102
002
002
003

o
1

o
o
o
o

003

5

2
2

103
103

4
4
4
4

2
2
2
2

1

3

1

3

1

0

1
1
1

2

16

4

0.8 0.2

0

0
0
0

0

o
o
o

2

102
102

o

1
1
1

o
o

2
2
2

o

40
2

16

4

0.8 0.2

For example, the sequence of 15
elements: 39,42,38,53,51,30,40,28,43,46,52,55,29,
2~34, leads to the following sequence of signs of
differences between successive elements:
+ - + -- + - + + + + - - +

The sequence is thus characterized by
4 runs up of length 1, one run u~ of length 4, 2
runs down of length 1, and 2 runs down of length 2,
giving a total of 9 runs.
Assuming that all n! possible arrangements of the n numbers occur with the same probability, we have the following mean number of runs.
-

2

r

ri=Ti+3YLn

_ i ...
i ~ n-2 (2.4)

and

Table 3 shows the expected number of runs up and
down of length k or more in rando~ arrangements of
n different numbers. The values RK computed from
generated pseudo-random numbers should not again
differ too much from the theoretical values.
5. Discussion. In a particular application
a block of digits of a definite length may be required to be random when consid~Oed in isolation.
The remark of Kendall ann Smith
is relevant here,
namely "if a series S is locally random in a Domain,
it does not follow that any part of S is locally
random in th~t Domain.f! They conclude that a set
of random numbers which is adeQuate for all requirements is impossible, and the only solution
is to carry out tests on blocks of numbers and
give the results of these tests, so that the prospective user can choose from the tables these

160
4.1
Table 3

venient for binary machines is
r

~

k

(1/3) (2n-1)
2

(1/12) (3n-5)

3

(1/60) (4n-11 )

4

(1/360) (5n-19)

5

(1/2520) (6n-29)

6

(1/20160) (7n-41)

7

(1/181440) (8n-55)

X. = (X. 1 + X. ) Mod 2
J
JJ-n
r
where 0 <. X. < 2 and r is the maximum number of bits
used to endode each fixed pOint number.
The reason that equation (2.7) is very
con~e~ienf8for binary machin~~s that t~e modulus
addltlons
are performed by slmply addlng and disregarding overflow.
To illustrate the behavior of this type
of sequence, consider the following simple example:
if

n

=2

X = 0,

X = 2477, the middle four digit~ of X;, as the
1
pseudo-random number.
2

Next X = 06135529 and X = 1355 the second pseudo2
random1 number. Similarly X3 = 8360, X = 8896, etc.
4
.

For the results of s9~e tests on numbers
generate~7this way see Mauchly
and Votam and
Rafferty • Unsatisfactory results have been observed if the number has less than eight digits and
the sequence may develop unsatisfactory properties
if extended beyond 700 or so eight-digit numbers.
2. The Congruential methodse
a. The additive process. Starting with
an initial set of n random numbers, X ' X ••• ,X in
2
n
1
the range
0(X<1
the procedure generates additional numbers, X.,
successively according to the following rule: J
X.

J

= (X.J-1

+ X.

J-n

) Mod 1 j>n
/

An equivalent statement of the process that is con-

1,

r

=3

Equation (2.7) becomes

blocks which are suitable for his problem.
2.3 Methods of Generation of Pseudo-Random Sequences
Methods of generation of pseudo-random
bers by high speed computers are usually divided
into two classes.
a. Those for which there is, or appears to
be, no way of determining the cycle structure other
than the brute force procedure of investigating the
sequence by actual computation. The best known
method of this type is the middle-of-square method.
b. Those for which mathematical analysis can
determine the cycle structure, and even suggest
suitable parameters in the recursion relation to
give the longest period and ~ost satisfactory output such as the congruential methods, additive and
multiplicative.
15
1. The Middle of Square Method • Von
Neumann and Metropolis first suggested the middle
of squares method. This process canbe exe~plified
in a special case as follows. Take a 4 digit number X , e.g., Xo = 2061. Square it to obtain
04247921. Define

X
2

1

Xj

(X. 1 - X.

X3

(X

J-

J-n

)

Mod 2 3

(2.8)

Then

X

4

3
2 - X1 ) Mod 2
(X - X ) Mod 23
2
3

2

By the same procedure, the following sequence is
obtained
0, 1,1,2, 3, 5, 0, 5, 5, 2, 7, 1

Then the sequence will repeat itself.
Length of Period. As just mentioned,
the additive process is neriodic; it will eventually repeat itself by generating the original n
numbers. The period depends m~inly on nand r. In
Green, Smith and Klem's paper, they state that the
period T

(2.9)
where K is a constant depending on n and is given
in TablR 4
Table 4
n

K
n

n

2

3

7

127

3

7

8

63

4

15

9

73

14

11811

5

21

10

889

15

32767

6

63

11

2047

16

255

For n

= 15

K

n

n

12

K

n

3255

2905

and r = 35 the period is
T = 32767 X 2 34

which is approximately equal to 5 X 0014.
Four statistical tests 1 were made by
Green, Smith and Klem of the apparent randomness
of numbers generated by the additive process of
equation (2.7) with various nand r.

161
4.1

The additive process passes the frequency, poker, and serial correlation tests, but it
fails the runs test for n less than 16. It passes
the runs test for n = 16, and presu~bly for n
greater than 16. If alternate numbers are discarded, then the additive process passes the runs
test for n = 6, and presumably for all larger n.
One wishing to use the additive process to generate random numbers has three alternatives. He may decide that he needs only apuroximate randomness, as in setting up ex~erimental designs, and may therefore ignore the runs difficulty.
He will then choose n according to his convenience.
The runs test is very sensitive and failure to pass
this test does not mean that the numbers are badly
awry.
However, if one is working on a Monte
Carlo problem, or some similar problem with stringent randomness recuire~ents, then he must either
choose an n of at least 16, or must plan to discard alternate number.
b. The multiplicative method. Although
the additive process for generating random numbers
has been found very convenient for use in digital
computers, there exists an even simf6er process
which was f~ost suggested b~ Lehmer ~ and later by
Greenberger and Rotenberg~.
Xi +
1

=

(2

a

+

1) Xi + C, Mod 2P

(2.10)

advantages of this process over
the additive ~rocess are as follows:
(a) The process does not require an
initial set of randoo numbers. This means that no
memory storage is needed.
(b) Multiplication by a power of the
base can be accomplished by shifting, which is comparable in speed to addition.
(c) This process requires essentially
three additions and it can be done in one logical
step by a special ?urrose digital computer.
(d) This process also permits cons truction of a special serial computer with very few
components.
Several empirical tests were made by
Rothenberg of the apparent randomness of the numbers generated by the process of Equation (2.10)
with a = 7, C = 1 and p = 35. These tests are:
the frequency test, the runs up and down test and
the runs above and below the mean test. The results show that the numbers are uniformly distributed and that there is no serial correlation in the
sequence.
The serial correlation coefficient between two consec~tive numbers of this sequence is
shown by Coveyou ,
~he

C
1-6 -C( 1 --)

P

P

(2.11 )
where P is the modulus of Xi' i.e. P
for a
7 1 C
_1_ _

J=

27 + 1-

0.008

2P

1

By taking a = 9, this correlatiJn coefficient can be recuced to ?~~roxi~tely 0.002.
It can be show~- tnat the secuence of
Equation (2.10) can generate the full period of 2D
numbers if

a

~

2 and C is OnD.

See appendix I for the proof.
3.

Designs of a Special Purpose
Random Number Genera tor

As a result of the increased interest in the
use of Monte Carlo methods of co~putation in high
speed digital com~uters a number of subroutines
have been written.
For exa~ple, an IEM 709 program for the additive process
x. = ( x. 1 + x.
J

J-

J-n

) ,Mod 2 ~"
-' ~

requires 11 instructions and twenty-two 709 m9.chine
cycles, i.e.,264 microseconds to generate one number. See appendix II for code.
An IBM 709 program for the multiplicative
process
a
x. = (2 + 1) x. 1 + C, Mod 235 (;.2)
J

J-

requires 7 instructions and fourteen 709 machine
cycles, i.e.,168 microseconds to generate one
pseu~o-random number.
The actual code is listed
in appendix II.
A Monte Carlo problem which demands one million random nUill~ers is not considered to be impracticable. Notice that this does not mean that
the problem has undergone one million comDutational
trials. There are usually many random variables involved in one trial run. To generate one million
random numbers by the multiplicative method requires 168 seconds on the 709.
The random number generation time can be reduced, if a fast ~~chine is used. The 7090 would
take 33.6 microseconds to generate one number by
using the multiplic~tive process, which is five
times faster than the 709. However, the percentage of the generation time consumed in solving the
problem is not going to be changed. For example,
in the matrix inversion problem it is 17~ of the
total problem running time no matter what computer
is being used. And in some other problems, it
could be even more Significant than this 1~~.
In the paper "Organization of Computer Systen:
The Fired Plus VaFiable Structure Co~puterlf by
Estrin, it is pointed out that If • • • The fastest
single compo!;gnt switching speeds being discussed
are about 10 second. For any significant p:1rallel
information transfers ~ost r searchers observ~ a
3
loss of a factor of 10 or 10 bE~n~ing ~7siC parallel computer operations to 10
or 10 second •• "
One cannot expect to do very much better with the
populations of switching elements without ~ change
in the method of organizing them ••• The pri-

162
4.1
mary goal of the Fixed Plus Variable Structure Computer is:
"1. To permit computations which are beyond 111e
capabilities of present systems by providing an inventory of high speed substructures ana rules for
interconnecting them such that the entire system
may be temporarily distorted into a problemorjented special purpose computer •••••• "
Special purpose circuits can be designed to
generate random numbers in parallel with other
'l~tivities in the computer such that these numbers
are available on demand in the same sense as any
other operand stored for use in the computation.
The special purpose generator can be some fraction of the variable structure inventory in the(F
+ V} system to essentially remove the time for random number generation from the overall co~putntion
time.
In the following sections, the design of a
serial and a parallel random number generator will
be discussed. The principal criterion for the serial generator is the use of a small number of components while maintaining a reasonable speed; for
the parallel generator, the achievement of high
speed with a reasonable number of components.
The serial generator would increase the performance in solving many Monte Carlo problems. It
is possible to generate one random number every
eight microseconds continuously.
The pArallel generator may achieve an access
time of less than 0.14 microseconds. Although the
par~llel generator does not appear to have immediate usefulness it may be compatible with future
super high speed computers and particular problems
formulated for s01ution by Monte Carlo techniques.
4.

Parallel and Serial Designs of a Special
Purpose Random Number Generator Utilizing A Multiplicative Congruence Method

4.1 FunctionAl Description
This pseudo-random number generator is essentially a special purpose co':router, which will produce long sequences of unifornly distributed random numbers. The algorithm used is based on the
Multiplicative Congruenti~l method,
x.

J

=

(2

a

+ 1) x. 1 + C, Mod 2P
J-

As indicated in section 3 ,the choice of the
parameters a, c, and p is governed by the period
desired ann the minimum acceptable value of the
serial correlaticn coefficient.
In this example the constants are chosen in
such for'71 a.s tv illustrate the computer design
anci may be modified accordn~ to the above cri teria •
The cons"'"nts of the above equation are chosen

as follows:*
a

11

C

20

p

20
The length of the period is equal to 2 which
is approximately one million. For a = 11, the serial ~~,relation coefficient is apnroximatelyequal
to 2
•
The generator can be set to any initial starting number. Since the process does not require
storage of the previous number, the generator will
go on to generate the next random number as soon as
the pr0viouR one has been transferred.
1. The Serial Gener~tor. The Functional
Block Diaeram is shown in Figure 1. The complete
functional flow includes two cycles:
11
Cycle 1: (2
+ 1) x. is formed
1

Cycle 2:

X. 1 =(2
J+

11

+ 1)

X.

1

+ 1 is cO Tl1rleted.

The R register is a 20-bit register in which
Rj initially contGtir:s the least Significant bit and
which s torE's the ranc am numller, X.. In every logical sequential step, the 20-bit fegiRter is shifted right one place. The output of the adder is
shifted into the most Significant bi t of the R register, R , and the least Significant bit of the R
20
register IS shifted into the I register.
The T register c~ntrols the cycles during
the flow. T = 0 corresponds to Cycle 1. T = 1
corresponds to Cycle 2.
The N register counts from 0 to 19 and provides thereauired sequence of 20 control steps.
The S register controls the start signal.
It also provides appropriate control signals to all
other registers in the generf'to.r. The S rep;ister
is set by an initial "Demand" signal from a supervisory control observing the needs of the rest of
the (F + V) system and is reset at the end of Cycle
2.
a. Input eo.us.tions. The input eouations
may be written by collecting together and combining the timing and truth tnble logic as follows:
S

(Deoand)S'· + (N1 + N2 + N3 + N4 + NS) S + T

N1

S [N1 t

:t-.2

S [N1N2' + N1'N2]

N3

S [N1N2 N3'N5' +

N4

S [N1N2N3N4'· + (1\1' + N2'· + N3') 1\4]

N5

S [N1N2N3N4N5' + (N1' + N2') NS]

T

J
(N1' + N2')

N3

J

T'N5N2N1 + T (N5' + N2' + N1')

* If C ~ 1, then C must be added into the proper
position of a parallel adder or at an appropriate
time in a serial adder rather than the Simple
forced carry into the least significant position.

163
4.1
R
n
J

n

R
(n= 2,3" ••••••• , 19)
n+1
11 12'C _ ' + 11' 12 C _ ' + 11' 12'C n _
1
n 1
n 1

+ 11 12 C _

n 1

where I

n

is the adder output
I1

= R1

I2 + R10 TI(N5 + N4 N3 + N4 N2N1)
C

n

I1 12 + C _ (11 + 12) + Tt N5 N2N1
n 1

b. Estimation of running time for the
full cycles. If five megacycle flip-flops are used,
the delay ~hrough each flip-flop is 200 millimicroseconds. It takes forty-clock times to complete a
cycle. The total delay, therefore, is eight microsecond..,.
2. The Parallel Generator. The functional
block diagram is shown in Figure 2.
The R register stores the random numberx.
which is 20 ~tts long. The two inputs to the addef
are x. and 2 x'i1 The constant C = 1, is added to
the slim x. and 2 x. by setting the initial carry.
As soon a~ the Start signal is received, the adder
outputs will transfer
1 into R register to reJt1
place x .• Since x., 2 i. and C are all available
at the ~ame time, the pro~ess cl1.n be accocnplished
in one lo,o;ical step. For cp..ses where C 1= 1 more
co~plete stages of the adder may be required to
permit injection of C. The access time of this
special computer will mainl;" depend upon the speed
of the adder.
A parallel adder2~u2~ be able to generate
parallel carry functions.-'
Since carry C is
k
an explicit function of C
,the parallel carry
functions can only be obt~i~ed by a method of substitution. In applying this method, one will soon
find out that the carry fl.;nctions contain a great
number of terms. This may make it electronically
impossible to mechanize and in any event the response of the carry functions becomes so slow that
it actually loses the effectiveness of a parallel
adder.
If an adder2~f the type proposed by
Weinberger and Smith
is used, the logical configuration of Figure 3 is obta~ned.
By studying the adder equ"l.tions, one finds
that the maximum fan-in is four. The maximum fanout is five. This results in a maximum of six inverter delay times and one flip-flop delay time.
If one uses as a basic building block, a current
mode, diode gate unsaturated inverter with a gainbandwidth product of approximately 400 megacycles,
then one may use inverters hElving a response time
of less than 15 millimicroseconds per stage. The
total delay through the adder is less than 90m~s.
If 20 megacycle flip-flops are used, the delay
through each fJ ip-flop is 50 mrs., Uncler sucr. conditions it is implied that the random number generator could generate a random number every 140 m~s.

4.2

solved, it is always possible to design a special
purnose computer which can perform the operation
faster and more efficiently than a general purpose
computer can. It is clear, however, that the
speed and simplicity of the special purpose computer has been achieved at a complete sacrifice of
flexibility. The lack of flexibility, in general,
means that the control unit for a special purpose
computer can be simplified since the sequence of
operations performed is fixed. The data to be
operated on can be arranged so that it becomes
available as required without the necessity for
addresses or even a memory cycle.

A general purpose computer is usually more
expensive than anyone special purpose computer.
However, aside from the fields of automatic control
and business data processing, one will hardly find
the extended use of a special purpose computer. A
general purpose cooputer is mostly used as anarithmetic tool, which can be programmed to solve thousands of functions separately and continuously. It
is p~ssible but more likely impractical to have a
great number of fixed special purpose computers
working Simultaneously together. Either it is a
tremendous waste in the dollar point of view or the
overall perfornance will probably be disappointing
due to the difficulty of data synchroni~ation among
machines.
In order to take advantage of the better technology of both types of computers; that is, the
speed and siTple organization of the special purpose computer and the flexibility of the general
purpose computer and still be within the realism of
electrcnic circuitry knowhow, the proposed "Fixed
Plus Variable Structure Comnuter" is advantageous.
The variable structure inventory in the (F +"V)
system \.:ould be expanda'hle ana could grow wi th expandable a.nd could p:rO'lt1 with experiences and needs.
The procedures which lead to the design of special
purpose modules from the vari8ble structure inventory can be set up in standard forms such'as the
SHARE program for a general purpose com~uter system.
The special purpose random number generator
will be considered as one of the specia.l modules
in the (F + V) system. It should be noticed that
the particular serial random number generator discussed in the previous section consists of only
five conventional switching elements. They are:
1•

A shift register

2.

A three input serial full adder

3.

A counter

4.

Three flip-flops

5.

Three "and" gates and three "or" gates

The pRrallel generator consists of the following
two parts:

(F + V) Structure

1.

A register

Given any sets of functions which are to be

2.

A high speed parallel adder

164
4.1
It is also suggested in the (F + V) system
that if the random number generator is used frequently enough, it should become part of the instructionset of a future general purpose computer or
even possibly become one of the mOdules in the V
inventory. ~ adding this special purpose random
number generator to any presently exisitng general
purpose computer, for solving Monte Carlo problems,
it becomes a minimum (F + V) system at the expense
of a fraction of the original cost but achieves a
significant gain in speed.
5.
5.1

Examples of Applications

Introduction

In this section, three problems using Monte
Carlo technique, i.e., gamma ray diffusion problem,
matrix inversion problem and design of electronic
curcuits, are considered. The role which is played
by sequences of pseudo-random numbers will be emphasized as well as the possibility of simultaneous
generations of other elementary functions such as
, logarithmic and exponential functions in the (F+V)
structure computer.
5.2

. 24 25
Gamma ray diffuslon '

In the Monte Carlo approach to the Gamma ray
diffusion problem, a beam of monoenergetic Gamma
rays is incident at a given angle on a plane parallel barrier of finite thickness in one dimension
and of infinite length in the other two dimensions.
The paths of the Gamma rays are simulated by
appropriate random walks. A set of three random
numbers is generated per random walk between successive events such as collision or absorption in
order to obtain energy and angular distribution of
transmitted and reflected Gamma rays.
In this problem it turns out that simUltaneous
computations of elementary functions In x and eX
can reduce the total computation time by 40~. The
generation of pseudo-random ~~m~5r36r3~u~8e397.4%
of the total computing time. ' , , , ,
5.3

An important consideration in the design of a
transistor resistor logic system is the delay in
propagating signals through various levels of these
circuits. The propagation delay is an implicit
non-linear function of many variables. In investigation of its statistical properties, two immediate difficulties are encountered. First, the distributions of the variables are not completely
known. While it is possible to make reasonable assumptions regarding resistor values, such as a uniform denSity function, the transistor parameters
present a different picture. Their denSity functions are not generally known in a form describable by any simple function such as the normal distribution. One approach to this problem is to derive approximations to these functions from empirical measurement. Even with the distribution of
all the variables known, the distribution of the
resultant delay cannot be readily determined and
the techniques of the Monte Carlo method become
very powerful aids in the solution of such complex
problems. This is particularly true since the development of models for the transient and steady
state analysis of transistor SW~9ching circuits
such as ~ose of Ebers and Moll
and Sparks and
l3eaufoy.
The flow chart below shows a typical computational flow for a conventional sequential machine
in applying the Monte Carlo technique to evaluate
the propagation delay.

Random Selection of
Power Supply voltages,
Transistor Parameters
andValues of Impedance
Steady State
Circuit Current
Computation

Matrix Inversion

Matrix inversion by the Monte Carlo method was
first suggested by Von Neumann an~6Ulam and later
described by Forsythe and Liebler • The method
provides a simple computational approach to the
statistical estimation of the elements of the inverse of a given matrix. Here, random walks are
again introduced in generating the matrix inverse
and the random walks are terminated when the statistical variation of the result is less than a
given tolerance. The percentage of time spent in
generating random numbe 2g ~?r2Shis type of program
is approximately 17.3~. ' ,
5.4 Electronic Circuit Design31 ,33,34,35
The Monte Carlo technique is finding wider
application in designing electronic circuits and is
illustrated by the problem of estimating propagation
delay time of transistor-resistor logic circuits. 31

repeat

The method of approach is: establish a mathematical model of the circuit; generate a set of
numerical values for all these variables according
to their respective distributions;* evaluate the
function using this set of values; and repeat the
above operations using another independent set of
data.

* One can obtain these values by actually measuring the parameters of a pa.rticula.r transistor randomly selected. Resistor values may be assumed to
have a uniform distribution.

165
4.1
Calculation of Rise Time, To
The time required for the collector current to
reach 90% of its final value is
1b1
To =

La

0.91 c

I b1--_

Pn

where four random numbers are generated to select
't
a

minority carrier lifetime in the active
region

1b1

step turn on current

I

collector current
normal common emitter current gain

R

the effective impedance looking out from the
base terminal
input capacitance
forward

bias voltage on base

reverse bias voltage on base
In estimating R, it may be more desirable to randomly select individual resistor values in the
circuit and then compute R, rather than picking R
randomly from some estimated distribution for R.
Turn on

c

(3n

Calculation of Storage Time -T
1
The time required to switch the transistor
from the saturation region to the edge of the active region is

+

In expression (5.2), T is the response of the transistor to a step gf base current.
The second term compensates the delay due to the
input capacitance.
Storage Time

where two new random numbers are needed to obtain
minority carrier lifetime in the saturation
region
1b2

step turn off current

Calculation of Decay Time T2
The time required for the collector current to
decay to zero 1s

Calculation of the Input Delay time TOD
The time required to charge or discharge the
input capacitance is

e

L
Re.

l.n -1

L
RC

in

where at least four additional random numbers are
used to select

T = T +L ln
s
s
1

e

T
on
1:'s
T
on
1:' s

-1

(5.6)

In expression (5.6) T is the time required
to remove excess carriers from the base with a
step of base current applied. The second term in
expression (5.6) is due to the fact that the base
of the transistor is discharged by a signal with a
finite rise time, Ton.

A simulation progr~r has been prepared by Y.
C. Ho and W. J. Dunnett. The entire program has
approximately 4,000 instructions and takes an average of three seconds per run on the IBM 709
Computer. In order to verify and establish the
accuracy of the mathematical model a statistical
experiment was also performed. In this experiment
measured data was compared with predicted statistical data provided by the computer program and in
each case, the predicted mean was quite close to
the measured mean. Moreover, the computed deviation in every case was more conservative than its
measured value.
One notices that in computing the propagation
delay, logarithmic anrl exponential functions are
used four times and three times respectively, in
addition to the required generation of a set of ten
random numbers per transistor per run.
The special purpose random number generator
does away with most of the time required to generate random numbers which become available essen-

166

4.1
tially in one access time to the high speed memory.
If in the same sense as the random number generation, a special purpose configuration isxestablished in V for computation of ln x or e concurrently with the program being executed in F, then
one may expect to achieve significant reduction in
total computation time.

"Empirical Tests of an Additive Random Number
Generator", Journal of the Assoc. Compo
Machinery, pp. 527-537, September, 1959.
9.

Rotenberg~

Consideration of the set of equations 5.1-5.6
shows that the seven natural logrithm and exponential functions must be evaluated in sequence.
Thus while F is calculating the operands to be
used in these functions, V might be generating the
random numbers and functions required in the following steps.

10.

Kendall, M. G. and Smith, B. B., "Randomness
and Random Sampling Numbers", Journal of the
Royal Statistical Society, Vol. 101, pp. 147166, 1938.

11.

Cramer, H., .Mathematical Methods of Statistics,
Princeton University Press, 1946.

A block diagram of the (F + V) structure computer illustrating a form for achieving the above
is shown in Figure 4.

12.

Rand Corp., A Million Random Digits With
100.000 Normal DeViates., Free Press, 1955.

13.

Dixon, W. J. and Massey, F. J., Introduction
to Statistical Analysis, McGraw-Hill, 1951.

14.

Hald, A., Statistical Theory with Engineering
Applications, Wiley, New York, 1952.

15.

Taussky, O. and Todd, J., "Generation of
Pseudo Random Numbers", Symposium on Monte
Carlo Methods, Wiley, 1956.

16.

Mauchly, J. W., Pseudo-Random Numbers, Presented to American Statistical Assoc.,
December 29, 1949.

17.

Votan, D. F. and Rafferty, J. A., "High Speed
Sampling", !!!!2. 5 (1951) pp. 1-8.

18.

Garner, Harvey L., "The Residue Number System",
Proceedings of Electronic Computers 8 June,
1959.

19.

Lehmer, D. H., "Mathematical Methods in Large
Scale Computing Units," Proc. of a 2nd Symp.
on Large-Scale Digital Calculating Machinery,
pp. 141-146, 1949.
Greenberger, M., Decision Unit Models and
Simulation of the United States Econo
Chapter III, Preliminary Draft 1958 •

6.

Conclusion

In all three examples considered in Section
5, serial generation of random numbers appeared
quite satisfactory, since, using inexpensive
transistors it is possible to generate one random
number in about 8 }Ns. No case has been found in
which it is necessary to employ the parallel method
of random number generation.
If it becomes necessary to generate random
numbers more rapidly than 8fs per random number,
one can either decide to employ more expensive,
faster components in a serial random number generator or utilize the fully parallel configuration.
References
1.

2.

3.

4.

5.

6.

7.

Estrin, G, I1Organization of Computer Systems The Fixed Plus Variable Structure Computer,"
Proceedings of Western Joint Computer Confer~, May, 1960.
Donsker, M. D. and !{ac, M., I1The Monte Carlo
Method and Its Applications", Proceedings of
Computation Seminar, IBM Corp., December, 1949.

20.

McCracken, D. D., "The Monte Carlo Method",
Scientific American, Vol. 192, Page 90, May,
1955.

21.

Brown, G. W. "Monte Carlo Methods", Modern
Mathematics for the Engineer, Edited by E.
Beckenbach, McGraw-Hill Co., New York, 1956.

Coveyou, R. R., "Serial Correlation in the
Generation of Pseudo-Random Numbers," J. Assoc.
Comp., pp. 72-74, January, 1960.

22.

Weinberger, A. and Smith, J., "A One-Microsecond Adder Using One Microsecond Circuitry,"
Proceedings of Electronic Computers, IRE
EC-5, June, 1956.

23.

Richards, R. K., Arithmatic Operations in
Digital Computers, D. Van Nostrand, 1955.

24.

Berger, M. J., "An Application of the Monte
Carlo Method to a Problem in Gamma Ray
Diffusion," Symposium on Monte Carlo Methods,
Wiley, 1956.

25.

Beach, L. A. and Theus, R. B., "Stochastic

Lytle, E. J., itA Description of the Generation
and Testing of a Set of Random Normal Deviates",
Symposium on Monte Carlo Methods, Wiley, 1956.
Butler, J. W., "Machine Sampling from Given
Probability Distributions", Symposium on Monte
Carlo Methods, Wiley, 1956.
IBM 709 Data Processing System Reference
1958.

~,

8.

A., "A New Pseudo-Random Number
Generator , Journal of the Assoc. for Compo
Machinery, pp. 75-77,January, 1960.

Green, B. F., Smith J. E. K., and Klem, L.,

167
4.1
Appendix I

Calculations of Gamma Ray Diffusion,
Symposium on Monte Carlo Methods, Wi1ey,1956.
26.

27.

28.

29.

30.
31.

32.

33.

34.

35.

Forsythe, G. E. and Leib1er, R. A., "Matrix
Inversion by a Monte Carlo Method," MTAC IV
31: pp. 127-129, 1950.
-Tang, T. "A Special Purpose Random Number
Generator" Master Thesis, U.C.L.A., January,
1960.

Period of Multiplicative Congruence Method
From the equation
Xi +1 = (2

Beaufoy, R. and Sparks, J. J. "The Junction
Transistor as a Charge Controlled Device,"
A.T.E.Journa1, October, 1957.
Ho, Y. C. and W. J. Dunnett, "Monte Carlo
Analysis of Transistor Resistor Logic Circuit, It
IRE Convention Record, March, 1960.
Edmonds, A. R., "The Generation of PseudoRandom Numbers On Electronic Digital
Computers," The Computer Journal, pp. 181-185,
January, 1960.
Hellerman, L. and Racite, M. P" "Reliability
Techniques for Electronic Circuit Design,"
Transaction on Reliability and Quality Control,
IRE September, 1958.
Parker, J. B., "The Accumulation of Chance
Effects and the Gausslal) freq.uency Distribution,"
Phil Mag., 38:681-682 (1947)
Silberstein, L. "The Accumulation of Chance
Effects and the Gaussian Frequency Distribution," Phil Mag. , 35: 395-404 (1944).

36.

IBM Share Program F114 Distribution No. 027,
Western Data Processing Center, U.C.L.A.

37.

IBM Share Program LAQ1 Distribution No. 525,
Western Data Processing Center, U.C.L.A.

38.

IBM Share Program AS09 Distribution No. 224,
Western Data Processing Center, U.C.L.A.

39.

IBM Share Program AS03 Distribution No. 224,
Western Data Processing Center, U.C.L.A.

P

+ 1) Xi + C, Mod 2

and
a

Ralston, A. and Wi1f, H. S. Mathematical
Method for Digital Computers, John Wiley and
Sons, Inc., New York, 1960.
Ebers, J. J. and Moll, J. L. "Large-Signal
Behavior of Junction Transistors," Proceedings
of IRE, December, 1954.

a

(2 + 1)

x.1+1

a
(2 + 1)

L (2a +1)

+ C
Xi +

c] + C

a
a
(2 + 1)2 Xi + (2 + 1) C + C
and
a
a
Xi+n = (2 + 1)n Xi -+ [(2 + 1 )n-1+

(2a + 1) n-2 + •

] C

(2a+1)n X. + C (2 a + 1)n _ 1
1

The sequence repeats itself when

Xi +n

= Xi

From (2) and (3) we have
a

(2 +1 )n-1

a2

[

a
]
2 Xi + C = 0

Mod 2P

If C is odd and a> 2 the sum of the terms in
the bracket is odd. Then the first term must be
divisible by 2P in order to satisfy (3). Thus it
is required to find the smallest n satisfying the
congruence
(2 a + 1)n_1

a
2

In Rotenberg's paper 9 , he points out that
from the Number Theory, ~ (2P ) is a solution of
(4) and the smallest n must be of the form 2r.

]
]
If a = 2, all terms in the bracket after the
first "l tt will be even and thus the Whole bracket
is odd. Then to satisfy (4) the factor 2r must
be divisible by 2P • Thus the minimum value for r
1s P, or n = 2P •

168

4.1
Appendix II
Examples of Computer Routines to Generate
Random Numbers
The IBM 709 MIdi tive Congruence Method 14
Location

Operation

Address Tag

00

SXD

Rand 1, 1

01

LXD

Rand 2, 1

02

CLA

Rand 3, 1

04

TIX

05, 1, 0

05

ADD

Rand 3,

Comments
The contents
of index reg.
1 is stored
in Rand 1
No. of initial
numbers is
loaded into
index reg. 1
x. 1 is added
ta-acc.
Transfer on
index (10)
x.
is added
t

06

STO

Rand 3,

07

SXD

Rand 2,

08

LXD

09

TOV

10

TRA

a-nx j _ 1

x. is stored
J

The content of
index reg. 1
is stored in
Rand 2
Rand 2, 1
Index reg. 1
is restored
10
overflow condition reset
Normal Return

Rand 1 Temporary Storage
Rand 2 No. of initial numbers
Rand 3 Address of the last initial random numbEr
(Random numbers are stored in consecutive
memory locations)
The two SXD and LXD (00, 01, 07, 08) instructions could be omitted if the index register 1 were
available for use by the subroutine, thus saving
four in~tructions per generated number.
The IBM 709 Multiplicative Congruence Method
Location

Address

Comments

CLA

Rand

x. 1 is added to
aac.
C is added to x _
j 1
x. 1+ C is stored

Operation

0

ADD

Rand 2

2

STO

Rand

3

ALS

Rand 3

4

ADD

Rand

2 x _ is. formed
j 1
Xj is formed

5

STO

Rand

x. is stored

6

TOV

Normal Return

Rand

Random number x. 1

Rand 2 C
Rand 3 a

J-

~-

J

CARRY

s

T

~~
~
N'19

I

N

ADD COMPLETE

(T=O)(N=19)

ADDER

REG~

0-19

s =I

Figure 1

-ADO COMPLETE

FUNCTIONAL BIDCK DIAGRAM FOR A SERIAL PSEUOO-RANroM NUMBER GENERA.'roR

~I-'

•

Ol

I-' (0

170
4.1

COMPLETION
TO S.C.
START
FROM S.C.

R REGISTER

I--~

RANDOM NUMBER TO
DATA REGISTER

PARALLEL ADDER

Figure 2

FUNCTIONAL BlOCK DIAGRAM FOR A PARALLEL PSElJ.OO-RANJX)M NUMBER GENERA.IDR

171
4.1

X7

I------------------~

Xa ~----------------~

Xg

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

X,o

~----------------__1

TIME

FiGClre:3

SYNEOLIC

I

Fum

TIME 2

TIME 3

TIME 4

CR4..."tT FOR PARALLEL ADDER INDICATING TINllm DURING ADDITION

t/:o.~

RANDOM NUMBER
GENERATOR

••

s.c.

"DEMANDII FROM

s.c.

..-. COMPLETION TO

"TRANSFERII FROM

JL

a:
w

s. C.

---.. RESULT TO F

l-

(/)

(!)

w
a:

~ 1-+ ARGUMENT
o

In x

OR

eX

GENERATOR

...

DEMAND FROM

~

COMPLETION

s.c.

TO

Figure 4 (F + v) STRUCTURE CONFIGURATION

s.C.

FROM F

•

....;j

~

to.;)

173
4.2
THE CELLS CAN

SYSTEM T.M.

A LEUCOCYTE PATTERN ANALYZER
K. Preston, Jr.
The Perkin-Elmer Corporation
Norwalk, Connecticut
Summary
Medical workers suspect that the incidence
of the binucleate lymphocyte in the peripheral
blood stream is an index of radiation damage. l
The fact that the incidence of this type of white
cell is of the order of one per each million
blood cells makes practical use of this index by
the human observer essentially impossible. The
Atomic Energy Commission has therefore arranged
for the construction of a blood cell scanning
system with which to determine the feasibility of
the semi-aatematic identification of the binucleate lymphocyte in glass-mounted smears of human
blood.
The CELLSCAN system consists of a closed circuit television microscope coupled to a special
purpose digital computer. The system produces a
quantized image of the leucocyte whose pattern is
to be analyzed. The computer program causes the
quantized image to be operated upon serially in
such a way that groupings of contiguous binary
"l's" are reduced to single l's. Sampling the
number of isolated l's at various instants during
the reduction process produces a histogram of the
leucocyte's various constituent parts. The
CELLS CAN system is intended to determine whether
the-resultant histogram of the binucleate lymphocyte is unique in the histogram population of
all leucocytes.
Introduction
The use of radioactive materials or of
radiation producing equipment continues to increase throughout the world. Users are confronted with the problem of measuring the effect of
radiation on workers in radiation environments.
To date no good index has been found which
measures the physiological changes produced in
humans by low level radiation. One hypothesis
that has been advanced is that the appearance of
a rare type of white blood cell in the blood
stream may provide such an-index. The rate of
occurrence of this white cell is extremely low.
I~ order to obtain sufficient statistical data to
prove the above hypotheses, it is essential that
automatic equipment be developed which is capable
of scanning and anaiyzing blood samples.
It is the purpose of this paper to describe
an experimental blood cell scanning system which
has been constructed for the Atomic Energy Commission. The equipment which has been assembled
is designed to scan individual white cells rather
than blood samples containing a multitude of
cells. Experimental results will be used to es-

tablish the feasibility of using electro-optical
scanning with data processing pattern recognizing techniques to identify certain types of
human blood cells.
Hematological Background
Figure 1 is a table showing some of the
classes of cells which are found in the human
blood stream, i.e. the peripheral blood. Blood
cells are divided into two major classes: The
red cells (erythrocytes) and the white cells
(leucocytes). It is the white cell population
with which we are concerned here. The two major
classes of white cells are lymphocytes and
granulocytes. There are about two granulocytes
for each lymphocyte. The sub-class of lymphocytes whose incidence may be an index of low
level radiation damage consists of lymphocytes
having two nuclei. These cells are called
binucleate lymphocytes. For brevity we shall
refer to them as "bilobes".
Some quantitative information on the incidence of bilobes has been obtained in experiments with animals. Figure 2 shows a graph prepared by Dr. Marylou Ingram at the University of
Rochester. 1 It shows the number of bilobes per
thousand lymphocytes in the peripheral blood of
dogs who were exposed to low level radiation
from the University of Rochester cyclotron. As
can be seen the normal incidence of bilobes in
the dogs tested is about one per 30,000 lymphocytes. There is an immediate increase in the
number of bilobes upon exposure by about an order
of magnitude. This increased incidence continues
in the peripheral blood for approximately one
month during which time its amplitude gradually
diminishes.
The normal incidence of bilo~es in man is
somewhat greater than that given in Figure 2 for
dogs. It is of the order of one per ten-thousand
white cells which is equivalent to about one for
each one million total blood cells. In order to
measure the index of bilobes for a particular
human being, the technician employed to process
a blood sample must count and catalogue cells
for several hours. This makes routine measurements of this index both practically and economically unattractive as long as human technicians
are required to perform the measurements. These
measurements are also subject to error to
technician fatigue. For this reason a study program has been instituted in order to determine
whether it is within the present state-of-the-art

174

4.2
to evolve a blood cell scanning and data processing system which will identify bilobes in typical
samples of peripheral human blood with reasonable
error rates. This system is called the CELLS CAN
system.
Figure 3 shows a portion of a typical blood
sample. The sample is prepared by squeezing a
drop of human blood between two glass microscope
covers lips. When the slips are separated the
surface of each is coated with a monocellular
layer of blood. The layer of blood is then
treated with certain biological dyes which selectively stain the different cells and cell constituents. The chemical reaction between the dye
and the cells imparts characteristic colors to
the cells which assist the hematologist in cell
identification. Figure 3, however, is purposely
black and white as the CELLSCAN system uses a
monochromatic input.
Different types of white cells appear in
Figure 3 interspersed with red cells. The red
cells are the regular disc-shaped objects. The
white cells are the irregularly shaped darker
objects. Each white cell consists of nuclei within the cell body or "cytoplasm". A bilobe is
shown as well as two granulocytes. Granulocytes
are characterized by the multitude of dark grains
or "granules" within their cytoplasm. It should
be noted that one granulocyte has two nuclei.
Therefore, in order to identify the bilobe, the
CELLSCAN system must differentiate between
granulocytes and lymphocytes as well as between
binucleate lymphocytes and lymphocytes with other
than two nuclei.
The technique used by the CELLSCAN system to
differentiate between classes of white cells is
to count and size the constituents of a given
cell. By "constituents" we mean the nuclei and
the granules. The output of the system is thus
a histogram from which cell type may be determined.
A digital approach to this problem of particle
counting and sizing has been chosen as the best
method for dealing with irregularly shaped
particles.
The "Shrink" Algorithm
The equipment which has been designed to
identify bilobes employs a television microscanner to deliver a binary video image to a
special purpose computer. The scanner is so designed that, whenever white cell nuclei or granules appear in the field of view, the digital
output is "1". All other areas of the field of
view are "0". A hypothetical binary image is
shown in Figure 4a. Since the computer is
digital, this image is quantized into a Cartesian
array of elements or "bits". Dark elements in
Figure 4 correspond to binary "l's".
The function performed by the special purpose computer is to count and size the image
areas which are comprised of contiguous "l'su.
The output of the computer is an image histogram

for the particular cell which is viewed by the
scanner. In order to produce this histogram the
computer first stores the entire binary image.
It then operates sequentially on each image bit
causing binary "l's" on the periphery of any
contiguous group of "l's" to be changed to "O's".
This operation has been called the "shrink" operation and was originally suggested by M. Golay.2
In general, the computer is required to make
several sequential examinations of the image before the image histogram is completed. Each
examination will be called a "pass". Figure 4b
indicates how the shrink operation modifies the
original image shown in Figure 4a after two
complete passes. Figure 4c indicates the re-·
sidual image after sufficient passes have been
made to reduce the original image to a single
isolated binary "I". The number of passes required to reduce the original image to an isolated
"1" is proportional to the maximum chord of the
original image. Thus the image histogram is computed by counting the number of isolated ones
present in memory after each pass. This information is then converted into a plot of the
total number of groups of "l's" in the original
image which fall into each of several maximum
chord ranges.
To program the computer to perform the shrink
operation an algorithm has been devised by L.
Scott and R. M. Landsman. 2 The algebraic expressions which define the shrink algorithm are
shown in Figure 5. The image bit operated upon
is designated as X and its neighbors as A, B, C,
••• ,H. Since the computer sequences from left
to right and top to bottom, bits A, B, C, and H
have previously been processed. Their processed values are designated Ap ' Bp ' Cp ' and Hp.
Three functions are derived from the values
of these 8 neighbors. The function f(ISO) indicates whether the bit under examination is an
isolated one while the function f(TAZ) indicates
whether the neighbors contain three adjacent
"O's". A further function is required whose
value indicates whether the X bit is a link between subgroups of a given grouping of contiguous "l's". This indication is provided by the
value of f(TUP) which is "1" when there are three
or more unlike neighbor pairs. This is necessary
so as to prevent the computer program from producing two isolated "l's" when operating upon
a dumbbell shaped original image.
The complete shrink algorithm is f(X ) and
is given in the last line of Figure 5. ~en this
algorithm is applied to a stored image, results
as previously described in Figure 4 are obtained.
It causes isolated "l's" to be retained rather
than being converted to "0". This permits periodic sampling of the number of isolated "l's" in
the computer memory. This periodic sampling takes
place at intervals determined by the scale required for the image histogram.

175
4.2
Technical Details
The first model of the CELLS CAN System has
been built in order to evaluate the feasibility
of recognizing bi10bes using the data processing
technique described above. A block diagram is
shown in Figure 6. The first model is not intended to be a machine capable of scanning complete blood samples. The micro-scanner is manually centered on the blood cell to be analyzed.
For reasons of economy the scanning and data processing rates are slow. About 10 minutes are
required to process one image. The periodic computer sampling of isolated "l's" is converted to
an image histogram by manual calculation.
The Scanner
The scanner standard employs a Leitz Ortholux microscope coupled to a Dage Data-Vision
scanner which has been modified so as to enhance
its signal to noise ratio. The output of the
video amplifier is digitized by means of a video
quantizing circuit.
The Data-Vision equipment scans at a rate of
60 horizontal scans per second which is about 300
times slower than standard closed circuit television systems. The period of the vertical scan
is 5 seconds. This produces a field of 300 horizontal scans for each image. This slow scanning
rate was chosen so that the output data rate of
the scanner could be matched to the data processing rate of the computer. It permits image
data recording directly upon magnetic tape at
2000 bits per second. Thus a relatively inexpensive audio tape machine can be employed.
White cells are typically 10 to 20 microns
in diameter. In order to restrict the field of
view to a single cell an area of blood sample 30
microns square is imaged by the micro-scanner.
Thus, with 300 scanning lines, the elementary
sample area or scanner "resolution" is one-tenth
of a micron square. This resolution is required
since the separation between nuclei in a bi10be
may be of this order. It implies, however, that
computer storage of almost 100,000 bits per field
of view is required. In order to reduce this
memory requirement a processing technique developed by W. K. Taylor is used in the scanner. 3
Circuitry is used which stretches white-polarity
video signals and equally shrinks dark-polarity
video signals. This causes the separation between bi10be nuclei to appear greater. However,
it simultaneously makes granules appear smaller.
In the CELLSCAN system a five fold stretching is
permissible. Although granules having maximum
horizontal chords less than one-half micron disappear from the image, sufficient granules remain
to make cell identification possible. In this
fashion the horizontal resolution is decreased
to one-half a micron and the computer memory capacity requirement becomes about 20,000 bits.
The Computer
A block diagram of the CELLS CAN system com-

puter is shown in Figure 7. The computer memory
is a continuously moving loop of magnetic tape
which is capable of storing 19,200 bits corresponding to a 64 x 300 field of view. Non-returnto-zero recording is used with automatic erase
after reading. One track is used for storing
"O's"; another, for "l's". Internal computer
timing signals are obtained from the data stored
on the tape. At the beginning of the data processing run the computer memory is erased. Then
under control of synchronizing signals from the
scanner the memory is filled with one field of
view. The image stored on the magnetic tape is
serially transferred to memory registers and
operated upon using the shrink algorithm under
the command of controls on the computer console.
The computer applies the shrink operation to the
image for a pre-set number of passes variable
geometrically from 4 to 128. After this number
of passes has occurred, the computer automatically reverts to a mode of operation wherein the
tape continues to be read and the image is rewritten unchanged. At this time the computer may
be ordered to count the number of isolated "l's"
in the residual image. By alternating between the
command to shrink and the command to count isolated "1' s", the Ote rator is able to complete the
image histogram for the field of view which has
been stored.
Another requirement of the computer is to
complement the image stored. This routine is
required in order to eliminate noise in the
original image. Consider for example, the problem of image inclusions as in applying the shrink
operation to the letter "0". Due to f(TUP) which
indicates an X bit which is a link between two
neighbors, this configuration of binary "l's"
cannot be reduced to a single binary "1" by the
shrink operation. This indicates that any inclusions which occur in a group of contiguous
binary "l's" will prevent this group from being
reduced to an isolated" 1". In or der to eliminate inclusions in the image received from the
scanner the CELLSCAN computer first complements
this image. In the complemented image inclusions
appear as isolated "l's" or small groups of contiguous of "l's". These are eliminated by applying a modified shrink operation wherein f(ISO)
= 0 so that isolated "l's" are not retained. Now
the image is again complemented in order to recr~ate the original image in an inclusion free
form.
Part of the computer is a test pattern generator which may take the place of the scanner as
a data source. It is used during periods allocated to cOmputer maintenance and debugging. The
action ot the computer is monitored by a video
monitor which displays the original or residual
image contained in computer memory. Furthermore,
the action of the scanner is observed by a second
monitor which is capable of displaying either the
analogue video signal "or the quantized signal.

176

4.2
Results
An example of an analogue signal from the
scanner is shown in Figure 8, which shows a
typical bilobe.
Figure 8 shows the analogue
signal before pulse stretching. It was obtained
using a glass mounted blood sample prepared by
the University of Rochester using peroxidase stain
counterstained with Wright stain. An oil immersed
90 power apochromatic objective lens having a numerical aperture of 1.32 was used with a 1.40
numerical aperture oil immersed condenser. The
virtual image formed by the objective was imaged
on the photoconductive surface of a General
Electrodynamic Corp. 7325 vidicon tube by means
of a Leitz widefield periplant eyepiece. The
light source was a tungsten spiral filament filtered by a 5300 Angstrom narrow band filter.
The analogue signal was quantized with the
result shown in Figure 9. Inclusions in the
nuclei due to non-uniformities in transmissivity
can be seen. After elimination of inclusions by
complementing the image and processing in the
modified shrink mode this image would shrink,
when recomplemented, to two isolated ones after
about 50 to 100 passes.
Other cell types have been imaged and quantized using the CELLSCAN system. At this writing
statistical information is being gathered on cells
from many blood samples in order to ascertain
what error rates can be expected in scanning
bilobes in the presence of all other cell types.
General Considerations
In the above discussion of automatic blood
cell scanning, we have described a machine which
instruments a particular data processing routine.
This special purpose machine is being used to
ascertain whether the particular routine suggested is sufficiently general to identify binucleate lymphocytes. As in all discussions of
special purpose pattern identifiers, the question
arises as to how one is to determine the optimum
approach to the problem at hand. Few treatments
of the general theory of pattern recognition
exist. Some that are known to the author are
listed in the bibliography.4,5,6 One of the more
general treatises in this field is the one written by A. Gill. 6 He treats the general problem
of pattern recognition by assuming a pattern set
S, consisting of M subsets or individual patterns.
He assumes that each of the M sub-sets contains N
features. If we confine ourselves to a binary
system, then the minimum value of N is (10g2M),
where the parentheses indicate the largest integral value of log2M. Furthermore, the maximum
value of N is equal to (M-l). In the latter case
the M x N matrix characterizing the S is the
unitary matrix, i.e., a matrix whose diagonal
values are all binary "l'sll and whose other elements are "0".
The more efficient the method of defining
each Mi, the lower will be the value of N. Gill

defines the efficiency of a noiseless pattern
scanning system as equal to the information content of S divided by the value of N. Defining
"e" as efficiency we thus have:

~ Pi 10g2 Pi

<

i

e

S.

M-l

~

Pi 10g2 Pi
(log2M)

where Pi is probability of finding the Mi in S.
Let us consider S as the population of white
blood cells in peripheral human blood. The probability of occurrence of mononucleate lymphocytes,
binucleate lymphocytes, and granulocytes respectively, are shown below:

PML

3 x 10- 1

PBL

3 x 10- 5

PG

7 x 10- 1

From the above figures, we can compute the
information content of the source as equal to
0.70. Since we are dealing here with three patterns, Gill's approach would indicate both a maximum efficiency equal to 0.35. However, in the
special purpose system which has been designed
it should be noted that a value of N equal to
19,200 has been used. This indicates a system
efficiency of 3.7 x 10- 5 • This quantity is completely outside the range of efficiencies delineated by Gill. This would seem to indicate
that a highly inefficient approach has been
adopted in the present pattern analyzing system.
Contemplating this problem further, we should
again note that, using the Gill approach, we
should be able to obtain a value of N ,equal to 2.
This implies that the existence or presence of
only two features need be recognized by the
scanner. For example, these two features could
be:
A

The existence of two nuclei.

B

The existence of many granules.

The ideal scanner would recognize the bilobed
lymphocyte as being characterized by AB. All
other lymphocytes would be characterized by AB.
Granulocytes would be characterized by (AB + AB)
or, merely, B. The problem now is how to design
a scanner whose output can be directly translated
into the existence or non-existence of the two
features mentioned above. Gill's theory that such
a scanner should exist does not give the designer
a clue as to how such a scanner can be constructed.
It appears clear to the author that existing contributions to the theory of pattern recognition .
require further extension in order to define
methods whereby the salient features of a multiplicity of patterns may be determined. An alternative is to term the special purpose computer

177
4.2
which accompanies the scanner as part of the scanner, rather than part of the recognition logic.
This semantic manipulation, however, is of little
value in solving the practical engineering problem.
Future Development
Some of the practical problems which must be
solved in building a practical blood cell pattern
analyzer may now be listed. As has been mentioned,
the present system is a research tool for use in
proving the value of certain data gathering and
processing concepts. It operates at low speed
and would, in fact, require over a month of continuous operation to process a single blood
sample. A useful system is one which can process
a blood sample (containing approximately 10,000
white cells) in about 15 minutes time. This rate
of data processing allows about 100 milliseconds
to process each cell. Taking into account present limitations of video scanning systems, about
30% of this time should be allocated to scanning
and storing the cell pattern. This leaves approximately 70 milli-seconds during which to perform
the shrink operation on the pattern stored for a
sufficient number of passes to prepare an image
histogram.
Let us assume that we extend the present
serial mode machine directly and further assume
that about 100 passes would be required for each
cell identification. Noting that about 20,000
bits must be stored in each image, this implies
that 2,000,000 bit operations must be performed
in 70 milli-seconds, i.e., an allowance of 35
nano seconds per bit. In order to do logical
operations at this rate we must work at the frontier of the state-of-the-art, using, let us say,
a microwave memory as a data storage medium and
high speed logic circuitry to process the data.
Such a technique would require that the machine
retrieve one bit of data from the delay line
store, cause the operation of a logical sequence
incorporating about 10 propagation times, and
read out the results of this operation all in a
35 nano second period. Such a feat may well be
beyond present day capabilities. 7
Another approach would be to substitute a
combined parallel-serial mode machine for the
serial mode machine described above. For example, one might contemplate storing the image
in a 300 word sequential access memory, having
64 bits per word. Parallel logic would be provided which would examine 64 bits of the image
simultaneously. This would require a 64 fold increase in the number of logical gates which would
be used to instrument the shrink logic. Some reduction in this figure is possible by cross connecting the f(TUP) logics. The computer would
operate by storing three 64 bit words in its
memory register and simultaneously storing 64
previously processed bits in an auxiliary register.
A further 64 bit register would be required to
store the output of the 64 shrink logic circuits.
Because of the characteristics of the shrink

algorithm, which at the moment is implemented by
about 50 logic gates, it can be shown that a 4 to
5 fold increase in total machine complexity is
indicated. The memory would perform a read-write
cycle for every 64 bits of image processed. Again
assuming 100 passes per image, this would imply
30,000 read-write cycles in 70 milli-seconds which
is about 2 micro-seconds per read-write cycle.
These specifications for a parallel-serial mode
machine seem more reasonable than those corresponding to the serial mode machine. They do,
however, imply a large increase in components
required due to the adoption of parallel logic
techniques.
Conclusion
In automatizing the analysis of blood cell
patterns the cQmputer engineer must devise
machine techniques which can reproduce the human
visual recognition process. It appears characteristic of such machines that vast amounts of
input data must be operated upon in order to
deliver a fairly elementary output. This is in
contrast to present scientific and business computers where input and output data rates are commensurate. For example, in the present CELLSCAN
system an input of approximately one quarter of
a billion bits would be gathered in scanning a
blood sample. This input is reduced to the
quantity of bilobes in the sample which can be
represented by a five-bit word. Even if the present system efficiency is improved to the maximum
predicted by Gill, the ratio of input to output
data ra es would be of the order of ten-thousand
to one.
Therefore, it is found that today's
applications of data processing technology in the
field of pattern recognition require us to take
full advantage of present circuit art. It is
here that use can be made of microwave logic
techniques. Furthermore, a better theoretical
grasp of pattern recognition problems is required
so as to guide the engineer towards the most efficient use of available logic circuitry and data
stores.

6

Acknowledgements
The author would like to acknowledge the
support provided the CELLSCAN project by Dr.
Marylou Ingram, of the University of Rochester,
who has been the prime mover behind the medical
investigation of the incidence and significance
of binucleate lymphocytes. The project itself
has been undertaken on a U. S. Atomic Energy
Commission contract.
References
1.

Ingram, Dr. Marylou, "The Occurrence and
Significance of Binucleate Lymphocytes in
Peripheral Blood After Small Radiation Exposures", Internationa1.Journa1 of Radiation
Biology, Special Supplement, Immediate and
Low Level Effects of Ionizing Radiations
(1959).

178

4.2
2.

United States patent application number
854254, filed Oct. 8, 1959 by the PerkinElmer Corp.

3.

Taylor, W. K., "An Automatic System for Obtaining Particle Size Distributions with the
Aid of the Flying Spot Microscope", Brit.
J. Appl. Phys., SUppa No.3, 173 (1954).

4.

Kirsch, R. A., et al, "Experiments in Processing Pictorial Information with a Digital
Computer", Proc.EJCC, 221 (1957).

5.

Unger, S. H., "A Computer Oriented Toward
Spacial Problems", Proc. IRE 46, 1744 (1958).

6.

Gill, A., "Theoretical Aspects of Minimal
Scan Pattern Recognition", Electronics
Research Lab., University of Calif., Series
60, Issue 233, March 23, 1959.

7.

Turnbull, J. R., "lOO-Mc Nonsynchronous
Computer Circuitry, Technical Digest, 1961
Internat'l Solid State Circuits Conf.,
Feb. 1961.

179

4.2

PERIPHERAL BLOOD

I
WHITE CELLS
(LEUCOCYTES)

RED CELLS
(ERYTHROCYTES)

I
LYMPHOCYTES

Figure 1.

GRANULOCYTES

COMPOSITION OF PERIPHERAL BLOOD

(/)

l.LJ

EXPOSURE

~

t;

0.8

~ 0.6

~::e

o

...J

~

EXPOSURE

~

o

CJ)

EXPOSURE

I

0.4

-0
en 0 0.2

o
; 0.0

~

, I'

24

0

TIME IN WEEKS

Figure 2.

INCIDENCE OF BILOBES IN DOGS

'I

28

,Ir--,--,

33

180
4.2

Figure

3.

PERIPHERAL BLOOD SMEAR

II

a. ORIGINAL

FIELD

b.

AFTER 2ND PASS

Figure

4.

THE "SHRINK" PROCESS

c.

AFTER 10 TH PASS

181

4.2

A B C

H X D
G F E
.f(ISO)

= ApBpCpDEFGHp.

~+(TAZ):: ABC+BCD+CDE+DEF+EFG+FGH+GHA+HAB.

of (TUP)

::

(AB + AB){(BC + Be) [(ei5+cD) + ... + (GH+GH)]}
+ .. " + ( E F+ EF) ( FG + FG) (G H+ GH).

FINALLY

~

f(X p ) =

x [.f(ISO)+.f(TUP)+-F(TAZ)]

Figure 5.

THE "SHRINK" ALGORITHM

MICROSCOPE

VIDICON
CAMERA

SCANNER
CONTROL

.-

_

SCANNER-COMPUTER
LINK

~
-

COMPUTER

~
VIDEO
MONITOR

Figure 6.

DATA DISPLAY
MONITOR

THE CELLSCAN SYSTEM

~

......
(X)

~

MAGNETIC TAPE STORE
( 19,200 BITS)

~
QUANTITIZED
VIDEO INPUT

--

EXTERNAL
SYNCHRONIZING
SIGNALS

TEST PATTERN
GENERATOR

-

CONTROL
~

--

--

,

DELAY REGISTERS
( 194 BITS)

SHRINK LOGIC

+
ISOLATED ONES
COUNTER

Figure 7.

..

COMPUTER BLOCK DIAGRAM

OUTPUTS TO
MON rTOR

~

183
4.2

Figure 8.

Figure

9.

ANALOG VIDEO IMAGE OF BILOBE

QUANTIZED IMAGE OF BILOBE

185
4.3
APPLICATION OF COMPUTERS TO CIRCUIT DESIGN FOR UNIVAC LARC
Gilbert Kaskey
Associate Division Director. Systems Design and Application
Remington Rand Univac
Philadelphia. Pennsylvania
Noah S. Prywes
Consultant to Remington Rand Univac
Assistant Professor. Moore School. University of Pennsylvania
Philadelphia. Pennsylvania
Herman Lukoff
Chief Engineer. Remington Rand Univac
Philadelphia. Pennsylvania
The design of circuits for computers has become in recent years a complex undertaking. The
problem is two-fold. On one hand optimization of
cost and speed is the prime objective. On the
other hand. complexity is increased through factors such as component charact~ristics and life
expectancy. manufacturing techniques. and the
suppression of noise in very large systems. The
complexity makes the use of computers as an aid to
design almost imperative; this was the case in the
desi.gn of circuits for Univac Larc.
Several applications of computers in circuit
design are reported here and demonstrated by case
histories for Univac Larc. The paper consists of
two parts. In Part I. a general description of
the problems and solutions is given. References
to available reports or publications are given
where a more detailed description can be found.
The problem areas can be divided into three categories: evaluation of components and life test;
design of circuits; protection against noise.
Part II consists of a detailed description of the
statistical techniques used in the circuit design.
Part I
A.

Evaluation of Components and Life Tests

Evaluation of Components. The components
evaluated included transistors. diodes. ferractors. ferrite cores. resistors. capacitors. etc.
As an example. the evaluation of Philco surface
barrier transistors will be reported here. This
transistor was selected during the first half of
1956 after a study of many other candidates in
respect to rise time. storage time. gain. current
level at optimum performance and cost. The evaluation program required the testing of a large
number of transistors to determine such parameters
as beta. rise time. storage time. and breakdown
voltage. These values were measured initially and
at various times throughout life test.
In the attempt to mechanize the test data
analysis. a complete library of Univac statistical
routines has been developed. These statistical
routines analyzed and evaluated the available
data. The statistical analysis of empirical data
is greatly simplified if the variate under analy-

sis is normally distributed. Since this is rarely the case in practice. the distributions were
transformed to ones that have Gaussian properties.
Close cooperation with manufacturers was
maintained to insure that the transistors received were the best that could be produced in
the manufacturing process used. Experimental
designs were made to aid in the determination of
the effect of various changes in production techniques. e.g •• resistivity and etching time. on
the several transistor parameters. A UNIVAC@
routine was then used to perform the analysis of
variance necessary for the identification of the
statistically significant variables. The results.
a joint effort between the manufacturer and user.
helped establish production control procedures
which virtually insured that component lots would
5
meet the required specification.
Evaluation of Life Test. Because of the
long life expectancy of transistors it was difficult to ascertain failure characteristics by life
test in a reasonable time period. In other words.
it was not possible to detect significant degradation of transistor parameters over thou~ands of
hours of life test. Since it was nevertheless
extremely important to be able to make a prediction as to the life expectancy of the transistor.
an attempt was made to run accelerated life tests
at elevated temperatures. under severe humidity
conditions and under vibration. with the purpose
of producing gradual deterioration in a reasonable
time. It was hoped that the correlation of such
results with deterioration under normal usage conditions would result in a reasonably accurate prediction of life expectancy. The tests under elevated temperature conditions were the only ones
that proved useful in this respect.
0

0

Transistors were placed on test at 55 c. 65 c.
85°c. and lOOoc. The transistors involved
were first tested for homogeneity by a study of
the distribution of breakdown vOltage and ~.
These parameters also appeared to be the major
cause of transistor failure in the circuits and
therefore are the subject of the investigation.
By studying the behavior of these homogeneous sets
over time, we hoped to obtain as estimate of transistor behavior at 25 0 c.
75 0 c,

186
4.3

To illustrate the statistical methods used,
the determination of transistor life using degradation of breakdown voltage as a criterion will be
discussed. The circuit design indicated that the
breakdown voltage degradation to 3 volts implied a
transistor failure. Theoretical studies had suggested the dependence of the breakdown voltage on
temperature and time as follows:
= A - Be -a/2T v't
(1)

vv:p

where Vp is the breakdown voltage, T is the absolute temperature and t is the age of the transistor in hours.
A least-squares fit was then made to the
data, using fixed values of T, to determine the
"best" values for A, B, and a. The results of
this analysis are given in Table I.
Table 1. Punch-Through Voltage
Accelerated Life Test
Group
No.

Temperature
(oc)

10

55
65
75
85

11

12
13

Least-Squares Equation
(t in hours)
10.6
9.57
11.9
10.9

-

0.0002t
0.0004t
0.0028t
0.0058t

(2)

where m is a slope of the linear least-square
equation and T is the corresponding absolute temperature as shown in figure 1-1.
This equation was used to estimate the slope
at 25 0 c. and the value was found to be approxi-5
.
mately 2 x 10 volts per ho~r. The results of a
fit by eye made prior to the regression analysis
made on the Univac System indicated an average
life of 211,000 hours as indicated in figure 1-2.
In addition to the prediction of accelerated
life tests, life tests under conditions similar to
operatinG conditions were conducted. Improved
predictions were only possible as more data became
available.
B.

1)

More efficient circuit optimization in terms
of the predetermined functionals. that is,
cost. by utilizing the speed of computers.

2) The generation of component specifications
which computer programs correlate with the
capabilities of the designed circuit.
Circuit design and optimization start with
given circuit schematic configurations and a
performance requirement. In the case of computer
circuits, the latter can be stated. for initance.
in terms of fan in (number of logical inputs),
logical operation, fan out (number of circuits to
which the output is connected), and the delay per
circuit. The objectives of the design are to
single out one of several suggested circuit configurations and determine component parameters so
that cost is minimized.
The process can be roughly divided into
three steps:

Because of the assumed exponential relationship, the least-squares straight line was obtained
for the logarithm of the slope as a function of
reciprocal temperature. The equation thus obtained
was
m = -1.3 x 1015e-(14278/T)

LARC attempting to mechanize all steps previously
performed manually. Our objective has been to
completely automate the design steps required in
going from proposed circuit schematic configurations to the development of an optimized circuit.
The process will consist primarily of computer
programs using a detailed mathematical model.
There are two advantages to such a process:

Circuit Design

Mechanization of the various steps involved
in circuit design for Univac Larc has served as
the basis for research whose objective is the
complete mechanization of design and fault diagnosis of transistor circuits. Research has been
in progress since the initial design phase of

1) Generation of d-c circuit equations.
2) D-c circuit design.
3) Optimization of the Circuit for Reduced

Dela~

Generation of d-c circuit equations. A code
was devised for transferring the circuit information in a schematic diagram into the computer.
This code is completely reversible; that is, the
original circuit diagram can be derived uniquely
from the computer code.
The nodes of the circuit are assigned a number mlm2. Each branch is uniquely determined by
its endpoints (that is, the branch with nodes mlm2
and nln 2 as endpoints is called branch
mlm2nln2PlP2). An additional set of numbers PlP2
is necessary to differentiate two or more branches
which have common endpoint notes. After listing a
number which identifies a branch, the components
of the particular branch are listed. When this
has been done for all branches. the circuit has
been completely described. Each component is
associated with a letter of the alphabet (for
example, resister R. emitter E, battery B, and
diode D). In the format used each letter is
followed by a tWO-digit number to differentiate
components which are of the same type.
The component closest to node mlm2 is listed
immediately after the description of the branch
mlm2nln2PIP2' followed in order by the remaining

187
4.3

components in the branch; thus, the component next
to node n n is the last in the series. As an example, th~ 10ll0Wing is the format for the circuit
shown in figure 1-3.
00
00
00

01
01
02
02
03
03

00

ROI

000

DOO
ROO

GOO
BOO

500
R02

04

00
01
00
00
00
00

04
04

00
00

R03
GOI

000
BOI
000
000
B02
000

01
04
04
02
04
03

QOO
EOO

This method of recording the information for
given circuit configuration gives a unique representation so that each component and its location is specifically described.

&

The two basic methods available for the generation of the circuit equations (that is, the
loop and the nodal-branch techniques) have been
considered. The loop method has the advantage of
yielding fewer equations, since not all of the
possible variables are included. If the,additional variables are eliminated from the nodal-branch
equations, the reduced set is identical with the
set derived by using the loop method.
The nOdal-branch derivation has a single advantage which is extremely important from the
standpoint of a computer solution: the equations
are derived very systematically. Thus the process
of mechanization, which will result in a set of
redundant equations, is easily implemented. On
the other hand, when using the loop equation
method, there are frequently too many loops to
allow extracting those which make up the system of
redundant equations.
It was decided, therefore, to concentrate on
the more systematic nodal-branch method, and then
eliminate any irrelevant variables.
If there are n nodes in the circuit, n-l independent nodal equations can be generated. (It
can be shown that if the n-th equation is generated, the result can be derived from the other n-l
equations.) Referring to the circuit in figure
1-3, the n-l equations generated are:

V - VI
4

=

+ V
V
BOI
R02

V3 - V2

VQo

V4 - '2

VE

V4 - V3

VR03 + VB02

These four nodal and nine branch equations.
then, represent a complete set of redundant
equations which fully describe the system.
The current-voltage relation of diodes and
transistors is nonlinear. In order to simplify
computation, these nonlinear curves have been
approximated and replaced by linear segments in
the regions of operation which are of interest.
Thus. whenever the voltage-current relationship
of a diode is considered. the following condition
is employed:

where VD and ID are the voltages and current respectively through the diode. (See figure 1-4.)
The constants DO and RD are unknown quantities to
be determined in the calculations. In effect. a
variable (Vo). which changes with input conditions. has been replaced by two quantities (DO.RD)
which remain constant through varied input
conditions.
In a similar manner the following substitutions can be made for the transistor currents:

1010 + 1040 .. 1041 • 0
-1 010 ' + 1120 .. 1140

=0

-1120 + 1230 + 1240

=0

-1 230 + 1340 + 1341

=0

The branch equations represent the total
voltage drop across each branch. For the sample
circuit in figure 1-3 the equations are as
follows:

where SO' RS' Kl and K2 are constants determining
the two straight line approximations; IS and 10
are the currents through 5 (base) and Q (collector)
respectively; and Vs and VQ are the voltage drops
across the base and collector. (See figure 1-4.)
As in the case of the diode. unknown quantities (VS' VQ)' which change with input conditions.
are replaced by quantities (K l • K2• SO, RS) which
remain constant through varied input conditions.

188

4.3
D-c circuit design. In the case of UNIVAC
LARC the optimization of cost consisted mainly of
reducing the required Beta of transistors used.
There are two criteria for calculating the required Beta: Worst case design, statistical
design.
In the case of worst case design, the parameters (such as resistances, supply voltages, etc.)
are multiplied by a factor which represents the
maximum tolerances allowed so that the Beta of the
transistor involved becomes minimum.
In statistical circuit design, Monte Carlo or
analytical 6 ,7, methods are used to obtain a distribution of the required Beta of the transistors
as a function of the distributions of the circuit
parameters (resistances. supply voltages, etc.)
Optimization of the Circuit for Reduced Delay.
Examination of the above circuit equations shows
that the number of unknown variables exceed the
number of equations. Therefore there is no unique
solution. Generally delay decreases with increase
in Beta, although the delay would depend on many
other parameters as well. The purpose of the optimization is then to determine a unique circuit
having the lowest Beta requirement such that the
maximum delay allowed in the circuit specifications is not exceeded.
The transient behavior of the circuit can be
determined experimentally, analytically, or through
statistical studies.
In the experimental transient study the unknowns in the circuit equations are divided into
so-called dependent variables and independent
variables. The number of the dependent variables
is equal to the number of equations. The determination of optimum values for the independent
variables inplies unique solution of the circuit
equation which represents the optimized circuit.
The problem then is to vary the independent variables and determine experimentally the values
corresponding to minimum delay. This can be an
iterative process where one of the independent
variables is varied while the others are kept constant. 8 Figure 1-5 illustrates such an experiment
where R3 is varied for a transistor Beta of 9 and
the on-base current equals 1.15 DB. The 'optimum
value of R3 is found to be approxiDBtely 750 ohms.
A large number of circuits have to be computed in the process of optimization using circuit
equations modified for worst case design. These
circuit equations are found to be nonlinear.
Solution of the system of equations by computer is
of significant advantage over DBnual computation,
especially when the system of equations is nonlinear.
The theoretical relationship between circuit
delay and the several circuit parameters has been
found to be extremely unreliable for prediction
purposes and therefore abandoned.

The third approach involves the statistical
determination of the relationship between the circuit parameters and the circuit delay. Specifically, the determination of the regression of circuit delay on the transistor parameters has been
e~tablished.
This method is the key to the circuit design procedure used and therefore it is
described in considerable detail in Part II of
this paper.
A functional relationship, developed statistically, serves two closely related purposes. It
is not only necessary for any work in statistical
circuit design but offers the following advantages:
1) Changes in production control, which, experience indicates, occur frequently, may improve
the parameters of the selected transistors in
some respects and degrade them in others.
Also, with the rapid developments in transistor production, newer, better and less expensive transistors become available. A correlation between transistor parameters and
circuit performances allows the transistor
manufacturer the freedom of changing production controls to improve one parameter at the
expense of others so that transistors improve
in yield and cost. Also a freedom is maintained to purchase transistors from many
sources.
2) The circuit that has been designed for a
particular specification may be useful in
other applications in which a less expensive
transistor would satisfy a functional specification calling, for example, for slower
speed or less gain. The regression of circuu
performance on the several component parameters allows such a change to be made without additional design or experimental check.
C.

Reduction of Noise and Delay in Backboard
Wiring

The transmission delay increases with the
lengths of the wire and distributed capacitances
representing connectors, wires, etc. The noise
pick up (that is, voltages and currents induced
in a wire by pulses in other wires in its proximity) increases with the lengths of the wires but
decreases with the total distributed capacitance
on the wire. The assignment of elements on the
backboard is made to reduce delay and noise pick
up. In Univac Larc, the logical designer decided
where groups of circuit elements were to be placed
on the backboard, based on his familiarity with
general information flow path among organs of the
computer. The circuit elements were then assigned
t? ~pecific printed circuit packages. Preliminary
wIrIng procedures were then run on Univac I to
determine whether bad cases, representing wires
exceeding length or capacitance, existed. This
was done on the basis of wire length calculation
between terminals, and calculation of total capacitance represented by connectors and wire lengths.
When bad cases existed, the logical elements were
moved (by decision of the logical designer) in an
attempt to reduce and/or eliminate the bad case

189
4.3
conditions. 9 lterations of this procedure,
partly manual and partly automatic, are continued
until wire lengths and capacitances are reduced to
a tolerable level.

A.

Research on computer placement of circuit
elements on the backboard has continued after completion of the layout of Larc-Univac. A suitable
algorithm has since been developed for performance
lO
of this task.
The algorithm is capable of minimizing the longest wires on the backboard or the
total length of all wires combined.

Driving circuit: Flip-flop whose output voltage
has exponential rise time to 70% in 40 ~s.

Part II
The engineer who has as his assignment the
design of a transistor circuit to perform according to a predetermined functional specification
has a choice of two courses in designing the circuit and specifying the transistor.
One approach, in general use, is to determine
by measurements the worst parameters of a selected
type of transistor, to employ these parameters as
the limiting criteria in a worst-case design, and
to ch~ose the other components in the circuit to
optimize speed or gain, for example, in a nonrigorous way.
In the second approach, discussed in this paper, the engineer designs a circuit for a typical
transistor which performs to the given functional
specification. The other circuit components are
selected to optimize the operation with this transistor. The dependence of the functional operation of the circuit (for instance, its gain or delay) upon the parameters of the transistor is determined over a wide range of variations of these
parameters through statistical studies. Transistor parameters are then determined for each range
corresponding to the functional specifications.
Information on the dependence or correlation
of parameters is valuable to both the transistor
manufacturer and the circuit designer. One transistor parameter can be improved at the expense of
another so that the transistor improves in production yield, cost, ete., without harmful effect on
the operation of the circuit. A change in production control, rather than being harmful, can be
helpful. Various types of transistors are candidates for use in the circuit without additional
deSign or experimental work, and the same circuit
can be used to satisfy a number of specifications,
changing only the type of transistor. The circuit
deSigner can use the same information for statis4
tical circuit desig~ as opposed to the worst case
deSign, thus effecting additional savings.
The subject approach will be illustrated by
a case history of a circuit design. To relate a
given circuit specification to the transistor
parameters involved, a considerable amount of computation is necessary, which has been carried out
on a UNIVAC I data-processing system.

Description of the Circuit

The functional specifications of the circuit
to be deSigned were as follows:

Input voltage:

Pulse from -2.9v to -o.3v

Input current:

Pulse from 0 ma to 4.5 rna.

Output voltage:

Pulse from -2.9 to Ov.

Output current:

Pulse from 0 ma to 52.0 mao

Maximum output capacitance:
Maximum load:

1000

~~f.

32 standard circuits.

Delay:
High-speed range
Medium-speed range
Low-s peed ra nge

Minimum
22 IDJ.I.s
33 Il\J.s
44 ~s

The circuit configuration
figure 2-1. Delay is measured
of the clock pulse driving the
beginning of the output of the
The delay-measuring circuit is

Maximum
165 m~
205m~
245 m~

chosen is shown in
from the beginning
flip-flop to the
leading circuits.
shown in figure 2-2.

Since the minimum delay is not critical in
this configuration, design effort continued with
the input and loading for maximum dela~as shown
in figure 2-3. Measurements were made when the
transistor was turning off, since maximum delay
occurs at that time. A typical transistor was
selected to give a delay in the medium range. The
values of the other components in the circuit that
would minimize delay were then determined experimentally. The Surface Barrier Transistor (SBT)
in the circuit (figure 2-1) has a relatively small
effect on the delay of the entire circuit; therefore, the determination of worst parameters for
the SBT was feasible. This paper will deal with
the regression of the parameter of the second
transistor and on the performance of the circuit
as a whole.
Like the choice of the circuit configuration
(figure 2-1), final determination of the values of
components other than transistors was based upon
other experimental work not relevant to the work
discussed here.
B.

Parameters of the Transistors

Four parameters and circuit delays were measured for each of 360 transistors. The six parameters normally specified by the manufacturer are
Breakdown voltage, Leakage current, Current gain
(~), Rise time (T), Storage time (S), and Peak
base-to-emitter voltage (V). The first two, which
affect mainly the dc operation of the circuit,
have no significant effect on circuit delay. The
remaining four parameters, assumed a priori to affect circuit delay significantly, were measured in
each transistor. Current gain (~) was measured at

190
4.3

a constant collector current of approximately 100
milliamperes and a collector voltage of -0.6. Because of dc considerations, gain must exceed 30
under these conditions. The circuits shown in
figures 2-4, 2-5, and 2-6 were used to measure T,
V, and S, respectively. Total circuit delay (6)
was assumed to be a function of V, ~, T, and S.
In addition, 6 was measured for each transistor,
with output loading which corresponds to maximum
delay (figure 2-3).
The transistors tested were in three groups:
236 transistors of type GT762 (taken from two production runs), 99 transistors of type CK762, and
25 transistors of type TA1830.
No theoretical relationship among the measured parameters was assumed. Each measurement
was performed twice. Transistors whose values did
not check within the accuracy limits of the measuring device, were eliminated from further consideration but the number of these was negligible.
C.

Statistical Studies

Interdependence studies between the parameters and delay values were undertaken first.
The tools of regression'analysis l were used to
ascertain whether, in general, circuit delay can
be predicted from known parameters of a transistor.
The second step was the establishing of a functional relationship between circuit delay and the
several known parameters. This function formed
the basis for the successful determination of
transistor parameters for the delay ranges.
Studies in Regression Analysis. To investigate whether there is any direct relationship between circuit delay (6) and any of the four transistor parameters listed, the measured value of
circuit delay for 236 transistors, assuming these
to be a representative sample of the population of
all GT762 transistors, was plotted against each of
the parameters. It was assumed that the regression of 6 on each of the parameters V, ~, T, and S
is linear. The scatter diagrams of 6 versus each
of the transistor parameters and fitting of linear
regression equations are given in figures 2-7 thru
2-10. The mean 8 values are connected in a line
of best fit, shown as a light line; the line of
regression, shown as a heavy line, is defined as
the linear function of the form y = mx + b, which
fits the means of arrays best, in the least
squares sense. The fitted linear regression equations are given below:

ships are obtained by using the last three of the
above equations. Though the estimated coefficient
of regression between 6 and V is greater than the
corresponding coefficient between 6 and any other
parameter, it is not statistically significant
since the estimated variance of this coefficient
is very high. Further investigation made to determine whether any direct relationship between
the parameters V, ~, T and S existed, indicated a
strong direct relationship between both V and T
and also between T and ~.
Delay as a Function of Sixteen Expressions.
A UNIVAC program was used to find the linear fit
and regression coefficients be.tween 6 and the following 16 parameter expressions:
V,

~,

2
T, S, 1I ~, TI ~,T 21~,T/~,

v/~,

VT, TS, S/~, T2, 1/~2, lis, VS
It was assumed that the coefficients of regression
between 6 and other parameter expressions were not
significant. The program revealed highest positive regression between 6 and terms T, T2, T2/~,
VT. S/~, and TS. These six parameter combinations
were chosen for a function with linear constants
as follows:
6· KIT + K2T2 + K3VT + K4T2/~
(1)

+ KsS/~ + K6TS + K7
A second program was subsequently written to
apply the least-squares fit criterion to the 234
sets of transistor data for the given equation.
The normalized equations (seven equations, seven
unknowns) of the fit were solved by the Crout
Method 3 • A third program was developed to test
the curve fit; that is, to compare the calculated
6 with the observed and to determine the individual term contributions. These programs revealed
that the KsS/~and ~ TS terms contributed little
to the value and could be dropped, thus simplifying the function to the following:
6

=

KIT + K2T2 + K3VT + K4T2/~ + K7'

where Kl - 1.666, K2 = 0.001, K3
K4

= -0.175,

and K7

(2)

= -2.717,

= 129.135

+ 197
+ 146.5
+ 171

A relatively simple evaluation of the normality of the distribution of errors based on this
regression equation is indicated in figure 2-11.
The cumulative dis,tribution of errors would appear
perfectly linear in the representation of a normally distributed population.

By using t tests 2, it was found that the regression between 8 and V is not statistically
significant but the regressions of 6 on the other
parameters are highly significant, i.~., at the 1%
level. There is a definite indirect relationship,
then, among ~, T, Bnd S. The indirect relation-

Discussion of Accuracy of Prediction Using
the Function. A lot of 99 \ype GT762 transistors
from a later shipment was measured to determine the
applicability of the derived functions. equations
0) and (2). The distribution of errors between
the equation prediction (2) and the observed

8

=

6
6 =
6 =

-33.4V + 194
-0.08l2~

0.577T
12.45

191
4.3
values of delay appeared normal, with a mean of
8%. A change in the constant term or inclusion of
dependence on S would correct the function as applied to this particular group and shift the mean
to zero.
D. Range Deter.iaation
Since the circuit under design specifies use
in one of three delay ranges, rather than a specific delay, a method for classifying transistors
into the three ranges according to known parameters would serve the purpose. The method capitalizes on the relationship established in the
search for a predictive function.
The 236 units first investigated were plotted
on a T-ordinate, ~-abscissa graph, and labeled
with their observed a values. Arbitrary 6 ranges
were found to separate themselves fairly well into
various regions of the plot; rough borders were
sketched between regions following the best range
separations. These T (~) curves descended exponentially at low ~ values, and leveled off horizontally as ~ increased, suggesting the functional relationship:
T

= Kl + K2e-K~

(4)

The a-labeled points were separated into the
three designated groups: 0-155 ~s, 156-195 m~s,
and 196-234 m~s; and the two borders were added.
A program was devised to fit the border ~, T data
to the suggested functional expression, yielding
the constants Kl , K , K3 for each curve. The
2
smooth exponential decay curves were drawn in to
separate the data. The results, for the first lot
of 234 type GT762 transistors, are described as
follows:
1)

2)

In the high range (196 < 5 < 235), eight
units out of 52 occurred which did not belong. Their values were 180, 180, 186, 192,
192, 192, 192, 194 m~. Thus there were only
two outside of the tolerance criterion,
± 10 m~. This tolerance was selected arbitrarily by adding the measurement tolerances
of T and 5, each ± 5 m~s.
In the medium range (156 S a S 195) nine
units out of 179 occurred which did not belong. One unit was below (152), and eight
units were above (196, 196, 196, 198, 198,
198, 200, 200). None of these was outside
the ~ 10 m~ tolerance region.

The results suggest a very accurate separation.
The lowest region, where there was insufficient
data available, was checked on another set of
transistors. The results are discussed below.
Figure 2-12 is a graph that can be used to
sort transistors by ~ and T measurements. Once
the measurements for each transistor are made, the
~-T point on the graph establishes the delay range
of the unit.
Discussion of Accuracy of Prediction Using
Ranges Determined. With the method just indicated,
using the transistor delay-range chart with the
originally derived borders (figure 2-12), the new
lot of 99 type GT762 transistors was plotted.
There were 28 transistors in the high range
(196 ~ 0 ~ 235).
In the medium range (156 ~ 0 ~ 195), 69 units
occurred, of which 11 did not belong. Nevertheless, of these 11, ten units were acceptable under
the tolerance limits, indicating only one misplaced.
In the low delay range (6 ~ 155), only two
transistors occurred, both of which were correctly
placed.
A linear shift in the borders of the a ranges
wobld take care of the errors of misplacement.
These results are strongly indicative that new
lots of the transistor have some property changes
that can affect our application, unless additional
parameters such as storage time (S) are considered.
An excellent prediction for the TA1830 data
was achieved by the transistor delay-range chart
(figure 2-12). Of the 25 units tested, 22 fell
within the predicted range and 3 were borderline.
The borderline cases were so close that, within
tolerance limits, they could be placed in the
correct categories.
In contrast to the broad range of delay
values in the original 236 type GT762 transistors,
these RCA TAI830 units were mostly confined to the
lowest delay range.
Acknowledgements
The authors gratefully acknowledge the
assistance of the many groups and departments of
Remington Rand Univac who contributed to the
success of this paper. Special acknowledgement
is given to P. Krishnaiah and P. Steinberg.

s

References

4) The two border equations are as follows:

lEzekiel, Mordecai. Methods of Correlation
Analysis. 2nd ed. New York: Wiley, 1941.

3) The low range (a 155) contained only two
units, both of which were correctly placed.

= 30.97 +
T = 62.18 +

T

60.58e-o·0157P at 5

= 155

112.41e-o,OI66~ at 5

= 195

(5)
~)

2Johnson, P. o. Statistical Methods in
Research Chap. V. Prentice Hall, Inc., 1944.
3nildebrand, F. B. Introduction to Numerical
Analysis. p. 429. McGraw Hill, 1956.

192
4.3

4Gray, Harry J., Jr. An Application of
Piecewise Approximations to Reliability and
Statistical Design and Proceedings of IRE. July,
1959.

~emington Rand Univac, Division of Sperry
Rand Corporation. Statistical Techniques in
Transistor Evaluation Final Report. Dept. of
the Navy, Bureau of Ships NObs 72382. Applied
Math Department: Remington Rand Univac, Philadelphia, Pa. April, 1959.
6Senner, A. H., and Meredith, B. Designing
Reliability into Electronic Circuits. Proc. Nat'l
Electronics Conf. Vol. 10, pp 137-145. Oct. 1954.
7Gray, H. J., Jr. An Application of Piecewise Approximations to Reliability and Statistical
Design. Proc. of the IRE Vol. 47, No.7, pp.
1226-1231. July, 1957.
8Remington Rand Univac, Division of Sperry
Rand Corporation. Univac Larc Highspeed Circuitry
Case History in Circuit Optimization. Prywes,
N. S., Lukoff, N., and Schwartz, J. Remington
Rand Univac, Philadelphia, Pa.
9Remington Rand Univac, Division of Sperry
Rand Corporation. The Univac Prepared Engineering
Document Program. Williams, T. Remington Rand
Univac, Philadelphia, Pa.

l~emington Rand Univac, Division of Sperry
Rand Corporation. The Backboard Wiring Problem:
A Placement Algorithm. Steinberg, L. Remington
Rand Univac, Philadelphia, Pa.

193

4.3

64 ~--------------------------------------------------------~

56

,,,
,,,
,,
,,,

--~--ACTUAL

THEORETICAL

48

I
I

40

w
:::to«

I

(!)

IJJ

z

32

v
~

P

)(

IJJ

0-

J

0

...J

(f)

24

15

25

35

45

55

65

TEMPERATURE (OC)

Figure 1-1.

PUNCH-THROUGH VOLTAGE SLOPE PREDICTION

75

85
701

,j:::..

I-'
(0

W,j:::..

12

II
4

25°C Vp =1O.6-0.36xl0- t

---

10

- - - ------.'-1.TLT2~',';~~ ~~~~~E
"

9

..........

"-

II)

>u 8

m

u>
~

7

0

>

~
x

6
65°C V p =10.6-4xl0- 4 t


0

0::

5

X
lX
0

4

z

""

75°C V p =10.6-28xl0- 4 t

85°C Vp =10.6-58xl0- 4 t

::>

Q.

3

2

o

~'------------------~

100

__________ ______

400

~

1000

~~

______

2000

-L~

__________

4000

~

________-L__

10,000

TIME (IN HOURS)

Figure 1-2.

PUNCH-THROUGH VOLTAGE ACCELERATED LIFE TEST

20,000

~

____

~

__

40,000

~

______

~

100,000
700-Rl

195

4.3

""....-----...--....,
,/

/
INPUT
TEST
POINT

"

BASE
CONNECTIONS

\

D

I

/

./

TRANSISTOR

/

......

OUTPUT
TEST

...... "

COLLECTOR
CONNECTION

"

\

POINT

NODE3~

/

\

\
\

I

\
I
I

,
I

R/

\
\

NODE 1

I

/

\
\

EMITTER
CONNECTION

\."

'R2

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

...........

/

I

_--

/
/
./

G

+

_./

TP

+
"-NODE 4
4402

Figure 1-3.

SAMPLE CIRCUIT FOR FAULT DIAGNOSIS

-Q
RS -ON

-'--~------------~IS

a. Diode

b. Transistor Base

c. Transistor Collector
4103

Figure

1-u.

VOLTAGE-CURRENT CHARACTZRISTICS OF A DIODE)
A1~ BASE AND COLLECTOR OF A TRANSISTOR

196
4.3

80~---------------------------------------------,

75 I-

0

TRANS I STOR /3 =9
BASE CURRENT 18 =1.15mo

3:'U

....

CI)
en

cn::t.
cnE
0D:cn
u .... 70


I-

~u
....J~

lLIU

0

65 IOPTIMUM

~~

0
0

I

I

0.4

0.8

I

1.2

1.6

2.0

-L
Xl03
R3

Figure 1-5.

DELAY AS A FUNCTION OF THE 1/R3

2.4
5634-Rl

197
4.3

+t2

+0.75
617

Figure 2-1.

SCHEMATIC DIAGRAM OF CIRCUIT

LOADING
CIRCUIT

FF
+
CLOCK
PULSE

MEASURED
DELAY
618

Figure 2 -2.

DELAY MEASUREMENT

198

4.3

r--i\----I
rV-I I
I

I

I

31 LOADING
CIRCUITS

FASTL-J'
________________~_+~
~--~ FF

~L~PC~~~~~;_ - - -

I UN DER TEST
I FF

--\J

I
I

I

I
I

CLOCK
PULSE

I

L ______

-~

~----------8mox.----------~
619

Figure 2-3.

INPUT AND LOADING FOR MAXIMUM DELAY

-4.1

MERCURY - RELAY
PULSE
~--~~~-+~
GENERATOR

a. Circuit

I-

5fLsec

-I

lL...--1t_ _ _----I1
b. Input Voltage

O-----~========~~~~~~~-­

_4.1---..:::31;.;.;~9::.;'lf~4~-------------------~
~r:~

c. Output Waveform
620

Figure 2 -4.

MEASUREMENT OF RISE TIME

-4.1

Vc
~-8V

50mJLsec TC

r----.,
MERCURY-RELAY~
PULSE

: 100Q

Wv

I

GENERATOR

I

I

I

"NY ._ .. -

I

Q

\.

I 50JLJLf -r- I
I
-L.. I
L ___~...J
a. Circuit

I'
•
I..

5~..c
b. Input Voltage

\

-v
_I

I
T

.,

a. Circuit

Vc

~Ts=E
b. Output Voltage

+3V----~--~

L

o

c. Output of RC Circuit

-12V

__- -__-

-....,""'--+----___
C. Input Voltage at A

622

d. Bose -to- Emitter Waveform
621

Figure 2-5.

MEASUREMENT OF PEAK BASE VOLTAGE

Figure 2 -6.

MEASUREMENT OF STORAGE TIME

~~
•
(0
(.to) (0

~

N

•we
e

240

230 l -

ONE POINT
TWO POINTS
• THREE OR MORE POINTS
e·

o

220

•

210

u
Q)

In

::l
E

ro
><{
...I

w

•

•

•

200

'90~

8 =- 33.4
r

V

+ '94

••

/

·•

=-0.0495

•

•

•

0

I-

~
0:

180

U
170

160

I
I

•

150

140
0.20

0.21

0.22

0.23

0.24

0.25

0.26

0.27

0.28

0.29

0.30

0.31

0.32

0.33

0.34

BASE-TO-EMITTE;R VOLTAGE: V(VOLTS)

Figure 2-7.

SCATTER DIAGRAM AND ROORESSION LINE FOR CIRCUIT DELAY AS A FUNCTION OF BASE-TO-EMITTER VOLTAGE

0.35
607

240

230

• ONE POINT
o TWO POINTS

220

210

..

'0
Ql

~
E 200

8 = -0.0812 f3 + 197

(.0

r = -0.393

;>.:

«
....J
lLJ

0

190

l-

S

(,)

a::

(,)

180

170

160

150

140 I
50

!

70

90

110

130

150

170

190

210

230

250

270

290

310

BETA:f3

Figure 2 -8.

330

350
608

SCATTER DIAGRAM AND REGRESSION LINE FOR CIRCUIT DELAY AS A FUNCTION OF BETA

~N

•w· ......
0

.l::>.~

eN

240

230

220~

• ONE POINT
o TWO POINTS

• THREE OR MORE POINTS

8 = 0.577

T

+ 146.5

r = 0.7895

210

'0
Q)

i

200

5

c:o
>-

«

...J

190

w

0

I-

:;

u

0:

180

U
170

160

..

150

140~1

o

______L -_ _ _ _-L______L -_ _ _ _-L______L -_ _ _ _-L______L -_ _ _ _-L______L -_ _ _ _-L______L -_ _ _ _- L______L -_ _ _ _- L_ _ _ _
10
20
40
30
50
60
70
80
90
100
110
120
130
140
150
~

TRANSISTOR DELAY: T (mJLsec)

Figure 2-9.

SCATTER DIAGRAM AND REGRESSION LINE FOR CIRCUIT DELAY AS A FUNCTION OF TRANSISTOR DELAY

609

o

~

240

230

220

• ONE POINT
o TWO POINTS
• THREE OR MORE POINTS

8=12.45+171

210

UQ)

1

200

.5

ro
><{

....J

190

lLJ

0

!::
::J
u
0::

180

u
170

160

150

140
0.50

0.90

1.00

1.10

1.20

1.30

1.40

1.50

1.60

1.70

1.80

1.90

STORAGE FACTOR: S (j-Lsec)

Figure 2-10.

2.00
606

SCATTER DIAGRAM AND REGRESSION LINE FOR CIRCUIT DELAY AS A FUNCTION OF STORAGE FACTOR

~N

•

0

ww

tIl-N
•
0
Wtll-

9

8

I--

7 l-

z
52

I-

6

I--

5

I--

4

I--

3

I-

2

'-

(,)

0

lIJ
0::

-

Cl.

~
0::

0

0::
0::
lIJ

I-

z

lIJ
(,)

0::
lIJ

Cl.

./

0
-t

--

-4 -5 -2
-3

-6

-

-7

l-

-8

I-

-9
O.Ot 0.050.t

v

v/

V

V

V

/

I

(

2

to

20

30 40 50 60 70 80

90

PROBABILITY OF OBTAINING GIVEN ER,ROR OR LESS

Figure 2-11.

I

//

/1)

-

l)

98 ' 99

99.9

99.99
610

CIRCUIT DELAY PREDICTION, 234GT 762 TRANSISTORS TEST FOR NORMALITY OF ERROR DISTRIBUTION

120

flO

100

\

90
UQ)

"l.

~

"'\.

'"E:t..

j::'

80

CD

~

w

::E

~

w
en
ii:

--

CIRCU IT DELAY RANGE:

196-235mJLsec

~

r-,...

70

t::::L

a::

0

--r...- r---a...--

I-

en

en
z

«
a::

60

I-

50

~h

40

~

--

®

CIRCUIT DELAY RANGE:

~h

156-195mJLsec
"-

~

1"1-I L - - -,
--

0

30

CIRCUIT DELAY RANGE:

0-155mJLsec

o
50

70

90

110

130

150

170

f3

Figure 2-]2.

I

I

190

210

230

(CURRENT GAIN)

250

270

290

310

330

350
605-R'

CIRCUIT DELAY RANGES

~N

•

w

0
C11

207
4.4

WIDE TEMPERATURE RANGE COINCIDENT CURRENT CORE MEMORIES
R. S. Weisz and M. Rosenberg
Ampex Computer Products Company
Culver City, California

A desired extension of man's intellect is to
extremes of temperature where the human brain
does not function effectively. Both military
and indus~rial applications call for digital
computers which can operate from polar cold to
equatorial heat. In the space age, these limits
must be broadened still further. Unfortunately,
core memories are generally somewhat less tolerant of temperature changes than are human beings.
Chiefly at fault is the ferrite core itself,
because of its high temperature coefficient of
coercivity, 0.5 to 0.7% per degree Centigrade.
Nevertheless, since ferrite cores have proven
to be highly reliable, reasonably fast, light in
weight, and modest in power requirements, some
means of overcoming the high temperature coefficient is needed to adva~ce the art of computing.
A number of such schemes have been proposed.
They include: (1) Compensating the drive circuits for the decrease in required current with
rising temperature; (2) putting the memory in a
thermostatically controlled chamber above the
highest ambient anticipated; (3) using multiapertured ferrite cores to provide a wide margin
of current tolerance; (4) using a "word select"
or linear select mode of address; (5) using
metal alloys which have inherently lower temperature coefficients. Each of these methods may
have some merit; certainly all have serious
limitations.
Clearly, what one would like is a ferrite
core having temperature independent properties.
We have recently developed such a core. Aside
from a low temperature coefficient. of coercivity,
the new core has a better disturb ratio than
that of standard ferrite memory cores over the
range -55° to +lOO°C. Valuable for other purposes as well, a high disturb ratio also aids
in decreasing temperature sensitivity. An array
of 1092 bits has been built from these cores
using a simple toroidal geometry. It has been
successfully operated over the range -55° to
+125°C in a coincident current mode without
temperature or current compensation.
In the remainder of this paper, we shall discuss in further detail the problem of wide
temperature memory operation, the prior art on
the subject, the new ferrite's characteristics,
and the new memory's operation.

Problem of Wide Temperature Operation
To match commonly called for military specifications, we shall discuss memory operation over
the range -55° to +125°C. It is important to
note that there already exist components other
than memory cores which can be operated over
this same range. For example, transistors,
diodes, resistors and capacitors, all necessary
in memory circuits, are available.
A fundamental problem involved in operating
a memory core over a wide temperature range will
be discussed in reference to Figure 1, which is
a superposition of hysteresis loops taken at
-55°, +25°, and +lOOoC for a conventional memory core. It is readily seen that a current
which is sufficient to switch the core at -55°
is more than twice the required current at 25°C.
Therefore, with the usual coincident current
scheme, if half such a current is applied at
room temperature, it will completely switch
what are intended to be half-selected (unswitched)
cores. Conversely, at low temperatures of
operation, it is impossible to switch a core
using the smaller current required at higher
temperatures.
In any usable memory core, the ratio of the
threshold to the full drive (or disturb ratio,
Rd, as it is commonly called) is actually'
greater than 0.5. The resulting current tolerance can be used to provide a margin against
drift in the drive circuits or changes in temperature. Conventional type memory cores are
known with Rd as high as 0.67. If no current
drift need be provided for, such a core can be
operated over a range of approximately 50°C.
A more realistic view is to allow for a drift
of ±10% in drive current leaving an operating
range of only 20°C.
One of the first methods suggested for extending the temperature range of operation to
the desired 180°C was to provide compensation
in the drive circuitry to decrease the current
with increasing temperature so as to match the
decreasing coercive force. Over a large temperature range a close match is difficult, but
perhaps possible. A more serious problem,
however, is also present. Those cores which
are being switched during a temperature change
will obviously settle on the appropriate new
hysteresis loop as shown in Figure 1. But what
happens to a non-selected core? In the absence

208
4.4

of an applied field, it is believed that it
remains at or close to its original remanent
point. Thus, at the end of the temperature
change there will be cores on two different
hysteresis loops. It follows that spurious
signals would then be obtained upon interrogating the memory. Despite this limitation,
current compensation is successfully used over
a narrower temperature range providing operation
from 0° to 60°C.

presence of moisture. None of the thin metal
film devices has yet found wide acceptance.
It is believed that uniformity of individual
storage elements has been a more serious problem here than with discrete ferrite toroids.
Furthermore, these elements have to date be~n
operated only in word select type memories.
This results once again in the use of a greater
number of semi-conductors than would be used in
an equivalent size coincident current memory.

Another approach that avoids the above problems completely, is to place the memory ia a
container which is kept at a controlled C' nperature above the highest anticipated. l
Although this scheme has the merit of simplicity,
i-t sacrifices some reliability due to the possibility of failure in the heater circuits.
There is a further limitation inherent in the
power requirement of the heaters. For example,
to heat a memory of 3200 bits from -55° to
+125°C, 10 watts are required. l A power level
of this magnitude is far beyond the supply
available in satellites now and in the predictable future; it is also an order of magnitude
above the power requirements of the memol.'~·
itself.

New Ferrite Cores

Still another partially successful approach
is to use multi-apertured cores or
"transfluxors. ,,2 Here one makes use of thr
geometry to provide a wider current tolerance
than is possible with simple cores. Memories
of this type have been operated over the range
-20° to +lOO°C with non-destructive readout.
The principal disadvantages of transfluxor
memories are: (1) Their complicated wiring,
and (2) increased use of semi-conductors. The
cores are also much larger than simple toroids.
Another approach has been to use ferrite
cores in a word select mode of operation. 3
This has the disadvantage of using a greater
number of semi-conductors, resulting in decreased reliability and increased costs.
Finally, it has been proposed to abandon the
square loop ferrites entirely and substitut~
metals or alloys in various configurations. 4 ,5
Permalloy and other nickel-iron alloys have very
low temperature coefficients of coercivity.
The principal disadvantage is their high electrical conductivity. One is forced to use very
thin film to avoid eddy-current losses and as a
result, two new problems arise. First, thin
films are not well suited to wide temperature
cycling because a difference in coefficient of
expansion between the film and substrate causes
cumulative stresses to be set up. Second,
metalic films are prone to destructive oxidation
at elevated temperatures, especially in the

Since none of the previously proposed methods
of wide temperature memory operation seemed
to be free of serious limitations, we embarked
on a program to develop a square loop ferrite
with a low temperature coefficient of coercivity.
A family of such materials has now been found.
Fortuitously, some of these also have a disturb
ratio higher than conventional cores over the
range -55 0 to +lOO°C. Details of the chemistry
and physics of the new core are outside the
scope of this article, but will be given in
another publication. 6 A particular core was
chosen for the memory to be described on the
basis of optimum squareness, temperature coefficient, and speed. The dimensions of the toroid
are standard: 50 mils O.D. x 30 mils I.D. x
15 mils thick.
Important characteristics of the cores are
given in Figures 2 through 4. Figure 2 shows
a superposition of 71 kc hysteresis loops of
the wide temperature range core taken at -55°,
+25°, and +lOO°C. For comparison purposes,
Figure 1 shows the corresponding loops of a
conventional magnesium-manganese ferrite core
with the lowest temperature coefficient we have
found among the presently used materials. One
will observe that the new core has a much lower
temperature coefficient of coercive force
(approximately 0.13% per degree Centigrade as
compared with 0.5% per degree Centigrade for
the magnesium-manganese ferrite). The usual
decrease in saturation flux density with increasing temperature is also much lower for
the new core. Undoubtedly, this is connected
with a high Curie temperature (greater than
500°C vs. 300°C or less for the magnesiummanganese ferrites).
More striking, and more to the point, are
the pulse characteristics of the new core.
Figure 3 epitomizes this data, giving the
following characteristics as a function of
temperature for a constant drive of 1.0 ampere
turn: Switching time, t s , and peaking time, t p '
in microseconds; undisturbed output uVl, and
noise, dV z , in millivolts; disturb ratio, Rd.
This data was obtained with a pulse rise time,

209

4.4

t r , of 0.2 microseconds, a pulse width, td, of
10 microseconds, and 20 repeats of the disturbing current, whose magnitude was 0.5 ampere
turn. The most striking feature is simply that
a core can be operated over the range -55° to
+lOO°C with a constant current. As stated pre-viously, conventional cores at a constant drive
cannot be operated over more than a 50°C spread.
Also important is the disturb ratio which is
greater than 0.69 up to 100°C. Conventional
cores have Ru up to 0.67. With Rd equal to 0.76
at room temperature, it is possible to operate
the new core easily on a triple coincidence
scheme. Although this possibility cannot be
exploited simultaneously with the wide temperature operation, it is a promising lead for future
work.
Figure 4 gives more detailed information on
the core characteristics as a function of
variable drive current. Other parameters of
the drive pulses were made the same as for
Figure 3. It should be stated at this time
that these figures show tentative values for
early cores. Recent improvements have lowered
the noise voltage from that shown with no
deleterious effects on other parameters.
Signal-to-noise ratios of 10:1 without time
strobing are now consistently obtained at room
temperature under normal drive conditions.
Since high reliability is one of the chief
attributes of ferrite memory cores, we felt
that a life test of the new type at an elevated
temperature was essential. In one such test,
cores have been aged in air at 100°C for over
eight months. Periodic test has shown no significant changes in pulse properties. In
another test, cores were aged in a vacuum at
10- 4 millimeters of Hg at 100°C for three months.
At the end of this period, the samples were
returned to room temperature conditions and retested. Again, no significant change was
observed.
Besides the 50-mil diameter toroids, sample
quantities of a 30-mil toroid, an 80-mil toroid,
and a multi-apertured structure have been made
of the new material. All have shown similar
temperature characteristics.
Memory Characteristics
Having established the unusual temperature
characteristics of the new core, system evaluation was carried out with the construction and
test of a model memory array. It was felt that
this-was an essential part of the evaluation
since, at times, components which have appeared
to be usable in individual tests have failed
when operated in systems. Memory elements in
particular are likely to show interactions,
delta noise problems, etc.

A standard type of coincident curre~t array
with 40 x 30 cores was constructed for the test.
With a constant, uncompensated drive, the array
was successfully operated in a temperature test
chamber from -55° to +125°C. Figure 5 shows
the direct output from the sense winding under
worst pattern conditions at +lOOoC; Figure 6
shows the direct output from the sense windLng
at -55°C. The pulse conditions in both cases
are as follows:
Pulse width = 3
Pulse rise time
Repetition rate

~s
0.75~.

= 50

kc

The signal-to-noise ratio at peaking time
of the "one" signal was greater than 50:1 in
both cases. The variation in sense amplitude
output over the temperature range-is less than
2:1 (90 millivolts to 60 millivolts). This
change is well within the dynamic range of a
properly designed sense amplifier, particularly
in view of the excellent signa1-to-noise ratio.
Since the cores nominally switch in a microsecond, they can be used in a 5-6. microsecond
memory system.
During test, the array was operated in a
3:2 selection mode by using 0.667 NI in the X
lines and 0.333 NI in the Y lines. At 25°C
and worst pattern conditions, a signal-to-noise
ratio of better than 10:1 at peaking time was
obtained. This simple experiment indicates
that it is possible to operate in a true threedimensional selection system. Further work is
planned along these lines.
Conclusions
It is believed that the described array is
a prototype for the first true wide temperature
range memory. The well-known resistance of
ferrites to oxidation, corrosion, and radiation
damage can also be used to advantage in adverse
environments. At the present time, a memory
for a satellite is being built with the new
type core. This memory has approximately one
thousand bits operating in a coincident current
mode over a temperature range of -55° to +lOOoc
without compensation of current or temperature.
Other interesting properties and applications
of the new memory core will be discussed in
later papers.

210
4.4

References

Captions

1. Miniature Memory Plane for Extreme Environmental Conditions, R. Straley, A. Heuer,
B. Kane, and G. Tkach. Journal of Applied
Physics, vol. 31, April 1960, pp. l26s l28s.

Figure 1 - Hysteresis Loops of Standard One
Microsecond Memory Core at -55°,
+25°, and +lOO°C. Vertical Scale
680 gauss/dive Horizontal Scale =
0.780e/div.

2. Temperature Characteristics of the Transfluxor, H. W. Abbott and J. J. Suran. IRE
Transactions on Electron Devices, vol. ED-4,
April 1957, pp. 113-119.

Figure 2 - Hysteresis Loops of Wide Temperature
Range Core at -55°C, +25°C, and
+lOO°C. Vertical Scale = 940
gauss/dive Horizontal Scale =
1.4 oe/div.

3. Design of a Reliable High Speed Militarized
Core Memory, M. Stern and H. Ullman.
Program of the Winter Conference on Military
Electronics (IRE), Los Angeles, Feb. 1, 1961.
(in abstract form)

Figure 3 - Pulse Characteristics of Wide
Temperature Range Core as a Function
of Temperature at a Constant Drive
of 1.0 Ampere Turn.

4. Recent Advances in Magnetic Devices for
Computers, D. H. Looney. Journal of Applied
Physics, vol. 30, April 1959, pp. 38s-42s.

Figure 4 - Pulse Characteristics of Wide
Temperature Range Core as a Function
of Drive.

5. Millimicrosecond Magnetic Switching and
Storage Element, D. A. Meier, Journal of
Applied Physics, vol. 30, April 1959,
pp. 45s-46s.

Figure 5 - Array Output at 100°C. Vertical
Scale
50 mv/div. Horizontal
Scale = 0.5 ~s/div.

6. Square Loop Ferrites With Temperature Independent Properties and Improved Disturb Ratio,
R. S. Weisz. Accepted for publication in
Journal of Applied Physics.

Figure 6 - Array Output at -55°C. Vertical
Scale
50 ~v/div. Horizontal
Scale = 0.5 ~s/div.

211
4.4

Figure 1.

HYSTERESIS LOOPS OF STANDARD O:NE NICROSECOND MEMORY CORE
AT -55°, f250J AND flOOD C.
Vertical Scale = 680 gauss/dive
Horizontal Scale = 0.78 oe/dive

Figure 2.

HYSTERESIS LOOPS OF WIDE TEMPERATURE RANGE CORE
AT -55°C J f250C J JLWD flOOoC.
Vertical Scale = 940 gauss/dive
Horizontal Scale = 1.4 oe/div.

.

.t:.N

......

.t:.N

~

~

~

~

~

I

0
~
(\J

0
~

-:

Q

0

0

Q

~

~

0

~

~

0

"":

f'.:

0

~

~

0

~

~
~

,-:.

en

uV,

00

I"-

-RO

W

0

10

-tp

0

"'"

~

0

~

0

-;

Q

0

0

t()

(\J

L-----

-dVz

1

-50

-30

-10

+10

+30

+50

+70

+90

+110

TEMP. (OC)
Figure

:3.

PULSE CHARACTERISTICS OF WIDE TEl:4PERATURE RANGE CORE AS A FUNCTION OF TEMPERATURE
Constant Drive of 1.0 Ampere Turn

fd

0

N

01=>0

W

~

214
4.4

Figure

5. ARRAY OUTPUT AT 100°C.
Vertical Scale = 50 mY/dive
Horizontal Scale = 0.5 j..t s/div.

Figure

6. ARRAY OUTPUT AT -55°c.
Vertical Scale
Horizontal Scale

= 50

mY/dive

= 0.5 fJ. s/ di v •

215
5.1

DESCRIPTIVE LANGUAGES AND PROBLEM SOLVING
Marvin Minsky
Dept. of Mathematics and Computation Center
Massachusetts Institute of Technology

Advances in machine problem solving
may depend on use of internal languages
for description and abstraction of the
outcomes of experiments. As more complex
problems are attempted there will have
to be less trial and error and more .
systematic analysis of the results of
each trial. Learning on the basis of
experience will require a phase of
refinement in which the machine will
attempt, by analysis and inductive inference, to get as much as possible from
each experiment.
Introduction
Work on artificial intelligence is
proceeding at a slow, apparently steady,
rate. The complexity of problems being
attacked is growing slowly, as is the
complexity of the successful programs
themselves. In the past it seems to
have taken two or three years for each
significant advance and one may ask why
progress is so slow. Much of this time
has been spent on the development of
programming languages and systems
suitable for the symbol manipulation
processes involved. But much of the
difficulty has been conceptual also.
The methods which worked quite well on
easy problems did ,not extend smoothly
to the difficult ones. Continued progress
will require implementation of new ideas,
for there are some very tough problems
in our immediate path. It seems to us
that solution of these problems will
require the use of non-trivial formal
and descriptive language systems. These
are only beginning to appear as a working
part of the problem solVing machinery
and it will take much ingenuity to
bring current notions into usable i"Ol'm.
The two papers of this session
represent important, and very different,
phases in the development of machineusable language systems. In one case
we have a system which can process, to
find the meaning with respect to a small
universe, expressions in a very lifelike fragment of ordinary language.
In the second paper we find an
ambitious attempt at the beginnings of a
Theory of Computation, based in part on
the use of a symbol manipulation language
suited at once for both theoretical
analysis and for practical programming use.

Our purpose here is to indicate a
few of the considerations that seem to
point toward the incorporation of
complex linguistic processes into the
next generation of heuristic programs.
Some of these difficulties have arisen
in the author's work, jOintly with 1
McCarthy, on the Advice-Taker system
In a recent paper 2 the author discussed
the principles and mechanisms of a
variety of problem solVing systems, but
did not dwell on the question of
extending these to really complex problems.
We assume the terminology of that paper.
When one attempts to apply the
teChniques aescribed there one discovers
that
1. The search problems become very
serious. One is faced not only with
greatly enlarged problem trees but also
with a greater variety of plausible
methods.
2. The problem of learning from
experience becomes qualitatively more
difficult. To learn the lesson of a
complex experience requires shrewd,
deliberate, analysis that ca~not be
approximated by any of the simple
learning models based on averaging or on
correlation.

3. The classification and pattern
recognition methods must be on a descriptlve level. Again, correlation or
matching methods must be replaced by
more sophisticated symbol-manipulation
processes.
4. Planning methods, Character
and Difference algebras, etc., threaten
to collapse when the fixed sets of categories adequate for simple problems have
to be replaced by the expressions of
a descriptive language. The use of lookup tables for choosing methods 't'lill have
to be supplemented by something more
like reasoning.
When we call for the use of
"reasoning we intend no suggestion of
giving up the game by invoking an
intelligent subroutine. The program
that administers the search will be just
another heuristic program. Almost
certainly it will be composed largely of
the same sorts of objects and processes
that will comprise the subject-domain
ll

216
5.1

programs. Almost certainly it will be
recursively applied to itself so that
the system can be finite. But it does
seem clear that the basic (non-recursive)
part of the structure will have to be
more complex than is any current system.
The Need for Analysis
The simplest problems, e.g., playing
tic-tac-toe or proving the very simplest
theorems of logic, can be solved by
simple recursive application of all the
available transformations to all the
situations that occur, dealing with subproblems in the order of their generation. This becomes impractical in more
complex problems as the search space
grows larger and each trial becomes more
expensive in time and effort. One can
no longer afford a policy of simply
leaving one unsuccessful attempt to go
on to another. For each attempt on a
difficult problem will involve so much
effort that one must be quite sure that,
whatever the outcome, the effort will
not be wasted entirely. One must become selective to the point that no trial
is made without a compelling reason;
Just as in any research, expensive
experiments must be carefully designed.
One must do a good deal of criticism and
analysis between experiments so that
each will be a critical test of a
significant portion of the search space.
The ability to solve a difficult
problem hinges on the ability to split
or transform it into problems of a lower
order of difficulty. To do this, without total reliance on luck, requires some
understanding of the situation. One must
be able to deduce, or guess, enough of
the consequences of the problem statement
to be able to set up simpler models of
the problem situation. The models must
have enough structure to make it likely
that there will be a v'lay to extend their
solutions to the original problem.
The construction of less difficult
subproblems will be useful, by definition,
only if one has already a very good chance
of solving them efficiently. Otherwise
the search tree will grow beyond bounds.
This means we must have already built
up adequate solution methods for the
lower order problems, e.g., as a set of
more or less packaged subroutines. This
entails soine formidable requirements:
Training Sequences
The machine is presumed to have
acquired its good subroutines through
earlier solution of less complex
problems. (We are not interested here

in the case in which these methods are
provided at the start.) Thus the machine
must have been exposed to a graded sequence
of problems. To be sure, given timelimits, a machine will select a graded
subsequence from an unorganized variety
of problems. But a careful arrangement
will be necessary to insure that methods
learned in the problems that the machine
does manage to solve will be useful on
more difficult problems met later. In
any case one cannot rely on making large
Jumps, either in machines or in humans.
Refinement Phase
Solving simpler problems is not
enough. To make progress one needs also
to package" the successful method for
effective later use. We are not interested in the trivial case of recognizing
a problem once before solved, though this
can be difficult enough when there is
some disguise. The success must be
generalized to cover a substantial
variety of situations. To do this it
would seem that there should be a phase
of exploration and consolidation in
which the successful method is refined-its central innovation (if any) isolated
and packaged, in terms as general as
possible., One must explore its range of
application and construct an expression
describing this range. This may involve
inventing similar problems on which the
method, or close variant, works; then
constructing a plausible generalization.
Il

Certainly people must go through
such phases. One cannot usually solve
hard problems with once-used but still
unfamiliar methods. One must first
"understand" the methods quite well; this
means becoming able to recognize situations in which they are applicable. It
is probably misleading to think of this
as 'Ipractice ll --acquisition of facility
through repetition. Exercise in, e.g.,
mathematical technique is probably very
different from exercise in weight-lifting.
Its effect is not so much in reinforcing
methods, or paths already weakly laid
down, but is rather to provide the necessary data for some Inductive Inference
technique. The latter will replace the
special method by one of somewhat greater
generality.
Failure of the refinement phase to
yield a precise, abstractly stated conclusion can be concealed to a pOint.
One often encounters mathematical
situations in which one can answer
particular questions quickly, yet is
unable to state a satisfactory formal
generalization. This can happen through
the assembly of a set of different models

217
5.1

or examples which, as a group, show most
or all of the features of the unformulated
general theorem. One can answer some
question in the negative, by finding
inconsistency with an example. Consistency with all leads one to the affirmative. Often the examples themselves
are not formulated clearly, or completely
consciously. In such cases one will
find some statements seem "obvious" yet
(because of the incomplete understanding
which precludes giving any precise
explanation) are also felt to be
"intuitive." An incomplete formalization
or conceptualization, e.g., such a set
of examples, can be very powerful when
used at or near the top level. But if
not understood or packaged" it could
become a serious nuisance later when,
because of its informality, it cannot
be used in deduction or in the construction of further abstractions.
l1

Coding and Retrieval Problems
The compact representation of
results of previous experience requires
an adequate descriptive language. This
language must permit general statements
about both problem-domain matters and
about the problem-solving methods. It
must permit logical deductions to be
made. This raises several problems.
One problem that has been a great
nuisance to us arises in connection with
non-mathematical problems in which actions
affect the state of some subject domain.
Thus a move affects the positions
of pieces in a board game. When this
happens, some statements formerly deduced
about the situation cease to be true.
(In a mathematical domain a theorem, once
proved, remains true when one proves
other theoremsl) One must then deduce
all the consequences of an action in so
far as it affects propositions that one
is planning to use. This might be done
through some heuristic technique which
can assess relevancy, or it could be done
through a logic which takes sU,ch consequences into account. The trouble
with the latter is that the antecedents
of all the propositions must contain a
condition about the state of the system,
and for complex systems this becomes
overwhelmingly cumbersome. Other
systematic solutions to the problem seem
about equally repellent. It is a problem
that seems urgently to require a heuristic
solution.
Our present proposal on this matter
is to make the system plan ahead. Whenever an important deduction is made, the
system is to try to discover which kinds
of actions could affect its validity.

Independent monitors are then set up to
detect when such actions are proposed.
The normal problem solving exploration
process proceeds independently of these
monitors, and is interrupted when one
of them detects a threat to the proposition it is defending. This model has
a certain introspectively attractive
character; it suggests a free conscious
exploration with more or less subconscious
trouble-detectors. Unfortunately, its
essentially parallel nature threatens
to make its use in serial computer programming rather expensive. We hope
someone will come up with a better idea.
In any case, the retrieval problem
has to be faced. The problem of making
useful deductions from a large body of
statements (e.g., about the relevance
of different methods to different kinds
of problems) raises a new search problem.
One must restrict the logical exploration
to data likely to be relevant to the
current problem. This selection function
could hardly be completely built-in at
the start. It must develop along with
other data accumulated by experience.
Another rather serious problem
centers around the problem of abbreviations,
or proper names. The language must be
used together with an abbreviative
technique so that the most useful notions
can be deSignated by reasonably convenient
(short) representations. This is not
only a matter of convenience and compactness; it is a more or less inescapable
requirement of known inductive inference
techniques and thus requisite for formation of hypotheses or generalizations.
Unfortunately an abbreviation cannot
show all of the structure of the longer
expression it designates. This seriously
limits the possibilities of making formal
logical deductions. Ultimately the
machines will have to use mnemonic codings
in their internal languages, just as we
need to do this when we use their external
languages.
The systematic solution to the
abbreviation problem is, again, to revise
the whole body of propositions in current
use, in so far as they are going to be
used in the same deductive operations.
All the alternatives to this that we
can envision are of somewhat stopgap
nature. We content ourselves with the
observation that it is equally a major
problem for humans to make substantial
changes in basic abstractions, ways of
classifying or perceiVing, and the like.
Once one has built up a structure depending
on a certain conceptual commitment, he
v'lill stave off a revision of its foundation as though the cost of changing

218
5.1

it were high. Otherwise, perhaps, people
would not argue so much. One may view
that phenomenon, if one likes, as a matter
of ego involvement. But it would be well
to remember that being wrong (and having
to change) has a real intellectual cost,
and not merely a social cost.
In any case, our ideas on this
subject are not yet in presentable
condition.
Conclusion
The need to be able to make abstractions in symbolic language is
already urgent in current attempts to
make machines prove theorems, 'play games,
etc. There are some very difficult
problems to be faced in this area. We
still maintain, with Mccarthy, that !lin
order for a program to be capable of --learning something1it must first be capable
of being told it."
Results on "selforganizing systems " without explicit
provision for such abilities show very
little promise to date, and systematic
attempts in the direction of internal
language processing should be promoted.
References
1.

J. McCarthy, Programs with Common
Sense," r,1echanization of Thought
Processes, H.M.S.O., London, 1959,
pp. 75-84 (Vol. I).

2.

M. L. Ivlinsky, 11 Steps TO\''lard
Artificial Intelligence, ?roc.
IRE, Vol. 49, no. 1, Jan. 1961,
pp. 8-30.

If

11

219
5.2

BASEBALL:

AN

AUTO~fATIC

QUESTION-ANSWERER

Bert F. Green, Jr., Alice K. Holf, Carol Chomsky, and Kenneth Laughery
Lincoln Laboratory*, Massachusetts Institute of Technology
Lexington 73, l~ssachusetts
Surmnary
Baseball is a computer program that answers
questions phrased in ordinary English about stored
data. The program reads the question from punched
cards. After the words and idioms are looked up
in a dictionary, the phrase structure and other
syntactic facts are determined for a content
analysis, which lists attribute-value pairs
specifying the information given and the information requested. The requested information is
then extracted from the data matching the specifications, and a~v necessary processing is done.
Finally, the answer is printed. The program's
present context is baseball games ; it ans\"ers
such questions as "Hhere did each team play on
July 7?"
Introduction
Men typically communicate vli th computers in
a variety of artificial, stylized, unambiguous
languages that are better adapted to the machine
than to the man. For convenience and speed,
many future computer-centered systems vTill
reqUire men to communicate with computers in
natural language. The business executive, the
military commander, and the scientist need to
ask ~uestions of the computer in ordinary English,
and to have the computer answer questions
directly. Baseball is a first step toward this
goal.
Baseball is a computer program that answers
questions posed in ordinary English about data
in its store. The program consists of two parts.
The linguistic part reads the question from a
punched card, analyzes it syntactically, and
determines what information is given about the
data being requested. The processor searches
through the data for the appropriate information,
processes the results of the search, and prints
the answer.
The program is written in IPL-Vl , an information processing language that uses lists, and
hierarchies of lists, called list structures, to
represent information. Both the data and the
dictionary are list structures, in which items
of information are expressed as attribute-value
pairs, e.g., Team = Red Sox.
*Operated with support from the U.S. Army, Navy,
and Air Force.

The program operates in the context of
baseball data. At present, the data are the month,
day, place, teams and scores for each game in the
American League for one year. In this limited
context, a small vocabulary is sufficient, the
data are simple, and the SUbject-matter is
familiar.
Some temporary restrictions ,,,ere placed on
the input questions so that the initial program
could be relatively straightforward. QQestions
are lliaited to a single clause; by prohibiting
structures with dependent clauses the syntactic
analysis is considerably simplified. Logical
connectives, such as and, or, and not, are prohibited, as are constructions implying relations
like most and highest. Finally, questions
involving sequential facts, such as "Did the
Red Sox ever win six games in a row?" are prohibited. These restrictions are temporary
expedients that will be removed in later
versions of the program. ~~reover, they do not
seriously reduce the number of questions that
the program is capable of ansvlering. From simple
questions such as "Who did the Red Sox lose to
on July 5? 11 to complex q,uestions such as "Did
every team play at least once in each park in
each month?" lies a vast number of answerable
questions.
Specification List
Fundamental to the operation of the baseball
program is the concept of the specification list,
or spec list. Tnis list can be viewed as a
canonical expression for the meaning of the
question; it represents the information contained
in the question in the form of attribute-value
pairs, e.g., Team = Red Sox. The spec list is
generated from the question by the linguistic
part of the program, and it governs the operation
of the processor. For example, the question
-"'{.here did the Red Sox play on July 7? 11 has the
spec list:
Place
Team
Month
Day

=?
= Red Sox

= July

=7

Some questions cannot be expressed solely
in terms of the main attributes (Month, Day, Place,
Team, Score and Game Serial Nillmber), but require
some modificatjon of these attributes. For
example, on the spec list of ";{.hat teams vTon 10

220
5.2

games in July?", the attribute Team is modified by
and Game is modified by Number of, yielding

~{inning,

Month

Month = July
Place = Boston
=7
Game Serial No. = 96
(Team = Red Sox, Score
(Team = Yankees, Score

Day

Team(winning) = ?
Game (number of)

values for each of six attributes, e.g.:

10

= July

5)
3)

Dictionary

The parentheses indicate that each Team must be
associated "lith its own score, "ifhich is done by
placing the~ together on a sublist.

The dictionary definitions, which are
expressed as attribute-value pairs, are used by
the linguistic part of the program in generating
the spec list. A complete definition for a ifOrd
or idiom includes a part of speech, for use in
determining phrase structure; a meaning, for use
in analyzing content; an indication of whether
the entry is a question-word, e. g., 'ifho or how
manyj and an indication of whether a-wDrd oCCUrs
as part of any stored idiom. Separate dictionaries are kept for words and idioms, an idiom
being any contiguous set of words that functions
as a unit, having a unique definition.

The processing routines are written to
accept any organization of the data. In fact,
they will accept a non-parallel organization in
which, for example, the data might be as above
for all games through July 31, and then organized
by place, ,nth month under place, for the rest
of the season. The processing routines will also
accept a one-level structure in which each game
is a list of all attribute-value pairs for that
game. The possibility of hierarchical organization
was included for generality and potential
efficiency.

The meaning of a word can take one of
several forms. It may be a main or derived
attribute with an associated value. For example,
the meaning of the word Team is Team = (blank),
the meaning of Red Sox is Team = Red Sox, and
the meaning of who is Team = ? The meaning may
designate a subroutine, together with a particular
value, as in the case of modifiers such as
'ifinning, ~, ~ or how many. For example,
winning has the meaning Subroutine Al = ITinning.
The subroutine, which is executed by the content
analysis, attaches the modifier 'dinning to the
attribute of the appropriate noun. Some 'ifOrds
have more than one meaning; the v,ord Boston
may mean either Place = Boston or Team = Red Sox.
The dictionary entry for such words contains, in
addition to each meaning, the designation of a
subroutine that selects the appropriate meaning
according to the context in which the word is
encounted. Finally, some ifOrds such as the, did,
play, etc., have no meaning.
-- -Data

The data are organized in a hierarchical
structure, like an outline, with each level
containing one or more items of information.
Relationships among items are expressed by their
occurrence on the same list, or on associated
lists. The main heading, or highest level of the
structure, is the attribute ~~nth. For each ~onth,
the data are further subdivided by place. Below
each place under each month is a list of all
games played at that place during that month.
The complete set of items for one game is found
by traCing one path through the hierarchy, i.e.
one list at each level. Each path contains

Details of the Program
The program is organized into several
successive, essentially independent routines,
each operating on the output of its predecessor
and producing an input for the routine that
follows. The linguistic routines include
question read-in, dictionary look-up, syntactic
analysis, and content analysis. The processing
routines include the processor and the responder.
Linguistic Routines
Question Read-in. A question for the program
is read into the computer from punched cards.
The question is formed into a sequential list of
words.
Dictionary Look-up. Each "ord on the
question list is looked up in the word dictionary
and its definition copied. Any undefined words
are printed out. (In the future, 'ifi th a directentry keyboard, the computer can ask the questioner to define the unknovrn words in terms of
,fOrds that it knows, and so augment its vocabulary.) The list is scanned for possible idioms;
any contiguous words that form an idiom are replaced by a single entry on the question list,
and an associated definition from the idiom
dictionary. At this pOint, each entry on the list
has associated with it a definition, including a
part of speech, a meaning, and perhaps other
indicators.
Syntax. The syntactiC analysis is based on
the parts of speech, which are syntactic categories assigned to words for use by the syntax

22l

5.2

routine. There are 14 parts of speech and
several ambiguity markers.
First, the question is scanned for ambiguities in part of speech, which are resolved in
some cases by looking at the adjoining vTords, and
in other cases by inspecting the entire question.
For example, the w·ord ~ may be either a noun
or a verb; our rule is that, if there is no other
main verb in the question, then score is a verb,
otherwise it is a noun.
Next, the syntactic routine locates and
brackets the noun phrases, [1 , and the preposi tional and adverbial phrases, ( ). The verb is
left unbracketed. This routine is patterned
after the work of Harris and his associates at
the University of Pennsylv~nia.2 Bracketing
proceeds from the end of the question to the
beginning. Noun phrases, for example, are
bracketed in the following manner: certain parts
of speech indicate the end of a noun phrase;
v1ithin a noun phrase, a part of speech ma;>7 indicate that the '\{ord is wi thin the phrase, or that
the word starts the phrase, or that the word is
not in the phrase, which means that the previous
word started the phrase. Prepositional phrases
consist of a preposition immediately preceding a
noun phrase. The entire sequence, preposition
and noun phrase, is enclosed in prepositional
brackets. p~ example of a bracketed question is
shown be1m{:
[ROvl many games] did

~he Yankees) play (in rJu1y1)?
limen the question has been bracketed, any unbracketed preposition is attached to the first
noun phrase in the sentence, and prepositional
brackets added. For example, 1I1ilho did the Red
Sox lose to on July 5?" becomes "(To [who] ) did
{ the Red Sox] lose (on [July 5 j )?"
Follov1ing the phrase analysis, the syntax
routine determines whether the verb is active or
passive and locates its subject and object.
Specifically, the verb is· passive if and only if
the last verb element in the question is a main
verb and the preceding verb element is some form
of the verb to be. For questions ,nth active
verbS, if a free noun phrase (one not enclosed in
prepositional brackets) is found between two verb
elements, it is marked Subject, and the first free
noun phrase in the question is marked Object.
otherwise the first free noun phrase i~
subject, the next, if any, is the object. For
passive verbs, the first free noun phrase is
marked Object (since it is the object in the
active form of the question) and all prepositional
phrases with the preposition by have the noun
phrase "inthin them marked Subject. If there is
more than one, the content analysis later chooses

among them on the basis of meaning.
Finally, the syntactic analysis checks to
see if any of the words is marked as a question
word. If not, a signal is set to indicate that
the question requires a yes/no answer.
Content Analysis. The content analysis uses
the dictionary meanings and the results of the
s;>~tactic analysis to set up a specification list
for the processing program. First any subroutine
found in the meaning of any word or idiom in the
~uestion is executed.
The subroutines are of two
basic types; those that deal with the meaning of
the '\{ord itself and those that in some way change
the meaning of another word. The first chooses
the appropriate meaning for a '\-Tord with multiple
meanings, as, for example, the subroutine mentioned above that deCides, for names of cities,
"ifhether the meaning is Team = At or Place = Ap.
The second type alters or modifies the attribute
or value of an appropriate syntactically related
'\-Tord. For example, one such subroutine puts its
value in place of the value of the main no~n in
its phrase. Thus Team = (blank) in the phrase
each team becomes Team = each; in the phrase what
team, it becomes Team =? Another subroutin-e--modifies the attribute of a main noun. Thus
Team = (blank) in the phrase innning team becomes
Team ( ,nnning) = (blank). In the question "1';Jho
beat the YanKees on July 4?11, this subroutine,
found in the meaning of beat, modifies the
attribute of the subject and ~bject, so that
Team = ? and Team = Yankees are rendered
Team(winning) = ? and Team(losing) = Yankees.
.Another subroutine combines these tvo operations:
it both modifies the attribute and changes the
value of the main noun. Thus, Game = (blank) in
the phrase six games becomes Game(number ~f) = 6,
and in the phrase hov many games becomes
Game(number of) = ?
After the subroutines have been executed,
the question is scanned to consolidate those
attribute-value pairs that must be represented on
the specification list as a single entry. For
example, in lI'Vlho was the winning team ••• Team = ?
and Team(winningJ = (blank) must be collapsed into
Team(winning) =.. Next, successive scans will
create any sub1ists implied by the syntactic
structure of the question. Finally, the composite
information for each phrase is entered onto the
spec list. Depending on its complexity, each
phrase furnishes one or more entries for the list.
The resulting spec list is printed in outline
form, to provide the questioner with some intermediate feedback.
11

Processing Routine
Processor. The specification list indicates
to the processor what part of the stored data is
relevant for answering the input question. The

222
5.2

processor extracts the matching information from
the data and produces, for the responder, the
answ'er to the question in the form of a list
structure.
The core of the processor is a search routine
that attempts to find a match, on each path of a
given data structure, for all the attribute-value
pairs on the spec list; when a match for the whole
spec list is found on a given path, these pairs
relevant to the spec list are entered on a found
list. A particular spec list pair is considered
matched when its attribute has oeen found on a
data path and, either the data value is the same
as the spec value, or the spec value is ? or each,
in which case any value of the particular attrroute
is a match. !',1ltching is not ablays straightfor,yard. Derived attributes and some modified
attributes are functions of a nuraber of attributes
on a -~th and must be computed before the values
can be matched. For example, if the spec entry
is Home Terun = Red Sox, the actual home team for
a particular path must be computed from the
place and teruns on that path before the spec
value Red Sox can be matched "Ti th the computed
data value. Sublists also require special
handling because the entries on the sublist must
sometimes be considered separately and sometimes
as a unit in various permutations.
The found list produced by the search routine
is a hierarchical list structure containing one
main or derived attribute on each level of each
path. Each path on the found list represents the
information extracted from one or more paths of
the data. For example, for the question "lihere
did each team play in July?", a single path
eXists, on the found list, for each team which
played in July. On the level below each team,
all places in ,{hich that team played in July
occur on a list that is the value of the attribute
Place. Each path on the found list may thus
represent a condensation of the information
existing on many paths of the search data.
Many input questions contain only one query,
as in the question above, i.e., Place =?
These
questions are ansvTered, with no further processing,
by the found list produced by one execution of
the search routine. Others require simple processing on all occurrences of the queried attribute on the generated found list. The question
"In how many places did each team play in July?"
requires a count of the places for each team,
after the search routine has generated the list
of places for each team.
Other questions imply more than one search
as well as additional processing. For a spec
attribute with the value every, a comparison with
a list of all possible values for that attribute
must be made after the search routine has
generated lists of found values for that attribute.

Then, since only those found list paths for which
all possible values of the attribute exist should
remain on the found list as the anS"Ter to the
question, the search routine, operating on this
found list as the data, is again executed. It
now generates a new found list containing all the
data paths for which all possible values of the
attribute vTere found. Likewise, questions
involving a specified number, such as 4 teams,
imply a search for which teruns, a count of the
teams found on each path, and a search of the
found list for paths containing 4 teams.
In general, a question may contain implicit
or explicit tlueries. Since these queries must
be ansvTered one at a time, several searches,
ivith intermediate processing, are required. The
first search operates on the stored data while
successive searches operate on the found list
generated by the preceding search operation.
As an example, consider the question liOn how
many days in July did eight teams play?" The
spec list is
Day (number of) =?;
MOnth
= July;
Team (number of) = 8 .
On the first pass, the implicit question which
teams is answered. The spec list for the-rirSt
search is

Day
= Each;
Month = July;
Team =? •
The found data is a list of days in July j for
each day there is a list of teruns that played on
that date. Follovdng this search, the processor
counts the teams for each day and associates the
count '\vith the attribute Team. On the second
search, the spec list is
Day
=?
Month
= July;
Team (number of) = 8 .
The found data is a list of days in July on which
eight teruns played. After this pass, the processor counts the days, addsthe count to the
found list and is finished.

Responder. No attempt has yet been made to
respond in grammatical English sentences. Instead,
the final found list is pri~t~d, in outline form.
For questions requiring a yes/no answer, YES is
printed along with the found list. If the search
routine found no matching data, NO is printed for
yes/no questions, and NO DATA for all other cases.

223
5.2

Discussion
The differences bet'YTeen Baseball and both
automatic language translation and information
retrieval should nO"l be evident. The linguistic
part of the baseball program has as its main goal
the understanding of the meaning of the g,uestion
as embodied in the canonical specification list.
Syntax must be considered and ambiguities resolved
in order to represent the meaning adequately.
Translation programs have a different goal: transforming the input passage from one natural language
to another. Meanings must be considered and
a.t"Ubiguities resolved to the e::tent that they
effect the correctness of the final translation.
In general, translation prograras are concerned
more ,Yith syntax and less "lith meaning than the
Baseball program.
Baseball differs from most retrieval systems
in the nature of its data. Generally the retrieval problem is to locate relevant doc~~ents.
Each docurnent has an associated set of index
nwnbers describing its content. The retrieval
system must find the appropriate index numbers
for each input request and then search for all
documents bearing those index numbers. The basic
problem in such systems is the assignment of index·
categories. In Baseball, on the other hand, the
attributes of the data are very 'tTell specified.
There is no confusion about them. HOHever,
Baseball's derived attributes and modifiers imply
a great deal more data processing than most
document retrieval prog~~. (Baseball does bear
a close relation 'Ylith the ACSI-r.IATIC system
discussed by Miller et al at the 1960 i·lestern
Joint Computer Conference.3)
The concept of the spec list can be used to
define the class of questions that the baseball
program can answer. It can answer all questions
whose spec list consists of attribute-value pairs
that the program recognizes. The attributes may
be modified or derived, and the values may be
definite or queries. Any combination of attributevalue pairs constitutes a specification list.
~i3.ny will be nonsense, but all can be ans'YTered.
The nmnber of questions in the class is, of
course, infinite, because of the numerical values.
But even if all numbers are restricted to two
digits, the program can answer millions of meaningful ~uestions.
The present program, despite its restrictions,
is a very useful communication device. Any
complex question that does not meet the restrictions can al'vays be broken up into several simpler
questions. The program usually rejects questions
it cannot handle, in which case the questioner
may rephrase his question. He can also check
the printed spec list to see if the computer is
on the right track, in case the linguistic program
has erred and failed to detect its O,in error.

Finally, he can often judge whether the answer
is reasonable.
Next Steps
No important difficulty is expected in
augnenting the prograrn to include logical
connectives, negatives, and relation words. The
inclusion of multiple-clause questions also seems
fairly straightfo~~rd, if the questioner will
mark off for the comuuter the boundaries of his
clauses: The progra,;.1 can then deal "Ti th the
subordinate clauses one at a time before it deals
\-lith the main clause, using existing routines.
On the other hand, if the syntax analysis is
req.uired to determine the clause boundaries as
well as the phrase structure, a much more
sophisticated program ,fould be required.
The problem of recognizing and resolving
semantic arabiguities remains largely unsolved.
Determining what is meant by the question "Did
the Red Sox 'rin most of their ga.t'1les in July?"
depends on a much larger context than the
immediate question. The computer might answer
all meaningful versions of the g,uestion (vTe know
of five), or might ask the questioner which
meaning he intended. In general, the facility
for the computer to query the questioner is
likely to be the most pmferful improvement.
This would allo'., the computer to increase its
vocabulary, to resolve ambiguities, and perhaps
even to train the questioner in the use of the
program.
Considerable pains ,{ere taken to keep the
program general. Host of the program 'YTill rema.in
unchanged and intact in a ne,{ context, such as
voting records. The processing program ,{ill
handle data in an;~r sort of hierarchical form, and
is indifferent to the attributes used. The syntax
program is based entirely on parts of speech,
which can easily be assigned to a new set of ,fords
for a new context. On the other hand, some of the
subroutines contained in the dictionary meanings
are certainly specific to baseball; probably each
new context would require certain subroutines
specific to it. Also, each context might introduce a number of modifiers and derived attributes
that would have to be defined in terms of special
subroutines for the processor. Hopefully, all
such occasions for change have been isolated in a.
small area of special subroutines, so that the
main routines can be unaltered. However, until
l..,e have actually s,. . i tched contexts, we cannot say
definitively that 'Yle have been successful in
producing a general question-answering program.
Acknowledgment
The Baseball program was conceived by
Fredrick C. Frick, Oliver G. Selfridge, and
Gerald P. Dineen, whose continued guidance is

224

5.2

gratefully

acknowledged~

References
1.

A. New"ell and F. Tonge, HAn introduction to
lni'ormation Processing Language yll, Commun.
Assoc. for Computing Mach., Vol. 3, pp. 205211; April, 1960.

2.

See the project summary, by Z. S. Harris, in
Current Research and Development in Scientific
Documentation No.6, pp. 52-53, Nat'l
Sciences Foundation, May, 1960.

3.

L. r.fi.ller, J. Minker, ~i. G. Reed, and 1,1. E.
Shindle, "A multi-level file structure for
ini'ormation processing", Proceedings \-lestern
Joint Computer Conference, Yolo 17, pp. 53-

59,

M3.Y,

1960.

225
5.3
A BASIS FOR A MATHEMATICAL THEORY OF COMPUTATION, PRELIMINARY REPORT
John McCarthy
M.I.T. Computation Center
Cambridge, Massachusetts
Abstract: Programs that learn to modify their
cwn behaviors require a way of representing
algorithms so that interesting properties and interesting transformations of algorithms are simply
represented. Theories of computability have been
based on Turing machines, recursive functions of
integers and COHlputer proGrams. Each of these has
artificialities which make it difficult to manipulate algorithms or to prove things about them.
The present paper presents a formalism based on
conditional forms and recursive functions whereby
the functions compEtable in terlLS of certain base
functions can be simply expressed. We also describe some of the formal properties of conditional
101 ms, and a method called recursion induction for
proving facts about algorithms.
1

Computation is sure to become one of the
most important of the sciences. This is because
it is the science of how machines can be made to
carry out intellectual processes. We know that
any intellectual process that can be carried out
mechanically can be performed by a general purpose digital computer. Moreover, the limitations
on what we have been able to make computers do so
far seem to come far more from our weakness as
programmers than from the intrinsic limitations
of the machines. We hope that these limitations
can be greatly reduced by developing a mathematical science of computation.
There are three established directions of
mathematical research relevant to a science of
computation. The first and oldest of these is
numerical analysis. Unfortunately, its subject
matter is too narrow to be of much help in
forming a general theory, and it has only recently begun to be affected by the existence of
automatic computation.
The second relevant direction of research
is the theory of computability as a branch of
recursive function theory. The results of the
basic work in this theory including the existence
of universal machines and the existence of unsolvable problems have established a framework
in which any theory of co~putation must fit.
Unfortunately, the general trend of research in
this field has been to establish more and better
unsolvability theorems, and there has been very
little attention paid to positive results and
none to establishing the properties of the kinds
of algorithms that are actually used. Perhaps
for this reason the formalisms for describing
algorithms are too cumbersome to be used to describe actual algorithms.
The third direction of mathematical research
is the theory of finite automata. Results which
use the finiteness of the number of states tend
not to be very useful in dealing with present
computers which have so many states that it is
impossible for them to go through a substantial
fraction of them in a reasonable time.

The present paper is an attempt to create a
basis for a mathematical theory of computation.
Before mentioning what is in the paper, we shall
discuss briefly what practical results can be
hoped for from a suitable mathematical theory.
This paper contains direct contributions towards
only a few of the goals to be mentioned, but we
list additional goals in order to encourage a
gold rush.
1. To develop a universal programming
language. We believe that this goal has been
written off prematurely by a number of people.
Our opinion of the present situation is that
ALGOL is on the right track but mainly lacks the
ability to describe different kinds of data,
that COBOL is a step up a blind alley on account
of its orientation towards English which is not
well suited to the formal description of procedures, and that UNCOL is an exercise in group
wishful thinking. The formalism for describing
computations in this paper is not presented as
a candidate for a universal programming language
because it lacks a number of features, mainly
syntactic, which are necessary for convenient
use.
2. To define a theory of the equivalence
of computation processes. With such a theory
we can define equivalence preserving transformations. Such transformations can be used to
take an algorithm from a form in which it is
easily seen to give the right answers to an
equivalent form guaranteed to-give the same answers but which has other advantages such as
speed, economy of storage, or the incorporation
of auxiliary processes.
4. To represent algorithms by symbolic expressions in such a way that significant ,changes
in the behavior represented by the algorithms
are represented by simple changes in the symbolic
expressions. Programs that are supposed to learn
from experience change their behavior by changing the contents of the registers that represent
the modifiable aspects of their behavior. From
a certain point of view having a conv~nient
representation of one's behavior available for
modification is what is meant by consciousness.
5. To represent computers as well as
computations in a formalism that permits a
treatment of the relation between a computation
and the computer that carries out the
computation.
6. To give a quantitative theory of computation. There ought to be a quantitative
measure of the size of a computation analogous
to Shannon's measure of information. The
present paper contains no information about this.
The present paper is divided into two
sections. The first contains several descriptive formalisms with a few examples of their
use, and the second contains what little theory
we have that enables us to prove the equivalence

226

5.3
of computations expressed in these formalisms.
The formalisms treated are the following:
1. A way of describing the functions that
are computable in terms of given base functions
using conditional expressions and recursive function definitions. This formalism differs from
those of recursive function theory in that it is
not based on the integers or any other fixed
domain.
2. Computable functionals, i.e. functions
with functions as arguments.
3. Non-computable functions. By adjoining
quantifiers to the computable function formalism,
we obtain a wider class of functions which are
not ~ priori computable. However, such functions
can often be shown to be equivalent to computable
functions. In fact, the mathematics of computation may have as one of its major aspects rules
which permit us to transform functions from a
non-computable form into a computable form.
4. Ambiguous functions. Functions whose
values are incompletely specified may be useful
in proving facts about functions where certain
details are irrelevant to the statement being
proved.
5. A way of defining new data spaces in
terms of given base spaces and of defining functions on the new spaces in terms of functions on
the base spaces. Lack of such a formalism is one
of the main weaknesses of ALGOL, but the business
data processing languages such as FLOWMATIC and
COBOL have made a start in this direction, even
though this start is hampered by concessions to
the presumed prejudices of business men.
The second part of the paper contains a few
mathematical results about the properties of the
formalisms introduced in the first part. Specifically, we describe the following:
1. The formal properties of conditional
expressions.
2. A method called recursion induction for
proving the equivalence of recursively defined
functions.
3. Some relations between the formalisms
introduced in this paper and other formalisms
current in recursive function theory and in
programming.
We hope that the reader will not be angry
about the contrast between the great expectations
of a mathematical theory of computation and the
meager results presented in this paper.
FORMALISMS FOR DESCRIBING COMPUTABLE FUNCTIONS
AND RELATED ENTITIES

1.

Functions Computable in Terms of Given Base
Functions

Suppose we are given a base collection ~
of functions having certain domains and ranges.
In the case of the non-negative integers, we may
have the successor function and the predicate of
equality, and in the case of the S-expressions
discussed in (7.), we have the five basic operations. ~r object is to define a class of functions C ~~ which we shall call the class of
functions computable in terms of ~ - . - - Before developing C~ formally, we wish
to gIve an example, and fn: 6rder to give the
example, we first need the concept of conditional
expression. In our notation a conditional expreSSion has the form
(Pl- e l ,P2- e 2 ,·· .Pn -en)
which corresponds to the ALGOL 60 reference
language (5) expression
if PI then e

l

else if P2 then e 2 ••• else if
Pn then en

Here Pl, ••• ,Pn are propositional expressions
takIng the values T of F standing for truth and
falsity respectively.
The value of (Pl- el'P2 - e 2 ,·· .,Pn -en)
is the value of the e corresponding to the first
p that has value T. Thus
(4< 3 -7,2> 3 -8,2< 3 -9,4< 5 -7) = 9
Some examples of the conditional expressions
for well known functions are
Ixl = (x< 0 _-x,x 1--0 _x)
oij = (i=j -l,i"tj - 0 )
and the triangular function whose graph is given
in figure 1 is represented by the conditional
expression
tri(x)

= (x-t'-l

-O,x$O

-x+l,x~ 1

y

-I-x,
x> 1 _0)

--------~~~----~----------~~-----x
Figure 1

In this part we describe a number of new
formalisms for expressing computable functions
and related entities. The most important section
is l~ the subject matter of which is fairly well
understood. Trre other sections give formalisms
which we hope will be useful in constructing computable functions and In proving theorems about
them.

Now we are ready to use conditional expreSSions to define functions recursively. For
example, we have
n = (n-O

-l,n~ -+n.(n-l)~)

227

5.3
Let us evaluate 2! according to this definition.
We have
2~

(2=0 ....... 1,2,,10 ....... 2·(2-1)~)
2.1
2(1=0 ....... 1,1'0 ....... l.(l-l)~)
2.1.0!
2.1.(0=0 ....... 1,0'0 ....... O·(O-l)!)
2.1.1
2

The reader who has followed these simple
examples is ready for the construction of C (~
which is a straightforward generalization of the
above together with a tying up of a few loose
ends.
Some Notation. Let ~be a collection
(finite in the examples we shall give) of functions whose domains and ranges are certain sets.
C ~~ will be a class of functions involving the
same sets which we shall call computable in terms
of rq::-'.
Suppose f is a function of n variables and
suppose that if we write y=f(xl, ••• ,x ), each Xi
n
takes values in the set U and y takes its value
i
in the set V. It is customary to describe this
situation by writing
f :U xU x ••• xU ....... V
l 2
n
The set Ulx ••• xU of n-tuplets (xl' ••• ,xn ) is
n
called the domain of f, and the set V is called
the range o~
Forms and Functions. In order to make
properly the-definitions that follow, we will
distinguish between functions and expressions
involving variables. Following Churchl the
latter are called forms. Single letters such as
!,~,~, etc. or sequences of letters such as sin
are used to denote funcgions. Expressions such
as f(x,~), f(g(x),X), x +y are called forms. In
particular we may re~er to the function f
defined by f(x,y)=x +y. Our definitions will be
written as though all forms involving functions
were written f(, ••• ,) although we will use
expressions like x+y with infixes like ± in
examples.
Composition. Now we shall describe the ways
in which new functions are defined from old.
The first way may be called (generalized) composition and involves the use of forms. We-8hall
use the letters x,y, ••• '(sometimes with subscripts) for variables and will suppose that
there is a notation for constants that does not
make expressions ambiguous. (Thus, the decimal
notation is allowed for constants when we are
dealing with integers.)
Tile class of forms is defined recursively
as follows:
i) A variable x with an associated space U
is a form, and with this form we also associate U.
A constant in a space U is a form and we also
associate U with this form.
ii) If e , ... ,e are fo:ms associated with
l
n
the spaces U ' •.• ,U respect1vely, then
l
feel' ••• ,e ) is a fgrm associated with the space
V. In thi~ way the form (f(g(x,y),x) is

built from the forms g(x,y) and x and the function f.
If all the variables occurring in a form
e are among Xl' ••• x
n we can define a function
h(xl, ••• ,x ) = e. We shall
n
assume that the reader knows how to compute the
values of a function defined in this way. If
fl, ••• ,fm are all the functions occurring in
h

by writing

e we shall say that the function h is defined
by composition from fl, ••• ,f • The class of
m
functions definable from given functions by composition only is narrower than the class of
functions computable in terms of the given
functJ.ons.
Partial Functions. In the theory of computation it is necessary to deal with partial
functions which are not defined for all ntuplets in their domains. Thus we have the
partial function minus, defined by
minus(x,y)=x-y, which is defined on those pairs
(x,y) of positive integers for which
x is
greater than y. A function which is defined
for all n-tuplets in its domain is called a
total function. We admit the limiting case of a
partial function which is not defined for any
n-tuplets.
The n-tuplets for which a function described by composition is defined is determined
tn an obvious way from the sets of n-tuplets for
which the functions,entering the composition
are defined. If all the functions occurring in
a composition are total functions, the new function is also a total function, but the other
processes for defining functions are not so kind
to totality. When the work "function" is used
from here on, we shall mean partial function.
Having to introduce partial functions is a
nuisance, but an unavoidable one. The rules for
defining computable functions sometimes give
computation processes that never terminate, and
when the computation process fails to terminate,
the result is undefined. It has been shown that
there is no effective general way of deciding
whether a process will terminate.
Predicates and Propositional Forms
The space ~ of truth values whose only
elements are T (for truth) and F (for falsity)
has a special role in our theory. A function
whose range is ,r,r-is called a predicate. Examples of predicates on the integers are prime
defined by
T i f x is prime
prime(x)
F otherwise
and less defined by
T if x< y
less(x,y) =
F otherwise
We shall, of course, write x 1-+3) = 3
(1< 2 -+4,1< 2 -+3) = 4
(2) 1 -+1,3> 1 -+3) is undefined
(0/(,< 1 -+1,1< 2 ~3) is undefined
(l < 2 -+0/0,1< 2 ~l) is undefined
(1< 2 -+2,1< 3 -+0/0) = 2
The truth value T can be used to simplify
certain conditional forms. Thus, instead of
= (x< 0 -+ -x, x ~ 0 -+ x)
we shall write
Ix) = (x< 0 -+ -x, T~ x)
The propositional connectives can be defined
by conditional forms as follows.
p /\ q
(p -+ q, T -+ F)
p v' q
(p -+ T, T -+ q)
1\.1 P
(p ~ F, T -+ T)
pC q
(p ~ q, T -+ T)
Conside~ations
of truth tables shows that these
formulas give the same results as the usual definitions. However, in order tQ treat partial
functions we must consider the possibility that
p or q may be undefined.
Suppose that p is false and q is undefined; then according to the conditional form
defini tion p/\ g is false and ql\ p is undefined. This unsymmetry in the propositional
connectives turns out to be appropriate in the
theory of computation since if a calculation of
p gives F as a result q need not be computed
to evaluate p 11 q, but if the calculation of p
does not terminate, we never get around to computing P.
-)0

pc'

It is natural to ask if a function cond

n

of

2n

variables can be defined so that
condn(Pl,···,Pn'

e l ,·· .,en )
This is not possible unless we extend our notion
of function because normally one requires all
the arguments of a function to be given before
the function is computed. However, as we shall
shortly see, it is important that a conditional
form be considered defined when, for example,
PI is true and e
is defined and all the
l
other p's and e's are undefined. The required extension of the concept of function would
have the property that functions of several
variables could no longer be identIfied with
one-variable functions defined on product spaces.
We shall not pursue this possibility further
here.
We now want to extend our notion of forms
to include conditional forms. Suppose
Pl, ••• ,Pn are forms associated with the space
of truth values and

el, ••• e
are forms
n
associated with the same space V. Suppose
further that each variable xi occurring in
PI' ••• Pn

and

el, ••• e

with the space U.

n
Then

is associated with the
(PI -+e , ••• ,Pn -+e )
l
n

is a form associated with V.
We believe that conditional forms will
eventually come to be generally used in mathematics whenever functions are defined by considering cases. Their introduction is the same
kind of innovation as vector notation. Nothing
can be proved with them that could not also be
proved wIthout them.
However, their formal
properties, which will be discussed later, will
reduce many case-analysis verbal arguments to
calculation.
Definition of Functions by Recursion. The
definitio-n----- n~ = (n=O -+l,T -+n·(n-l)~)
is an example of definition by recursion. Consider the computation of O~
0: = (0=0 ~1,T -+O'(O-l)~)
We now see that it is important to provide that
the conditional form is defined even if a term
beyond the one that gives the value is undefined.
In this case (O-l)~ is undefined.
Note also that if we consider a wider
domain than the non-negative integers, n~ as
defined above becomes a partial function, since
unless n is a non-negative integer, the
recursion process does not terminate.
In general, we can either define single
functions by recursion or define several functions together by simultaneous recursion, the
former being a particular case of the latter.
To define simultaneously functions
fl, ••• f k , we write equations

229

5.3
f

Let I be the set of non-negative integers
[0,1,2, .• !? and denote the successor of an
integer r( by n' and denote the equality of
integers n
and n
by n =n • If we define
l
2
l 2
functions succ and ~ by

(x , ••• , x

lIn
e

k

The expressions cl, ••• ,e

k

must contain only

known functions and the functions f , ••• ,f •
k
l
Suppose that the ranges of the functions are to be
VI' ""V respectively; then we further require
k

that the expressions e , ••• e be associated with
k
l
these spaces respectively, given that within
e , ••• e the f's are taken as having the V's as
l
k
ranges. This is a consistency condition.
fi (xi'" .x
is to be eval uated for given
k
values of the x's as follows.
1. If e is a conditional form then the p's
i
are to be evaluated in the prescribed order
stopping when a true p and the corresponding e
have been evaluated.
2. If e has the form g(e*, ••• , e*) then
i
1
m
ei, ••• ,e~ are to be evaluated and then the funcg

applied.
3. If any expression f. (e*, ••• e*) occurs it
~

1

nj

is to be evaluated from the defining equation.
4. Any subexpressions of e that have to be
i
evaluated are evaluated according to the same
rules.
5. Variables occurring as subexpressions are
evaluated by giving them the assigned values.
There is no guarantee that the evaluation
process will terminate in any given case. If for
particular arguments the process does not terminate, then the function is undefined for these
arguments. The possibility of termination
depends on the presence of conditional expressions in the e. 's.

~unctions

i~f

The class of
C
computable in
terms of the given base functions ~ is defined
to consist of the functions which can be defined
by repeated applications of the above recursive
definition process.
2.

R0cursive Functions of the Integers
In Reference 7 we deVelop the recursive functions of a class of symbolic expressions in terms
of the conditional expression and recursive function formalism.
As an example of the use of recursive function definitions, we shall give recursive definitions of a number of functions over the integers.
We do this for three reasons: to help the reader
familiarize himself with recursive definition, to
show how much simpler in practice our methods of
recursive definition are than either Turing
machines or Kleene's formalism, and to prove that
any partial recursive function (Kleene) on the
non-negative integers is in C (T) where e::p:- contains only the successor function and the
predicate equality.

succ(n) = nt
eq (n ,n )
T if n = n
l 2
l
2
F if n I- n
2
l
then we write ~ =fsucc,e~. We are interested
in C [~~. Clear11 all functions in CE~;1
will h~ve'either integers or truth values as
values.
First we define the predecessor function
pred (not defined for n=O) by
pred(n) = pred2{n,0)
pred2(n,m) = (m'=n - m. T - pred2(n, m'».
W3 shall denote pred(n) by n-,
Now we defi ne the sum
m+n = (n=O ~,T - m' +n-),
th~ product
mxn = (n=O -,~ 0, T -+ m+mxn-)
the difference
m-n = (n=O.....>· m, T -+ m--n-)
which is defined only for m n, the inequality
m =n = (m=O) " (.-1/(n=O)/\ (m-!: n-)
the strict inequality
m < n = (m:::..n) 1 ~ (m=n)
the inte~er valued quotient
min = (m < n -->- 0, T ->, «m-n)/n) '),
the remainder
rem(m/n) = (m < n -+- m, T -+ rem(m-n/n»
and the divisibility of a number n by a number
m
min = (n=O) "'«n~m)A (ml(n-m»).
The primeness of a number is defined by
prime(n) = (n~),If (nl-l) 1\ prime 2(n,2)
where
prime2(m,n) = (m=n) V «mtn)1l prime2(n,m'»
The Euclidean algorithm defines the
greatest common divisor if we write
gcd(m,n) = (m > n - gcd(n,m) ,rem(n/m) =
-r m,T gcd(rem(m/m) ,m»
and we can define Euler's ~ -function by
fen) =~2(n,n)
where
«2(n,m) = (m=l- 1, gcd(n,m) = 1-+~2(n,m-) I ,
T -> ~2(n,m-»

°

The above shows that our form of recursion
is a convenient way of defining arithmetical
functions. \'Ie shall see how the properties of
the arithmetical functions can conveniently be
derived in this formalism in a later section.
Computable Functionals
The formalism previously described enables
us to define functions that have functions as
arguments. For example,
3,

n

.'2- a i
l.=m
can be regarded as a function of the numbers m
and ~ and the sequence
If we regard the

fai1'

sequence as a function f
recursive definition

we can write the

230
5.3
sum(m,n,f) = (m > n -- O,T

~

f(m) + sum
(m+l,n,f»
or in terms of the conventional notation

+.l:.

~ f(i) = (m > n ->- 0, T~' f(m)
If(i))
l.=m
l.=m+
Functions with functions as arguments are called
functionals.
Another example is the functional least(p)
which gives the least integer n such that pen) for
a predicate p. We have
least(p) = least2(p,0)
where
least2(p,n) = (p(n) - l ' n,T~ least2(p,n+l»
In order to use functionals it is necessary
to have a notation for naming functions. We use
Church'sl lambda notation. Suppose we have a
function f defined by an equation f(x , ••• ,x )=e
..
1
n
where ~ i s some expressl.on 1.n xl, ••• ,X •
n
The name of this function is /}«x , ••• x ),e). For
1

n

example, the name of the function f defined by
f(x,y) = x 2 +y is ~ «x,y) ,x2 +y). We have
~«xly),x2+y)(3,4) = 13 but
~«y,x),x2+y)(3,4) = 19.
The variables occurring in a rI defini tion are
dummy or bound variables and can be replaced by
others without changing the function provided
the replacement is done consistently. For example, the expressions ~«x,y),x2+y) and
.::7«u,v),u 2 +v) and ri«y,x),y2+x ) all represent
the same function.
n. 2
In the notation ~l 1. is represented by
sum(l,n ~«i),i2» and the least integer n for
which n 2 > 50 is reBresented by
least( A «n) ,n > 50)
When the functions with which we are dealing
are defined recursively, a difficulty arises.
For example, consider factorial defined by
factorial(l1) = (n=O ~ 1,T -+ n'factorial(n-l»
The expression
;J.( (n), (n=O -l- I, T -." n·factorial(n-l»)
cannot serve as a name for this function because
it is not clear that the occurrence of "factorial"
in the expression refers to the function defined
by the expression as a whole. Therefore, for
recursive functions we adopt an additional convention. Namely,
label(f, ~ «xl' ••• xn ) , e»
stands for the function
equation
f (xl" •• xn) = e

f

defined by the

where any occurrences of the function letter f
within e stand for the function being defined.
The letter f also serves as a dummy variable.
The factorial function then has the name
label(factorial, ~«n), (n=O ~ I,T ~ n'
factorial(n-l»»
and since factorial and ~ are dwmny variables
the expression
la be 1 (g, ~ « r) , (r=O -+ 1, T - r. g (r-l) ) ) )
represents the same function.
If we start with a base domain for our
variables, it is possible to consider a hierarchy
of functionals. At level I we have functions
whose arguments are in the base domain. At

level 2 we have functionals taking functions of
level 1 as arguments. At level 3 are functionals
taking functionals of level 2 as arguments etc.
Actually functionals of several variables can be
of mixed type.
However, this hierarchy does not exhaust
the possibilities, and if we allow functions
which can take themselves as arguments we can
eliminate the use of label in naming recursive
functions. Suppose that we have a function f
defined by
f(x) = t(x,f)
where~(x,f) is some expression in
x and the
function variable f. This function can be
named
label(f, -=l «x) ,e(x,f»)
However, suppose we define a function
g by
g(x,~) =
(x, ;j «x), y,(x,f'?»)
or
g = ;) «x, ce) '€'( x,~( (x) ,If'(x,h)))
We then have
f(x) = g(x,g)
since g(x,g) satisfies the equation
g(x,g) = E:-(x, ;::t«x),g(x,g»)
Now we can.write f as
f = -1 «x), ~«y, 1:"),~(y, A«U), ~(u,t""»»
(x, rl «y, ~)
(y, A «u),

e

e

~u, ~»»

This elinunates label at what seems to be an
excessive cost. ~ly, the expression gets
quite complicated and we must admit functionals
capable of taking themselves as arguments thus
escaping our orderly hierarchy of functionals.
4.

FW1ctions and Functionals
It nught be supposed that in a mathematical
theory of computation one need only consider
computable functions. However, mathematical
physics is carried out in terms of real valued
functions which are not computable but only
approximable by computable functions.
'Ve shall consider several successive extensions of the class C (~~ • First we adjoin
the universal quantifiefV to the operations used
to define new functions. Suppose e is a form
in a variable x and other variables associated
with the space ~ of truth values. Then
~-ComputabHl

"( (x) ,e)

is a new form in the remaining variables also
associated with ~. ~«x),e) has the value T
for given values of the remaining variables if
for all values of x, e has the value T.
V«x),e) has the value F if for at least one
value of x, e has the value F. In the
remaining-case, i.e. for some values of x e has
the value T and for all others e is undefined,
Y «x) ,e) is undefined.
If we allow the use of the universal
quantifier to form new propositional forms for
use in conditional forms, we get. class of functions Ha f9j which may well be called the class
of functions hyper-arithmetic over ~ since in
the case where ~= ~successor, equalit
on the
integers, Hatry consists of Kleene's hyperarithmetic functions.
Our next step is to allow the description
operator ~ • L«x),~(x» stands for the unique

S)

231

5.3
x such that ~(x) is true. Unless there is such
an x and it isunique (: «x) ,~(x» is undefined.
In the case of the integers««x),~(x» can be
defined in terms of the universal quantifier using
conditional expressions, but this does not seem to
be the case in domains which are not effectively
enumerable, and one may not wish to do so in
domains where enumeration is unnatural.
The next step is to allow quantification over
functions. This gets us to Kleene ' s 5 analytic
hierarchy and presumably allows the functions used
in analysis. Two facts are worth noting. First
~«f), ~f»
refers to all functions on the
domain and not just the computable ones. If we
restrict quantification to computable functions,
we get different results. Secondly, if we allow
functions which can take themselves as arguments,
it is difficult to assign a meaning to the
quantification. In fact, we are apparently confronted with the paradoxes of naive set theory.
Ambiguous Functions
Ambiguous functions are not really functions.
For each prescription of values to the arguments
the ambiguous function has a collection of possible
values. An example of an ambiguous function is
lessen) defined for all positive integer values
of n. Every non-negative integer less than n is
a possible value of lessen). If we define a-basic
ambiguity operator amb(x,y) whose possible values
are x and I when both are defined otherwise whichever-is defined, we can define lessen) by
lessen) = amb(n-l,less(n-l».
lessen) has the property that if we define
ult (n) = (n=O --.. 0, T --.. ul t (less (n» )
then
'V«n),ult(n)=O) = T.
There are a nUlaber of important kinds of
mathematical arguments whose convenient formalization may involve aml?iguous functions. In order to
give an example, we need two definitions. If f
and g are two ambiguous functions, we shall say
that-f is a descendant of g if for each x every
possible value of f(x) is also a possible value of
g(x). Secondly, we-shall say that a property of
ambiguous functions is hereditary if whenever it
is possessed by a function g it is also possessed
0y all descendants of g.
The property that iteration of an integer valued flllction eventually
gives 0 is hereditary and the function less has
this property. So, therefore, do all i~
descendants. Thus any function, however, complicated which always reduces a number w ill if
iterated sufficiently always give O.
This example is one of our reasons for hoping
that ambiguous functions will turn out to be
useful.
With just the operation amb defined above
adjoined to those used to generate C ~~~ , we
can extend
to the class C* [~ wb.ic~ may be
called the computably ambiguous functions. A
wider class of ambiguous functions if formed using
the operator Am(x,~(x» whose values are all
x's satisfying ~(x).
5.

r

6.

Recursive Definitions of Sets
In the previous secti~s~rccursive definition of functions the domains and ranges of the
basic functions were prescribed and the defined
functions had the same domains and ranges.
In this section we shall consider the definition of new sets and the basic functions on
them. First we shall consider some operations
whereby new sets can be defined.
I. The Cal'tes~an product AxB of two sots A
and B is the set of all ordered pairs (a.b) with
a(A and beB. If A and B are finite sets and
n(A) and nCB) denote the numbers of membors of
A and R respectively then l1(AxB=n(A)·n(B).
Associated \uth the pair of sets (A,B) are
two canonical mappings
~
:AxB ->- A defined by IT
{(a.b» = a
(A,B:AxB -+ B defined by ~A,B{(a.b» = b
,A,B
A,D
The word "canonical" refers to the fact that
~
and
are defined by the sets A and B
A,B
A,B
and do not depend on knowing anything about the
members of A and B.
The next canonical function riS a function
of two variables ~:A B~ AxB defined by

e

~

'A B

I

(a,b) = (a'b)
A,B
For some purposes functions of two variables x
from A and y from B can be identified with functions of one variable defined on AxB.
2. The direct union AGB of the sets A and B
is the union of two ,non-intersecting subsets one
of which is in 1-1 correspondence with A and the
other with B. If A and B are finite, then
n(AGB) = n(A)+n(B) even if A and B intersect.
The elements of A@B may be written as elements of
A or B subscripted with the set from which they
come, i.e. a or b •
A
B
The canonical mappinGs associated with the
direct union AGB are
i
: A -)- AG8 defined by iA (a)
A,B
,B
j
: B .,.-)- A@B defined by jA B(b)
b'
,
A,B
B

p
: AGB -> ~ defined by PA,B(x)
T if and
A,B
only if x comes from A.
q
: AGB -'r ~ defined by q
(x) = T if and
A,B
A,B
only if x comes from B
There are also two canonical partial functions r
: AGB --Jo- A which is defined only for
elementsAc~ming from A and satisfies
r
(
(a»=a. Similarly SA A@B-l-B satisfies
A,B A,B
,B
sA,B(jA,B(b»=b.
3. The power set ,/' is the set of all
mappings f:B ->- A. The canonical mapping
./-,
:ABXS-)-A is defined byqt..
(f,b) = f(b)
A,B
A,B

Canonical Mappings
We will not regard the sets Ax(BxC) and
(AxB)xC as the same, but there is a canonical 1-1
mapping between them
r:.:
: (AxB) xC -+ Ax(exc)
~A,B,C

232
5.3
defined by

~A,B,C(U) = dA,BXC(rrA,B(7TAxn;C~»' ~,C
~(A,B(rrAXB,C(U»,

AxB,C(u»).

We shall wri te
(AXB)xc~Ax(BXC)

to express the fact that these sets are
canonically isomorphic.
Other canonical isomorphisms are
1. t
: AxB -~. BxA defined by
A,B
t(u) = y. (e
(u),rr
(u»
B,A A,B
A,B
2. d : Ax(nOC) ._)- AxB@AXC
l
;j.
a: (AGl3)OC -> A0(BOC)
2

d: p~ xrF ._). (AxBP
2
d : ABxl~ -> lJGC

4.

5

•
6.

3

s: (AB'P
1

_+./'xC

We shall denote the null set (containing no
by 0 and the set consisting of the
integers from 1 to n by n. We have
A@ oj!::A
element~

A9 O~
Axl~A

Ax2~~@A

AO~

(n terms,associate to left by
convention)
(by convention)

Al,-vA
A ~AX ••• xA (n terms, associate to left by
convention)
Suppose we write the recursive equation
S =(.A} GAxS
we can interpret this as defining the set of
sequences of elements of A as follows:
1. In terpret j\.. as denoting the null sequence,
Then the null sequence (strictly an image of it)
is an element of S.
2. Since a pair consisting of an clement of
A and an element of S is an element of S,a pair
(a,A) is an element of S. So,then,are
a l • (a '11» and a • (a • (a ,,») etc.
Z
l
2
3
Thus S consists of all sequences of elements of A
including the null sequence.
Suppose we substi tute
~
@AxS
for S in the right side of S =.
GAxS. We get
S = tA1GAX( (AJ@AXS)
If we aga1n ~bstitute for S.and expand bv the
distributive law we ~et
S =
GAx U)fAxAX [-:Ii) + ••••
which if we denotft6.e set~"!J becomes
S ~ IGA9A 20A3G •••

f.("1

[:!J

which is another way of wri tillg the set of
sequences. We shall denote the set of sequences
of elements of A by seq(A).
Another useful ~ursiye construction is
S = AGSxS
Its elements have the forms a or (aI' a ) or
2
«a 'a )'a ) or (a '(a 'a » etc. Thus we have
2 3
I 2
3
I
the set of S-expressions on the alphabet A which
we may denote by sexp(A). This set is the subject
matter of Reference 7 and the following paragraph
refers to this paper.

When sets are formed by this kind of
recursive definition, the canonical mappings
associated with the direct sum and Cartesian
product operations have significance. Consider,
for example, sexp(A).
We can define the basic operations of LISP,
i.e. atom, eq, car, cdr and cons by the equations
atom(x) = p
(x)
A,SxS.
eq(x,y) = (~,SXS(x) = ~A,SXS(Y»
assuming that equality is defined on the space A
car(x) = rrS, S CAA, SxS (x»
cdr(x) = es, s (~
-A,SX Sex»~
cons(x,y) = jA,SXS(~,s(x,y»
Definition of the set of integers.
Let-'\,. denote the nullset and 1/1) be the
set whose sole element is the null set. We can
define the set of integers I by
I = f,4 G 11 xl
its elements are then
/J, (A, 4), (A, (I(,A», (A, (A, (/) ,A») etc.
which we shall denote by 0,1,2,3 etc.
The successor and predecessor functions are then
definable in terms of the canonical operations
of the defining equation. We have
succ(n) = ~J,IVf,n)

J f J

pred(n)

= e}

fA

(,a....-).~ (n»

,I

'f,1y'r I
I

PROPERTIES OF CO~WUTABLE FUNCTIONS
The first part of this paper was solely
concerned with presenting descriptive formalisms,
In this part we shall establish a few of the
properties of the entities we previously introduced. The most important section is the
second dealing with recursion induction.
7.

Formal Properties of Conditional Forms

~heory of conditional expressions

corresponds·to analysis by cases in mathematics
and is only a mild generalization of propositional
calculus.
We start by considering expressions called
generalized Boolean forms (gbf) formed as follows:
1. Variables are divided into propositional
variables p,q,r, etc. and general variables
x,y,z, etc.
2. '/ie shall write (p -i- x, y) for
(p -+ x, T -> y) • (p...,. x, y) is called an elementary
conditional form (ecf) of which p,x, and yare
called the premise, conc~usion and the
alternative, respectively.
3. A variable is a gbf and if it is a
propositional variable it is called a pf
(propositional form).
4. If rr is a pf and .,.Land fJ are gbfs, then
('lr~,~) is a ~bf~ If, in addition,A..and 13
are pfs so is (rr -;- pI-,JJ) •
The value of a gbf ~for given values
(T,F or undefined) o~he propositional variables
will be T or F in case ~ is a pf or a general
variable otherwise. This value-rs determined for
a gbf (7T -r.,L,.. ,3) according to the table

233
5.3
value

value «7T -+""- , JJ »

('IT)

T

value (0')

F

value (1)

~.0'0
cd~

~

undefined

undefined

~
~

~

t

-;:;
T

pqr
TTT
TTF
TTu

.e
T
T

T

-;
H

(I$~

~

~

(I$~

(I$~

-r

l'

t

0'

l

.e
'-'

~

'-'

a
a
a

a
a
a

a
b
u

b
b
b

a
b
u

b
b

u

a
b

u
u
u

H

'0

(I$~

We shall say that two gbfs are stron:,:;ly
equivalent if they have the same value for all
values of the propositional variables in them including the case of undefined propositional
variables. They arc weakly equivalent if they
have the same values for all values of the
propositional variables when these are restricted
to T and F.
The equivalence of gbfs can be tested by a
method of truth tables identical to that of
propositional calculus. The table for
«p -)- q, r) - l - a, b) and (p -~ (q -) a, b) (r ->- a, b» is

(Ij~

-- I' - - t
--t t t
- - --

0.

a
a
a

C)

'0

(Ij~

.o~

0.

0.

0.

0'

a
a
c
c

b
b
d

a
b
c

d

d

i'

~

~

'-'

T
T
F
T

a
b
a
b

c
d
c
d

a
b
c
d

'0

.0
;

T

0'0"

P q
T

(I$~

C)~

C)

t

F
T
T

~

C)

0.

i

which proves that (p -+ (q -} a, b), (q -+ c,d» and
(q ...... (p -+ a,e), (p -, b,d» are weakly equivalent.
They are also strongly equivalent. '.'/e shall
wri te
and -=for the relations of
-s
w
strong and weak equivalence.
There are two rules whereby an equivalence
can be used to generate other equivalences.
1. If ~
13 and 0'-1
.JJ 1 is the
result of substituting any gbf for any variable
• This is called
in ~ =...B
• then...< 1 El

=-

==

T

the rule of substitution.
2. If"",
1) and.?- is sub-expressioh of
"""and & is the result of replacing an occurrence
of"+'" in y
by an occurrence of.J3 , thenr-::=' 6
This is called the rule of replacement.
These rules are applicable to either strong
or weak equi valence and in fact to .l1uch more
general situations.
Weal< equivalence corresponds more closely
to equivalence of truth functions in propositional
calculus than does strong equivalence.
Consider the equations
1)
(p --> a,a)
wa

==-

TFT
TFF
TFu

F
F
F

b
b
b

TuT
TuF
Tuu

u
u
u

u
u
u

u
u

u

FTT
FTF
FTu

T
F
u

a

a
a
a

a
b
u

a

2)

(T- a,b) =-sa

b
u

3)

(F-l-a,b):;:"sb

4)'

(p

FFT
FFF
FFu

T
F
u

a
b
u

b

a

b
b

5)

(p.-~-

(p->-a,b),c) ==-s(p->a,c)

u

a
b
u

6)

(p

a, (p

FuT
FuF
Fuu

T
F
u

a

a
b
u

a
b
u

7)

u

u
u
u

«p->q,r) -) a,b);::.s(P- (q-)"a,b),
(r -)- a, b»
( (p -~:- - (q --. a, b) , (q -" c, d» ===- s

uTT
uTF
uTu

u
u
u

u
u
u

a
a
a

a
b
u

u
u
u

uFT
uFF
uFu

u
u
u

u
u
u

b
b
b

a
a
u

u
u
u

uuT
uuF
uuu

u
u
u

u
u
u

u
u
u

a
b
u

u
u
u

b

u

b

b

b

According to the table «p -+ q, r) - ? a, b)
and (p -- (q
a, b), (r -+ a, b» are strongly
equivalent.
For weak equivalence the u case can be
left out of the table. Consider the table
-)0

=

8)

-l

-r

T,F)

== sP
->

b,c» :::'s (p -;. a,c)

(q -;.. (p

-3'

a,c), (p -;. b,d»

All are strong equivalence except the first and
can be proved by truth tables.
These eight equations can be used as axioms
to transform an ::;bf into any weakly equivalent
one using substitution and replacement. In fact,
they can be used to transform any gbf into a
canonical form. This canonical form-is the
following. Let PI' ""Pn be the variables of the
gbf a taken in an arbitrary order.
be transformed into the form
(PI -> a ' a )
O l
where each a
a l =(P2

l

has the form
-l

aiO,a il )

Then

a can

234

5.3
and in general for each k = l,n-l
a. , •••. = (P. i<>a. , ••• ,. ,O,a. ···.1)
11
1k
1+
11
1k
11
1k
and each a , ••• ,
is a truth value or a general
il
in
variable.
For example, the canonical form of
q,r) -> a, b)
with the variables taken in the order r,q,p is
(r -- (q -r (p -). a,a), (p~. b,a», (q ~ (p -~ a, b),

«p -

n

(p~·b,b»)

In this canonical form, the 2 cases of the truth
or falsity of Pl, ••• ,Pn are explicitly exhibited.
An expression may be transformed into
canonical form as follows:
1) Axiom 7 is used repeatedly until in
every sub-expression the 7f in (IT -- t:A ~ ) consists

replacement on formulas we need the additional
axioms
(9) (p - (q - a, b) ,c) ~ (p - (q -~. (p - . . a,a),
~s
(p - b, b» ,c)
and
(10) (p-,.a,(q-l b,c»~s(p-a.(q-.;. (p-, b,b),
(p ->- c, c»)
Suppose there is an occurrence of PI in the
conclusion; we want to replace it by T. To do
this, we use axioms 9 and 10 to move in a PI

until the objectionable PI occurs as the inner
PI of one of the forms
(PI-> (Pl-l"o a,b),c)
or
(PI -',. a, (PI

->

b,c».

In either case, the objectionable PI can be

of a single propositional variable.
2) The variable PI is moved to the front by

removed by axiom 5 or 6 and the PI'S that were

repeated application of axiom 8. There are three
cases: (q-> (Pl-:·a,b),(Pl-c,d» to which axiom

moved in can be moved out again.
Thus we have (PI -+ e:A--,fi) with PI missing

8 is directly applicable.
(q -+ a, (PI -+ c,d» where axiom 8 becomes applicable

from ~ and jJ' •
4)
Inevitable variables are then brought to
the front of t?'-and.J? u!ld so forth.
Two gbfs are equivalent (weakly or strongly)
if and only if they have the same(weak or strong)
canonical :form. One way this is easy to provej
if two gbfs have the same canonical form they
can be transformed into each other via the
canonical form. Suppose two gbfs have different
weak canonical forms when the variables are taken
in the same order. Then values can be chosen
for the p's giving different values for the form
proving non-equivalence. In the strong case,
suppose that two gbfs do not have the same inevitable propositional variables. Let p be inevitable in a but not in b. Then if the other
variables are assigned suitable values b will
defined with p undefined. However, a wIll be
undefined since p is inevitable in a-which proves
non-equivalence. Therefore, strongly equivalent
gbfs have the same inevitable variables, so let
one of them be put in front of both gbfs. The
process is then repeated in the conclusion and
alternative etc.
The general kind of conditional expression
(P -+ e , ••• ,Pn - en)
l
1

after a~dom 1 is used to make it
(q --> (PI -,. a,a), (PI -) c,d» and then axiom 8 is
applied, the case (q

->-

(PI - a, b) I c) is handled

similarly.
Once the main expression has the form
(p - d-. ,ft) we move any PI's which occur in
, 1

c;A....

and~ to
5 and 6.

the front and eliminate them using axioms
We then bring P to the front of c:I- andJ!
using axiom I if neces~ary to guarantee at
least one occurrence of P2 in each of ~ and~ •

The process is continued until the canonical form
is achieved.
There is also a canonical form for strong
equivalence. Any gbf a is strongly equivalent
to one of the form(Pl--> ~, .Jl), where ~nd ..ff
do not contain PI and are themselves in canonical
form.

However, the variable PI may not be chosen

arbitrarily but must be an inevitable propositional
variable of the original gbf and can be chosen to
be any inevitable variable. An inevitable
variable of a gbf (IT -If?'-Jl) is defined to be
either the first propositional variable or else
an inevitable variable of bot~ and~ •
A gbf a may be put in strong canonical form
as follows:1) Use axiom 7 to get all premises as
propositional variables.
2) Choose any inevitable variable, say PI'
and put a in the form (PI -0'--,3) by using
axiom 8.
3) The next step is to eliminate occurrences
of PI in A and!3. Thi s can be done by the
general rule that in any ecf occurrences of the
premise in the conclusion can be replaced by T
and occurrences in the alternative by F.
However, if we wish to use substitution and

can be regarded as having the form
(PI -> e l , (p 2 - e 2 ,·· .(Pn -)0 en,u) ••• »
where u is a special undefined variable and
their properties can be derived from those of
ece's.
The relation of functions to conditional
expressions is given by the distributive law
f(xl, .. ·,xi_l,(Pl-el, .. ·,Pn ~·en)'
x

, ••• x )

=

i+l
k
(p ->f(x , ••. x.
,e ,x.
, ••• ,x ), ••• ,p 1
1
1-1 1 1+1
n
n
f(x , ••• ,x
,e ,x
, ••• ,x »
1
i+l n i+l
n

235
5.3
The rule of replacement can be extended in
the case of conditional expressions. Suppose ~
is an occurrence of a sub-expression of an expression
We define a certain propositional
expression 7f called the premise of ~in ]3 as
follows:
I) The premi se Of.-- in
i sT.
2)" The premise of pin f(x ,···, n'·· .x )
l
n
where j3 is part of~ is the premi se of t7'-- in J?
3) The premise of
in
(p -;.. e , ••• p -r e , ••• p -;.. e )
I
Iii
n
n
where [3occurs in e , and the premise of~ in e

rr-

¥

i

is

7f,

is "1..Jp ;\. ••• II £...-)P.

where

~-

I

4)
(p

f

I

I"

i

p. A
~

The premise of p in
- e , ••• p. -)- e. , ••• p

-)0

e )

n
occurs in Pi' and the premi se of O"-"in Pi
~

I

~

n

7f.

, is /!Vp I A ••• Il t'VP.~-l/\ 7f.
The extension of the rule of replacement is
tha t an occurrence o~ inl1 may be replaced by 0('
if (7f -;.pZ.)-== (7f ->P< ') wliere 7f is the premise of
-s
oL-in F. Thus in a subcase one need only prove
equivalence under the premise of the subcase.

is

m and n and hence is defined by th~ equation.
(m+n) I
(n=O -;.. m,T -;.. ml +n-) ,
== (n=O -)- mI ,T - l (m I +n-) I )
m+n
= (n=O -l- m' ,T -;.. (m') +n-)
It is easily seon that the functions g and h
defined by the equations g(m,n) = (m+n) , and
h(m,n) = m'+n both satisfy the equation f. For
example, it is clear that ~(m' ,n-) = (ml+n-), and
hem' ,n-) == (m')I+n-. Therefore, by the principle
of recursion induction hand g are equivalent
fUnctions on the domain of where f is defined,
bu~ this is the set of all pairs of integers.
The fact that f(m,n) converges for all m and
n is a case of the fact that all functions defined
by equations of the form
Hn,..x_ .... x) = (n=O-.. g(xl ••• Xn),T--h(f(n-,ri
(x 1 ' .•• xn) , •.. l' l~ (Xl ••• xn) ) , ••• ,
f (n - ,1'1k (xl' •.• , xn ) , ••• , rnk

7f

S.

Recursion Induction
Suppose a function

by
1)

f

is defined recursively

f (x , •• ., x ) = e ~. xl'.'. x , f)
1

C

n

{

n

where£" is an expression that in general contains
f. Suppose that C{ is the set of n-tuplets
(x , ••• , x ) fOl' which f is defined. Now let g and
I

n

-

h be two other functions with the same domain as f
and which are defined for all n-tuplets in GL •
Suppose further that g ant: h satisfy the equation
which defined f. We assert-that
g(x , ••• ,X ) ,== h(x , ••• x )
n
n
l
l
for all (x., ••• ,x ) in C{. This is so simply
~

n

because equation 1) uniquely determines the value
that any function satisfying it has for arguments
in 4. which in turn follows from the fact that
1) can be used to compute f(x , ••• ,x ) for
1

n

(x , ••• x ) in C{
I
n

We shall call this method of proving two
functions equivalent by the name of recursion
induction.
We shall develop some of the properties of
the elementary functions of integers in order to
illustrate proof by recursion induction. We recall
the definitions
m-tn = (n=O -> m, l' -;.. mI +n-)
ron == (n=O -+ 0, T --,' n1+nlll -)
Th 1. m+O
m
Proof m+O = (0=0 -> f.1, T - m1+0-)

conver!~8.

(xl" .xn ) ,n, xl" .x »
n
We shall postpond discussinG formal

proofs of conver:~8nC2.
In prcs~~tinG further proofs we shall be
more terso.
Th 3. (1I1+~)+P = (m-tp)+n
Proof. Ld't f(m,n,p) = (p=O~-m+n,T--f(m',n,p-»
AGain f conver~es for all m,n,p. We have
(m+n) +p
(p=O...., nun, T - r (m+n) '+p-)
(p=O->-m-l-n,T->- (m'-tn)+p-) using Th 2.
(m+p) -:-n
(p=O - r 111, T -- m' +p-) +n
(p=O-+m+n,T-- (m'+p-)+n)
Each of these forms satisfies the equation for
f(l11,n,p).
Settin~ m=O in Th 3 gives
(O+n)+p = (O+p)+n
so that if we had O+m = m we would have
commutativity of addition.
In fact, we cannot prove O+m = m wi thout
makin~ some assumptions that take into account
that we are dealing with the integers. For
suppose our space consisted of the vertices of
tho binary tr~~ where ml is the vertex just above
and to th~ left and m- is the vertex just below
and 0 is the bottom of the
tre'c. m+n can b0 defined as
above and of course satisfies
Theorems 1, 2 and 3 but does
not satisfy O+m == nl.
We shall make the following assumptions:
1. ml p 0
2.

(m') -

=m

3. (m#O);:>«m-)1 -= m)
which embody all of Peano's axioms except the
induction axiom.

Only the definition of addition and the properties
of conditional expressions were used in this proof

Th 4. O+n = n
Proof. Let fen) = (n=O - r 0, T -;.. f(n-) I)
O+n = (n=O -- 0, T - r 0 1+n-)
= (n=O -r O,T -r (O+n-)')
n = (n=O -r n,T - r n)
(n=O -+ 0, T -1- (n-) I)
axiom 3

Th 2 (m+n) I = ml+n
Proof Define f(m,n) == (n=O -;.. m',T

Th 5. m+n
n+m
Proof. By 3 and 4 as remarked above

=m

-~

f(m,n-»

It is easily seen that f(m,n) converges for all

236

5.3

Now
Th (). (m+n) +p = ul+(n+p)
(m+p) +n
Proof. (m+n) +p
(p+m) +n
(p+n) +m
m+(n+p)

[X*y] *z
Th
Th
Th
Th

3
5

3
5 twice

Th 7. m.O = 0
(0=0 -+ O,T -+ m+n·O-)
Proof. m.O

=0
=0

Th 8. O·n
Proof. Let f(n)

O·n
o
TIl 9.

Proof.

mn'
mn'

(n=O -+ O,T -+ f(n-»
(n=O -> 0, T -)- 0+0 ·n-)
(n=O -+ O,T -+ 0)

m+Jnn
(n'= 0 -+ O,T-+ m+m·(n')-)
nl+lnn
axioms 1 and 2

Now let
f( XjYj z]

=£ null[ x*y ] -+

Zj T -+ cons[ cal[ x*y];
cdr [x*y]z
=[null[ x] -;. [null{ y] --+ z; T -+ cons
[ca:r{ x*y]; cdr [x*y]*z ]] T-+ cons
[car[ x*y ]j cdr [x*y] *z] ]
=fnull[ x] -+ [null( y] -+ z; T -+ cons [car
y]; cdr [y J"z T -+ cons [car [x ]j[ cdr
[x J*y] *z]]]
=[null[x] -+ y*zjT -+ cons[ car [x] [cdr
[x] *y]*z ]]

=[

null [x ]-+ y*zjT

[x]

-+

cons [car[ X]j f( cdr

jYi Z ] ] ]

From the above steps we see that [x*y)*z satisfies
the equation for f. On the other hand
x*[ y*z ]:;::: [nuIl[x] -+ y*Zj T -+ cons[ car ~]; [cdr
[x }[ y*z ] ]]
satisfies the equation directly

Th 10. m(n+p):::: mn+mp
Proof. Let f(m, n,p) = (p=O -+ nm, T -l> f(ln, n' , p-»
m(n+p)
m(p=O -)- n, T -+ n' +p-»
:::: (p::::O -) nm,T -l> m(n' +p-»
mn t-mp = mn+(p=O -+ 0, T ~. m+mp-)
(p::::O -+ mn+O, T -> nm+(m+mp-»
(p=O -> mn,T -+ (mn+m)+mp-)
(p=O -I- nm, T -) mn' +mp-)

NIL*x = x
x*NIL = x
Proof. NIL*x = [null fiIL]~' Xj T -;. cons par [NIL] j
cdr [NIL]*x ] ]
=x
x*NIL = [null ~ }-~ NILj T -> cons ~ar [x 1
cdr [x] *NIL ] ]
Let f(x) = [null[ x] -+ NILjT -+ con~aar [x] jf[cdr

NoW we shall give some examples of the
application of recursion induction to proving
theorems about functions of symbolic expressions.
The rest of these proofs depend on an acquaintance
with the LISP formalism
We sta~t with the basic identities.
car [cons[ Xj ~] :;: : x
cdr[ cons[ Xj ~] = y
a tom [x] [ cons[ car [xj : ccl r [ x]
x
a tome cons[ Xj y ] ]
nul~ x] :;::: eq[XjNIL]

x*NIL satisfies this equation. We can also write
for an~ list x
x = lnull[ x] -,. Xj T -+ x]
= [null[ x] ._.- NILj T -> cons[ car [x] j cdr[ x] ] ]
which also satisfies the equation.
Next we consider the function reverse x
defined by
reverse W
/pull [,a -> NIL; T -+ reverse[ cdr [x] ]
*cons [car[ x]; NIL] ]
It is not difficult to prove by recursion
induction that
reverse[x*~ = reverse &]* reverse~]
and
reverser reverse [,q] = x.

Let us define the concatenation x*y of two
lists x and y by the formula
x*y :;::: [null[ x}--l- YjT -- cons [car[x]jcdr[x }l y,T-)l
n
2
n
l
h (y' ,xl' ••• ,xn »
2
The converse statement that all functions in
C
suce, eq "') are partial recursive is presumably
also true but not quite so easy to prove.
It is our opinion that the recursive function
formalism based on conditional expressions
presented in this paper is better than the formalisms which have heretofore been used in
recursive function theory both for practical and
theoretical purposes. First of all, particular
functions in which one may be interested are more
eas~ly written down and the resulting expressions
a~e briefer and more understandable.
This has
been observed in the cases we have looked at and
there seems to be a fundamental reason why this is
so. This is that both the original Church-Kleene
formalism and the formalism using the minimalization operation use inte~er calculations to control
the flow of the calculations. That this can be
done is noteworthy, but controlling the flow in
this way is less natural than using conditional
expressions which control the flow directly.
A similar objection applies to basing the
theory of computation on Turing machines. Turing
machines are not conceptually different from the
automatic computers in general use, but they are
very poor in their control structure. Any programmer who has also had to write down Turing
machines to compute functions will observe that
one has to invent a few artifices and that constructing Turing machines is like programming. Of
course, most of the theory of computability deals
with questions which are not concerned with the

'l

l

.3

particular ways computations are represented. It
is sufficient that computable functions be
represented somehow by symbolic expressions, e.g.
numbers, and that functions computable in terms
of given functions be somehow represented by expressions computable in terms of the expressions
representing the original functions. However, a
practical theory of computation must be applicable to particular algorithms. The same objection applies to basing a theory of computation on
Markov's 9 normal algorithms as applies to basing
it on properties of the integers; namely flow of
control is described awkwardly.
The first attempt to give a formalism for
describing computations that allows computations
with entities from arbitrary spaces was made by
A. P. Ershov 4 • However, computations with the
symbolic expressions representing program steps
are also necessarily involved.
We now discuss the relation between our formalism and computer programming languages. The
formalism has been used as the basis for the LISP
programming system for computing with symbolic
expressions and has turned out to be quite practical for this kind of calculation. A particular
advantage has been that it is easy to write
recursive functions that transform program which
maltes genera tors easy to wri te •
The relation between recursive functions and
the description of flow control by flow charts is
described in Reference 7. An ALGOL program can
be described by a recursive function provided we
lump all the variables into a single state vector
having all the variables as components. If the
number of components is large and most of the
operations performed involve only a few of them,
it is necessary to have separate names for the
components. This means that a programming
language should include both recursive function
definitions and ALGOL-like statements. However,
a theory of computation certainly must have
techniques for proving algorithms equivalent and
so far it has seemed easier to develop proof
techniques like recursion induction for recursive
functions than for ALGOL-like programs.
References
1.

2.
3.
4.
5.
6.

Church, A., The Calculi of Lambda-Conversion,
Annals of Mathematics Studies:-no. G,
Princeton, 1941, Princeton University Press.
Church, A., Introduction to Mathematical Logic,
Princeton, 1952, Princeton-University preSS:-Davis, lvI., Computability and Unsolvability,
New York, 1958, McGraw-Hill.
Ershov, A.P., On Opera tor Algori thms (Russian),
Doklady Akademii Nauk, vol. 122, no. 6,
pp. 967-970.
Kleene, S.C., Recursive Predicates and Quantifiers, Transactions of the American-Mathematical Society, vol. 53, 1953, p. 41
McCarthy, J., letter to the editor,
Communications of the Association for Computing Machinery, vol. 2, August, 1959, p. 2.

238

5.3
McCarthy, J., Recursive Functions of Symbolic
Expressions and Their Computation by Machine,
Part I, Communications of the ACM, vol. 3,
April, 1960, pp. 184-195.
8. McCarthy, J., The LISP Programmer's Manual,
M.I.T. Computation Center, 1960.
.
9. Markov, A.A., Theory of Algorithms (Russian),
Moscow, 1954, USSR:Academy of Sciences,
Steklov Mathematical Institute.
10. Naur, P., et al., Report on the Algorithmic
Language ALGOL 60,Communications of the ACM,
vol. 3, May 1960.
11. Turing, A.M., On Computable Numbers ~ ~
Application to the Entscheidungs Problem,
Proceedings of the London Mathematical Society,
sere 2, vol. 43, 1937, p. 230; correction,
ibid, vol. 43, 1937, p.544.

7.

239
6.1

:INFORMATION RETRIEVAL; STATE OF TIlE .ART

Don R. Swanson
SUMMARY

Certain aspects of science communication are
of especial importance to information retrieval.
The exponential growth of science raises many
questions related to increasing specialization.
"Quality identification" is suggested as a critical issue. All approaches to information retrieval share a common set of basic problem areas
and solutions. Semantics and redundancy are key
conceptual issues and give rise to difficulties
more likely to be overcome by meticulous thesaurus
compilation than by any sudden insight or "breakthrough. " The effectiveness of present retrieval
capabilities is largely unknown, though certain
recent studies are illuminating. Presentation
and display to the user are suggested as important
approaches to problems of information digestibility.

******

Information retrieval embraces only one
aspect of a broad class of science communication
problems. A proper perspective for assessing the
state of the information retrieval art can best
be achieved through considering first the broader
problem context. Certain elements of that context will be discussed before the subject of
information retrieval itself.
Important Aspects of Science Communication
Who Pushed the Panic Button?
The obvious fact that the world's store of
scientific information is increasing at an exponential rate has apparently enjoyed independent
discovery by scores of science information writers
during the past few years. On a suitably scaled
graph, any rising exponential curve tends to produce an hypnotic effect of impending crisis. Now
an atmosphere of alarm, since it is conducive to
action, is no doubt beneficial, but only provided
such action is properly directed. Increasing
volume of information is not in itself the real
problem. It can be inferred in fact that in the
context of increasing world scientific population,
the information increase rate is not at all out
of balance. The number of published papers, the
number of journals, the number of scientists, and
expenditures for research, all have been increasing atlan annual rate of 5-7% for the past 250
years.
Since the earth's population increases
at only 1.6% annually, indiscriminate extrapolation of the exponential growth of science
clearly is absurd. ll Limiting factors will of
course set in. Prof. Bar-Hille12 has challenged
the view that anything at all is worsening (so far
as basic long term trends are concerned), and
rightly insists that the prophets of calamity must
produce more convincing evidence than that which
has yet come to light. Since on a scientist per

capita basis the quantity of scientific information remains about constant, the case for pushing a panic button is at best obscure. The real
issue must be identified as a changing partition
of an increasing supply of human knowledge among
the increaSing number of scientists who generate
and assimilate that knowledge. More simply, a
scientist cannot now realistically expect to keep
on top of as large a portion of any particular
field or discipline (such as "physics" or "logic")
as he could some years ago. The problem must be
considered in terms of "backlog" as well as
"current production."
Many profound questions, presently unanswerable, should be diligently explored, not in an
atmosphere of crisis but in recognition of their
long term fundamental importance. Should scientific journals be organized to meet increasingly
more specialized needs? Does the manner of
journal organization itself influence the course
pf scientific research? Can excessive specialization inhibit scientific creativity? Should
steps be taken to "purge" the world's scientific
literature, or at least to separate the important
from the unimportant? Are more surveys, review
papers, and monographs needed? How should education systems and information retrieval systems be
kept flexibly responsive to the increasing segmentation of scientific knowledge?
These and other issues must be identified so
that guidelines to future research may be
developed but further pursuit of the matter is
inappropriate here. So far as information retrieval itself is concerned, it will be assumed
from this point on that things are bad enough
and ought to be improved, regardless of the rate
at which they mayor may not be getting worse.
The conceptual, rather than equipment, aspects
of the subject will be emphasized.
Mooers' Law
Mooers 3 has postulated that "an information
system will tend. not to be used whenever it is
more painful and troublesome for a customer to
have information than for him not to have it."
By way of explanation, he observes that the
penalties for not being diligent in library
research are minor if they exist at all. Clearly
a greater amount of recognizable and rewarding
(even though duplicative) work can be accomplished if time is not spent in the library.
Effective and efficient information retrieval
systems have been installed and then removed for
lack of use; the customer's ability to retrieve
information outstripped his motivation. It is
suggested in effect that the diligent finding and

240

6.1
use of information should be rewarded and failure
to do so punished; in these circumstances we can
expect better retrieval systems.
Though Mooers may have overstated the case,
his observations are provocative. If this state
of affairs indeed exists, and certainly it must
to some extent, then the question of remedial
action is difficult and important. In addition
to the concepts of IIrewardll and "punishment",
two additional directions for improving "motivation" seem to be promising.
First, more attention ought to be paid to
active dissemination of scientific information
rather than passive interment in libraries wherein
the resurrection of information requires initiative and ingenuity on the part of the customer.
Perhaps the scientific community is in need of
something akin to IIscreening panels" whose job
is to direct newly acquired information to the
appropriate potentially interested users.
Secondly it is suggested that beyond, after,
and independently of "information retrieval"
per se, the proper presentation and display of
information to the customer may play an important
role in the palatability and digestibility of the
retrieved information. The extension of present
information retrieval boundaries in this direction
will be discussed later in the framework of research trends and goals.
Birth Control of Technical Reports
Scientists are painfully aware that much, if
not most, scientific literature is either so
repetitive,verbose, or of such low quality, that
it ought not to have been written in the first
place. In the opening address at the 1958
International Conference on Scientific Information,
Sir Lindor Brown made a sparkling appeal for the
exercise of res~raint on the part of both authors
and publishers.
In professional journals, quality of information is at least partly controlled
through editorial sanction, but the now comparable
volume of unpublished technical reports are totally immune to such controlling influence. Dwight
Gray5 pinpoints the lack of bibliographic control
as an especially unfortunate aspect of technical
reports.
It is doubtful that any system of "informat:iDn
birth control" can be made practical, particularly
one that appeals to voluntary self restraint; involuntary control however may be . a cure worse than
the disease. "Quality Control" applied more'
liberally may be a part of the answer, but any
connotation of "control" raises the question of
information suppression without competent and
diligent exercise of the judgmental function.
"Quali ty Identification 11 is suggested here as a
more appealing approach. Clearly in a situation
in which the volume of information exceeds the
individual scientist's digestive capacity, some
method of assisting him to discriminate between
the important and the unimportant literature can
be counted among the most crucial of requirements.

The referee system for professional journals
imposes a modest degree of quality control on
scientific information. It is at least thinkable
that recognized leaders in the various technical
fields could be imposed upon to implement a
"quality identification" system on a sizable
scale. However, a more practical technique
immediately at hand may lie in the idea of
"citation indexing." It is not unreasonable to
measure the importance of a published article or
technical report by the frequency of subsequent
citations to that article or report (possibly
taking into account who did the citing and why.)
This suggestion was ably described by garfield
in an article published six years ago.
Experimental studies of citation patterns in physics
literature has been carried out by Kessler. 7
The idea of automatic abstracting has enjoyed considerable publicity and popularity during
recent years as an approach to information compression. Techniques for so dOing have somehow
managed to come into being without the benefit of
any adequate description of requirements. The
situation can be rather easily understood by
first considering one plausible elementary rule
for "automatic abstracting" (or rather "automatic extracting"): "Select title and first
sentence of each paragraph." This rule is
obviously mechanizable. Furthermore, at the time
of this writing, no other rules have been demonstrated on the basis of which a convincingly
better abstract can be produced. The problem lies
in the fact that adequate measures for the
"quali ty" of an abstract (human or machine) have
never been developed. Ideally, of course, one
would like to cut the length of an article drastically without appreciable loss of information,
but unfortunately it cannot yet be demonstrated
that the percentage of information loss is
Significantly smaller than the percentage of text
reduction. The question of a need for text
reduction even at the expense of information loss
is nonetheless open.
An interesting and constructive approach to
"volume reductions is contained in a recent Case
Insti tute report, a part of which concerns the
effect of condensation on reader comprehension of
journal articles. If indeed one can somehow
single out "measurable" factors in the area of
reader comprehension, then certainly experimental
observations of this kind are crucial to questions
of redundancy elimination and abstracting.

"Communication Habits of Scientists"
The scientific community itself has been the
object of attention of a number of "opinion
surveys" and scientific communication llbehavior
studies." These studies represent a commendable
attempt to describe characteristics and attributes
of our present system of scientific communication.
A comprehensive review of several dozen such
studies has been reported by Menzel. 9 The stated
purpose of this review is to display the variety
of research that has been done on the flow of
informationamongscientists; it does not attempt,

I

241
6.1

however, to evaluate or recapitulate that which
is worthwhile and deserving of future emphasis.
The total amount of information of this kind
presently available tends to cause mild indigestion when one attempts to assess the relevance of
the whole matter to the implementation of practical improvements in scientific communication. In
their present state, however, these "behavior"
studies are intended only to be descriptive rather
than diagnostic. It is premature to attempt to
translate a description of present behavior into
future requirements; to discern any relationship
between the two requires considerable further
study.
certain fragmentary results of a rather
general nature are worth noting, however, since
they have an interesting bearing upon some of
the questions raised here. Let us consider a
few excerpts from the Menzel review.
"The amount of time chemists devote to
reading on the job was found·directly related to
the ease of their access to scientific literature.
Reading time was directly related to the availability of journals at chemists' desks, to the
location of library facilities in the chemists'
bUilding, and to the existence of company library
facilities."

of current literature for news rather than
reference reflects a requirement or whether it
simply reflects the fact that current literature
information systems are inadequate for reference
purposes, and hence can be used only for news.
This identification of problem areas at
least suggests action in two directions. First
since the stimulant effect of scientific publication is known to be important, steps to enhance
such stimulant capability can be taken. If this
idea were coupled with "quality identification"
we might reasonably equate "quality" with
"stimulation ability." Once stimulating material
has been identified, provision could then be
made for wider circulation within the scientific
community. The conclusion that current literature
is not really used very much for reference, if
correct, is dramatic in its implication of a
requirement for a profound look at and perhaps
overhaul of our literature reference systems.
The spectrum of behavior studies reviewed
by Menzel does not apparently include any reports of active controlled experiments which
exert a perturbing influence on the scientific
communi ty • Experiments with a fully controlled
model system have been suggested by Kessler,7,lO
and seem worthy of serious consideration.
The Conceptual Nature of Information Retrieval

"'one of the greatest stimulants to the use
of information is familiarity with its source.'"
"Sometimes pieces of work which have been
ignored by the scientific community prove to be
highly significant when someone finally stumbles
upon them in the back volumes ••••• It is suggested
tentatively that it is often necessary to publicize information repeatedly, lest it fail to enter
the stream of communications which will lead to
its ultimate user. From the point of view of the
consumer of the information, it seems sometimes
necessary to be exposed to the information
repeatedly before it will make an impact."
"On the basis of the inforr.lation on chemists'
reading time and on the number of articles abstracted in Chemical Abstracts in a given year,
it was concluded that only one half of one percent of the articles published in chemistry are
read by anyone chemist."
"a large portion of articles read and considered useful have been met with by chance;"
" 'there is thus a good deal of circumstantial evidence for the hypothesis that the
literature is used very much more for news than
for reference. ' "
The foregoing conclusions are provocative
and clearly have some bearing on the issue of
"literature search" versus "specific-information
retrieval," as well as on the matter of Mooers'
Law. It should be realized that the results as
they stand shed no light on cause-effect
relationships. It is not clear whether the use

All approaches to information retrieval are
reducible to a common set of elements, both from
the point of view of problems and solutions.
The objective of any information retrieval system
is to permit the originator of information to
communicate with an unknown user at some unknown
future time. We may presume that the user or
customer is faced with a large volume of information, i.e., a library which by many orders of
magnitude is too large to permit an exhaustiye
direct examination search to meet whatever
requirements for information he might have. Thus
communication from originator to "user" must take
place· via some abbreviated "representation" of the
contents of the library. This representation may
take many forms, for example: classification
systems, alphabetic subject indexes, coordinate
systems, facet classification systems, full text,
and others yet to be invented. They all incorporate certain common problems, and these problems have a common origin.
When any document is indexed, cataloged,
classified, or otherwise tagged, this process
constitutes an attempt to represent "information
content" of the article or document in a hignly
abbreviated forin. This "representation" is
intended to serve the purpose of information
retrieval but contains only a minute fraction of
the information contained in the document. Now
it is not usually supposed that the "representation" should reflect the total "information
content", yet at the same time there is no solid
theoretical reason that the purpose of information retrieval can be fully served in any other
way. Fortunately, the whole matter need not rest

242

6.1
on theoretical proof since the objective at hand
is of a purely practical nature.
As a practical matter, the encoding of
"information content" or IImeaning" of a document
in any highly abbreviated representation is an
intuitive process not amenable to formal description. Experiments have indicated that the
reproducibility of the intuitive judgment of
indexers or catalogers is relatively low. A
subject heading, index term, or descriptor, has
a IImeaning" which is a function of the vie'WpOint
of the observer or equivalently a function of the
context in which that heading, term, or descriptor is utilized.
The effectiveness of an information retrieval
system depends on the success with which the indexer and user can pursue the following strategy.
In creating an lIabbreviated representation II of an
article being indexed, an attempt is made by the
indexer to foresee essentially all ways in which
some unknown user might ~sh to recover the
information. A rather high degree of redundancy
is a result of a deliberate attempt on the part
of the indexer to predict the customer's viewpOint. Similarly the search is conducted on a
redundant basis since in effect the user hunts
in all likely (and in some not-so-likely) places.
Both indexer and user must therefore play a
"guessing game ll to ascertain the vie'WpOint of
the other. This game results in redundancy, but
not so much so that large quantitites of retrieved irrelevant information thwart the purpose
of the search. The foregoing description applies
to a successful information retrieval system.
The fact is, it is possible to find many systems
which seem to have no elements of such guessing
or redundancy and which, at the same time, are
not very successful.
Consider for the moment a retrieval system
in which the library IIrepresentationll is based
solely on key words or descriptors. For this
Situation, the llguessing game", played by both
the indexer and searcher, can be aided by means
of a thesaurus. For our purpose here, we define
a thesaurus to be a collection of groups of words
wherein the various members of a group tend to
mean about the same thing for purposes of information retrieval. These word groups thus serve
as a reminder list to assist the indexer and
searcher in conjecturing how the same idea might
be expressed in many ways. (Experimental use of
a thesaurus is described for full text search in
references 11, 12.) In one form or another,
something akin to a thesaurus must be implicitly
present in any good retrieval system. The IIsee
also" portion of an alphabetic subject index or
of a classification system can be looked upon
as serving the purpose of a thesaurus.
The total amount of desirable redundancy in
any retrieval system, and the detailed manner in
which redundancy should be distributed between
indexing and searching, depends upon economics
and on the question of irrelevant retrieval.
For a system in which the encoding process is hi~

ly efficient, redundancy may be cheaper to come
by as a part of the input rather than as part of
the output. In other systems, however, such as
those in which the search workload is light, it
may be that redundancy can be purchased more
cheaply in the searching process.

An important problem area, more or less
distinct from questions of meaning, has not been
covered in the foregoing discussion. This area
concerns syntactic relationships among the various
index terms that might be assigned to a given
document. The effect of failure to utilize such
syntactic relationships is to increase the retrieval of irrelevant material but with no loss
of that which is relevant. To illustrate, if the
phrase "the effect of radiation on mutation" is
an appropriate description of the content of an
article, then the independent aSSignment of
descriptors corresponding to "radiation" and
I1mutation" would not by itself preserve the
syntactic relationships between these two terms
implied by the original phrase. Irrelevant
retrieval of documents in which the descriptors
"mutation" and '!radiation" were simply included
among a number of other descriptors in such a
way that those two were not related to one
another, could clearly result. The degree to
which this whole question is of importance in
terms of the amount of irrelevant retrieval that
might be caused by ignoring syntax has not been
experimentally established. In principle though
the problem is of considerable importance and a
number of approaches to its solution should be
mentioned.
First of all, a considerable amount of work
has been in progress for some years at the 13
University of Pennsylvania under Z. Harris.
The objectives of this effort are broader and
more fundamental than the question of preserving
syntactic relationships among index terms, but
the subject is nonetheless of considerable relevance. Eugene Wall14 ,15 has presented a lucid
categorization of information retrieval problems
in terms of vie'WpOint, generiCS, semantiCS, and
syntax. He describes the use of "role indicators'~
as a solution to problems of syntax, and claims
that effective and successful technical information systems which employ only twelve such
indicators have been in operation.
So far as full text searching is concerned,
an especially simple substitute for syntax,
namely, the "proximi t~" of terms has been suggested by the author l • This approach depends
on the fact that two terms which appear in the
text of an article in reasonable proximity to
one another (i.e., within a few sentences) have
a high probability of being syntactically
related.
"Effectiveness ll and "economy II can be identified as the two fundamental objectives of information retrieval research. "Effectiveness II has
to do with how well a system works, in terms of
both the percentage of relevant material retrieved and the accompanying amount of irrelevant

243

6.1
material. "Economy" of course refers to the cost
of operating the system, including indexing,
storage, file maintenance, and searching. In
particular, the "cost" to the user of reading
irrelevant material must be included. (Otherwise
the distinction is subject to total confusion by
the existence of a hypothetical library system of
100% effectiveness and essentially zero cost.
This library is an unorganized, unindexed, unmanned warehouse which the customer reads through
from beginning to end for each information request.)
Any research project on information retrieval
should be assessed in the light of its objectives,
and the division of objectives into the two categories of effectiveness and economy usually provides a perspective not otherwise apparent.

for its final justification on overall economy of
the resulting system. Up to the present time
large scale systems based on the merger philosophy have not been conspicuously noted for
economy and simplicity (to risk a serious understatement). (That fact was of course known from
the deSign stage on--what's worse though, at
least some of these systems after installation
have been victims of a rather severe onset of
Mooers' Law). The question of mechanization in
libraries is full of economic pitfalls, but not
to the extent of preventing soundly designed
systems to be implemented with present hardware.
Since this paper is intended to focus on conceptual problems, the matter of eqUipment will
not be further pursued.

The issue of mechanization is, of course,
tied to economy. It is obvious, though sometimes
overlooked, that computing machines cannot in
principle improve the "effectiveness" of anything
at allover what can be done with people, unless
response times or environmental factors are of
importance. The basic question is one of economy
alone. Speed itself may be of some relevance but,
in principle at least, one can duplicate the speed
of any automatic process with enough people working in parallel. Admittedly this argument is
oversimplified (though the remark on response
times and environment is a strong hedge), but
justified as a counteraction to the unfortunately
prevalent belief that mechanization of a process
that doesn't work to begin with will improve
matters.

A considerable amount of reported work in
the information retrieval field is indirectly
addressed to the question of economy rather than
effectiveness though not in an obvious sense and
not necessarily in connection with mechanization.
A number of ingenious logical-mathematical models
of information retrieval systems have been devised.
Many of these papers have interesting implications
on questions of file org~ization and search
strategies, (e.g. Estrinl~, Moore 19 ) and possibly
on the design of future machines; their promise
for leaoing to new insights with respect to the
more basic semantic problems that lie at the core
of "retrieval effectiveness" is less clear.

Apart from its intimate relationship to
economy, one other especially important factor
in mechanized information retrieval should be
recognized. The machine handling of the physical
contents of a library involves problems totally
different from the machine handling of a representation which permits that library to be
searched. Present general purpose computing
eqUipment, provided it is used with some ingenuity, is not badly suited to machine search of
representations, i.e., index and catalog type
information. So far as handling the total information content of a library is concerned, the
required storage jumps several orders of magnitude, and existing computers are largely inappropriate. The output of computers which handle and
search only an index-representation is necessarily
just a bibliographic listing of responsive documents (accompanied possibly by titles or brief
abstracts). Retrieval of the document itself
(from shelves, filing cabinets, or microfilm
storage) must then follow as a second stage process. Special purpose retrieval machinery can
be relatively efficient for this second stage,
though generally speaking well designed manual
filing systems offer serious economic competition.
Some equipment designs have been based on a
merged record (usually recorded on i'i1m) of the
machine-readable coded representation of a document with a non-machine-readable full text document image. This merger imposes awkward machin~
design problems from the beginning and must depend

It is at least plausible, however, that some
of the mathematical models might eventually lead
to fruitful results in a more fundamental sense.
Mooers 20 has developed a model in which certain
mathematical properties of different types of
retrieval systems can be compared. It would seem
useful to extend a model of this kind to include
formal representation of redundancy, and then to
investigate the relationship between retrieval
effectiveness and redundancy. In the terminology
of Mooers' model, the transformation relationships between the space of all retrieval prescriptions and the space of all document subsets
should be formulatable in such a way that the
effect of redundancy is clearly brought out. It
would be overly bold to predict at this point
whether such an approach would yield any new
fundamental inSights, better file organization
and storage strategies, both, or neither.
A key element in understanding the "nature
of information retrieval" must certainly be the
types of questions posed by users of information
systems. It is in fact far more reasonable to
design library representations on the basis of
the way in which the users tend to organize the
subject matter rather than the way in which
indexers imagine that it ought to be organized,
yet it seems that this procedure is seldom
fOllowed. This approach is embodied in a study
carried out by Herner2l on the information system
of the U.S. Atomic Energy Commission. (Several
interesting results with immediate practical
implications, so far as machine design is concerned, were obtained. Ninety percent of the
questions involved three or fewer distinct

244

6.1
concepts and of all of the multiple concept
questions, ninety-eight percent involved logical
products rather than logical sums or differences.)
In connection with the Herner study, it would be
interesting to know the extent to which the user's
questions were predicated on some prior notion of
how the AEC library was organized; that is to say,
did their questions really reflect what they
wanted or were they conditioned by what they
thought the library could provide?
How Effective are Present Information Retrieval
Systems?
It is a curious fact that the above question
is essentially unanswerable in terms of any
objective "measurables." For any given information
retrieval system, those concerned with it are
generally not at a loss for an opinion as to how
well it works, but such opinions are seldom
accompanied by evidence. Some recent experimental research has been seriously addressed to
this point, and insight is being acquired. Large
scale experiments which attempt to measure retrieval effectiveness and at the same time compare
various factors crucial to such effectiveness, are
presently under way (under the direction of
Cleverdon) at the College of Aeronautics, Cranfield, En~~and. Preliminary results have been
reported.
Within the framework of a small
scale experimental system, considerable attention
is given to the question of defining measures for
retrieval effectiveness in the author's investigation of text searching. ll The results of several
retrieval methods are compared with Itdirect examination" of the whole document collection.
The emerging results reported in these two
recent studies seem to be taking on a rather
interesting pattern. The ASLIB Cranfield project
had for its objective the measurement of three
variables: the indexing system (four specific
systems were compared), the time spent indexing,
and the experience or qualifications of the indexer. Preliminary results presented by Cleverdon
in a recent seminar1 7 indicate that it doesn't
much matter which indexing system is used, how
much time is taken in the indexing process, or
whether the indexer did or didn't have a lot of
experience. Nothing seemed to depend critically
on the searcher either. Similar "invariances"
were encountered in the reported text searching
experiments. ll
On the whole retrieval effectiveness in these
experimental systems was relatively low (Cleverdon
reports recoveries in the neighborhood of eighty
percent but it should be noted that these figures
refer to It source " documents--Le., those which
directly inspire the retrieval question. Retrieval
percentages for non-source documents have not yet
been reported for that project. The same distinction between source and non-source documents
was used by the author, 11 and it was found that
recovery of source documents was about twice as
effective as that for non-source documents.) In
general these experiments seem to indicate
mediocrity of retrieval effectiveness and

insensitivity to parameters that might reasonably
be supposed important. Though further analysis
is necessary before firm conclusions can be drawn,
certain hypotheses to explain the observed mediocrity and insensitivity can be suggested. If we
suppose that the thesaurus used by the author and
a "see also" structure of the indexing systems
tested by Cleverdon were about equally primitive
(insofar as covering any substantial range of
language redundancy is concerned) then one would
have expected the results to come out about as
they did. It has already been found that a substantial portion of the text searching ineffectiveness could reaso~~bly be,attributed to a
deficient thesaurus.
An interesting sidelight brought out by
Cleverdon in the NSF Seminar is that there presently exists no book or well organized doctrine
on how to compile subject heading lists nor does
there exist a systematic body of opinion on the
application of Itsee also references. Cleverdon
observes that the latter particularly tend to
be applied in a haphazard manner. With inadequate
thesaurus groups, cross references, and haphazard
Itsee alsol! application, it should come as no surprise that the indexer and user find difficnlty
in communicating via an information retrieval
system.
lt

A few remarks can be made on the state of
automatic indexing. It has been pOinted out that
the language redundancy problems associated with
retrieval effectiveness must be approached through
thesaurus-like compilation techniques. On the
basis of all present evidence, there is no
particular reason to believe that the process
is any more difi'icult i"or automatic indexing than
it is for human indexing. The more subtle aspects
of semantics and redundancy in a system based on
automatic indexing must in that case be thrown
into the search process, but with no obvious disadvantages. The ~eal question is not so much
whether automatic indexing can be made to work
(though to be sure the matter must still be left
open) but whether it can be made economical.
Generally speaking, with present computing equipment it cannot, except for certain special
applications where input costs can be amortized
in other ways. Engineering achievements in the
area of direct printed page input and higher
speed memory readout, may hold the final answer
to this question.
Trends and Goals
The nature of the basic problems of information retrieval is such that no sudden conceptual
breakthrough seems likely. Current inventories of
mathematical theories and techniques are applicable
to information systems only to a limited extent.
The tasks that clearly lie ahead must include
large scale language stUdies and laborious experimental investigations. The past several years
have seen increasing awareness of the significance
of some kind of thesaurus approach to problems of
multiple meaning, viewpoint, generics, and redundancy. This approach is not limited to natural

245

6.1
language search techniques or to key-word descriptors, but finds its counterpart in the "see
also ll cross references of all types of subject
headings and classification systems. These compilation efforts which in effect attempt to build
"structures of relatedness II in indexing and retrieval systems should be closely tied to studies
of the types and forms of questions which users
ask of the system. Studies of users' questions
should be patterned not only after the kind that
are presently asked (e.g., Herner 21 ) but should
be addressed also to the kinds of questions Which
users ought to ask if they had different and
better information retrieval systems.
Retrieval techniques work best when they
deal with a limited subject area and are tailored
to the requirements of a limited group of users.
Working systems of this kind should form the
nucleus for experimental investigations so that
deeper inSights can be obtained into the question
of how and why they behave as they do.
Syntactic studies should continue though it
may be anticipated that their practical relevance
to problems of information retrieval may not
materialize to the degree that many workers presently hope. It seems to the author that problems
of semantics tend to dominate practical requirements. It is a more or less obvious phenomenon
of language that essentially the same concept
can be expressed in many dozens of ways which
are not syntactic transformations of one another
and this fact alone suggests that syntacticians
may be eventually disappointed in the extent to
which their work finds practical application.
Storage and retrieval of condensed or IIkernelized ll
sentences 13 suffers from the same illness as do
other methods of automatic abstracting; at present
there are no acceptable measures for the amount of
information loss in the kernelization process.
This entire area may well emerge as being of
considerable future importance however if for no
other reason than the keen interest of a considerable number of competent workers; it holds much
potential for taking off in new directions.
The presently primitive state of the automatic indexing art was mentioned briefly in the
last section, and it is. clear that studies of
this kind shall and should continue. It is in
this area that advances in ~quipment capabilities
are urgently needed. A portion of these studies
should be addressed to the interplay between
natural language queries &ld mechanized information retrieval systems. The formulation of a
request in natural language for computer proceSSing is not subject to the current limitations
of high volume storage that natural language text
searching involves.
The extent to which one should be sanguine
about continued studies of the information
gathering habits of scientists, as though they
were a colony of bees or ants, is not really clear.
Certainly a rash of these studies has broken out
during the past four years or so and many of the
results have been interesting. It j.s possible

that these have largely run their course and that
now emphasis ought to shift to experimental studies
that are subject to greater control. Certainly
more investigation should be made of the obviously
observable parameters within the science communication sphere such as those provided by the technique of citation indexing. Other techniques
exist for "hooking" documents together by purely
pragmatiC and intuitive estimates of relevance
within the framework of some particular purpose.
A set of citations as useful "relevance hooks"
would be based on an assumption that co-citation
implies similarity of meaning from the point of
view of the author dOing the citing. In an
analagous sense, Fan022 suggests that the recogni tion of similarity of two documents in the
process of request formulation (i.e., the requestor asks for a document "like" one he already
has) provides relevance hooks similar to those
provided by co-citation. Networks of relevance
chains once assigned are susceptible to highly
efficient machine manipulation and have the added
virtue that the profound questions of semantics,
viewpoint, ambiguity, generics, and syntax, are
involved only to the extent of directly observable
external effects resulting from a large collection
of instances in which professional judgment is
applied to the question of "relatedness."
It was mentioned earlier in connection with
Law that extending the boundaries of
information retrieval into the area of presentation and display may provide an approach to
solVing problems of information indigestion on the
part of the user. Strictly speaking, an information retrieval system has served its purpose
once responsive documents have been delivered to
the customer, but if indeed the customer's attitude
is characterized by indifference, to th~ extent
that the retrieval system is not effectively used,
then we might surmise that a highly important and
critical problem area has been ignored. This area
is one which begins where information retrieval
ends. Let us imagine that a stack of several
dozen documents of several thousand words each has
been retrieved. Now the customer's original
motive in re~uesting those documents must be examined. If he seeks specific pieces of related
information, a considerable effort on his part may
be required to extract what he wants. If that is
indeed the case then further machine proceSSing of
such retrieved data may serve a useful purpose.
It is of particular significance to note that
computer handling of the total retrieved text (in
contrast to the total text of the library) is
reasonable in terms of storage requirements. His
original requirement may of course have been of a
different kind, such as acquiring general familiarity with a subject area, but we shall consider
further only that situation which calls for specific small fragments of information to be brought
together and "correlated" in some abstract intellectual way. This type of requirement typically
occurs in a business or military intelligence
application. The question now is whether or not
the proper presentation, rearrangement, and display of numerous fragments of retrieved data may
play a significant role in stimulating the user to
Moo~rs'

246
6.1
perceive relationships that are not otherwise
apparent in a more laborious process of directly
examining the total contents of the retrieved
material. Experiments carried out by the author
and co-workers at Ramo-Wooldridge during the last
year or so have demonstrated that such stimulation
is indeed possible. (A report on this work is in
preparation.) When combined with one 0f the conclusions quoted in the Menzel review, 9 namely
that the periodical literature tends to be used
more for "idea stimulation If than for reference,
these concepts of presentation and display take
on added significance. It is suggested that
future information retrieval work may exhibit
an interesting trend in the direction of autmatic user stimulation.
The articles cited in the follOwing list of
references themselves contain references to 160
other articles (probably some are duplicative).
If succeeding generations of citations are
counted, one wonders whether a closed system will
be encountered (it seems likely) and whether
subsets of those cited many times are in some
way more "central" or "important" to information
retrieval.
REFERENCES
1. Scientific and Technical Information as One
of the Problems of Cybernetics by G.E.Vleduts,
V.V.Nalimov, N.I.Styazhkinj SOVIET PHYSICS
USPEKHI, Vol. 2 No.5, p 637, Sept.-oct. 1959
2. The Real Problems Behind Information Retrieval
and Mechanical Translation, Y• Bar-Hillel , Seminar
Feb. 13, 1961, Sponsored by the Office of Naval
Research
3. Information Retrieval Selection Study, Part II:
Seven Retrieval System Models, by Calvin N. Mooers,
Zator Company, Report No. RAOC-TR-59-173 Contract
AF30(602)-1900
4. Opening Address, Sir Lindor Brown, ICSI* p.3
5. Technical Reports I Have Known, and Probably
Written, by Dwight E. Gray, PhYSics Today, V.13
No. 11, p.24 Nov.1960
6. Citation Indexes for Science by Eugene
Garfield, SCience, V.122 No. 3159 p.108 July15,'55
7. Technical Information Flow Patterns,
M. M. Kessler, WJCC May 1961
8. An Operations Research Study of the Dissemination and Use of Recorded Information by
Operations Research Group, Case Institute of
Technology, December 1960 (Sponsored by National
Science Foundation)
9. Review of Studies in the Flow of Information
Among Scientists, Bureau of Applied Social
Research, Columbia University, January 1960
(Sponsored by National Science Foundation)

10. Background to Scientific Communication by
M. M. Kessler, IRE Translations, Vol. EWS-3 No. 1
April 1960
11. Searching Natural Language Text by
Computer, by Don R. Swanson, SCience, Vol. 132
No. 3434 p.1099 October 21, 1960 (Sponsored by
Council on Library Resources)
12. Research Procedures for Automatic Indexing
by Don R. Swanson, presented at American University Third Institute on Information Storage and
Retrieval, Feb. 16, 1961, (Sponsored by Council
on Library Resources)
13. Linguistic Transformations for Information
Retrieval by Z. S. Harris, ICSI* p. 937
14. Documentation,Indexing,and Retrieval of
Scientific Information, prepared by Committee
on Government Operations, U. S. Senate, I:ec. No .113
(report by Eugene Wall p. 175-202)
15. Summary of Area 4 Discussion, ICSI*
p. 804 805
16. ASLIB CRANFIELD RESEARCH PROJECT, Report
on the First Stage of an Investigation Into the
Comparative Efficiency of Indexing Systems by
Cyril W. Cleverdon, Sept. 1960 (Sponsored by
the National Science Fbundation)
17. Seminar February 17, 1961, by Cyril W.
Cleverdon, held at National Science Foundation,
Washington, D.C.
18. Maze Structure and Information Retrieval,
Gerald Estrin, ICSI* p. 1383
19. A Screening Method of Large Information for
Retrieval Systems, Robert T. Moore, WJCC May 1961
20. A Mathematical Theory of Language Symbols
in Retrieval, by Calvin N. Mooers, ICSI* p.1327
21. Determining Requirements for Atomic Energy
Information from Reference Questions, by Saul
Herner and Mary Herner, ICSI* p.181
22.

Summary of Area 6 Discussion, ICSI* p.1407

*ICSI : Proceedings of the International Conference on Scientific Information National Academy
of SCience, National Research CounCil, Washington,
D.-C. 1959

247

6.2

TECHNICAL INFORMATION FLOW PATTERNS
M. M. Kessler

Lincoln Laboratory,

* Massachusetts

Institute of Technology

Summary
A study of the bibliographies of a large
number of articles in physics and electrical
engineering indicates that definite patterns exist
for the flow of technical information. Quantitative data are pre sented on the flow of information
between countries, between cultural and functional groups, and between past and present. An
analysis of the numerical data indicates that
these flow patterns are deeply rooted in the dynamics and evolution of scientific thought and
engineering development. The analysis also
discloses that extreme asymmetry exists between journals in their capacity as carriers of
scientific information.

This study is concerned with the literature of physics and closely related fields of application. Data are presented on the flow of information across political boundaries, from a
basic science to an applied technology, from
past to present, and concludes with some remarks on the efficiency of various journals as
carriers of information. The results are then
discussed in terms of their application to the
design of a system for scientific retrieval and
communication.
Num.e rical Data
Communication across Political Boundaries

Introduction
Every project in the field of information
retrieval must face up to the following question:
"Will it promote the flow of meaningful information from originator to consumer? II No idea,
scheme or component, no matter how intellectually clever or technically elegant, has any worth
except insofar as it contributes to an information
flow system. This flow system has its carriers,
channels, sources, and sinks. It defines the
coupling between individual scientists across
field boundaries, political groupings, time, and
habits of tradition. It is clear that any new
scheme or component of information retrieval,
in order to be effective, must mesh into the flow
pattern and be properly matched to it. And yet
we know remarkably little about the information
circuits as they now exist. 1 This paper attempts
to map the flow of information by analyzing the
statistics of reference s that authors include in
their published papers. It is assumed that the
published journal article is the message unit and
that its citation by an author is recorded evidence
that the message has found a meaningful target.
The strength of this method is that within its
area of limitations it is quantitative and unambiguous. It doe s not depend on subjective opinions
and questioners, and very significantly, a wealth
of recorded data exists in journals of all countries, all sciences, and as far into the past as
we care to go. The weakness of the method is
that it certainly does not measure all scientific
communication. Other modes and circuits exist
that are not reflected in bibliographic citations
and conversely some citations may be irrelevant
to the information transfer process. The method
is nevertheless particularly applicable to the
problems of retrieval systems because the latter
is largely concerned with communication through
written papers.

A number of journals were analyzed for
purpose of obtaining a rough measure of the flow
of scientific information across national boundaries. The bibliographies in the indicated journa
were sorted on the basis of country of origin.
Table I shows the results for the January 1957
issue of the Physical Review.
Table I
Geographic Distribution of References
in the Physical Review (January 1957)
Reference to
Physical Review
Other American
British
European
Russian
All others

No. of
References
994
558
198
258
28
33

% of
Total
48.0
27.0
9.5
12.4
1.4
1.6

We see that roughly half of the references
in the Physical Review are to papers published in
the same journal. Three quarters of all the references are to American journais. The remaining 250/0 of the references are distributed among
a variety of European journals, mainly British.
Note in particular the vanishing call on Russian
reference material (1.40/0).

*

Operated with support from the U. S. Army,
Navy and Air Force

248

6.2
We now consider the same process as it
operates on Russian physicists. Table II gives
the results of a count on the June and October
1957 issues of the Journal of Theoretical and
Experimental Physics.
Table II

JETP
Other Russian
Physical Review
Other American
British
European
All Others

Geographic Distribution of References
in the Physica (1957)
Reference to

Geographic Distribution of Refere~ces
in the Russian Journal of Theoretical and
Experimental Physics (June and October 1957)
Reference to

Table IV

No. of
References

% of
Total

102
201
148
57
63
66
22

15.4
30.5
22.4
8.7
9.5
10.1
3.3

Note that on its home grounds the Russian
JETP actually runs second to the Physical Review. The Rus sian physicists depend on British
and other European literature roughly to the
same extent as do the Americans, but they do
not have the strong partiality to their own chief
journal nor to any combination of journals in
their political group.

Reference to

No. of
References

% of
Total

Nuovo Cimento
Othe r Italian
Russian
Physical Review
Other American
British
European
Others

344
38
84
771
331
223
245
135

15.7
1.7
3.9
35.6
15.4
10.5
11. 2
6.2

225
71
30
249
85
159
121
108

21.6
5.9
3.0
23.9
8.2
15.2
11. 6
10.4

In view of the political and cultural polarization of modern society into East and West
groupings, it is interesting to present the data
of Tables I to IV in such a way as to illustrate
the flow of physics from East to West and vice
versa. Table V groups all references to American and European journals and compares them
with those to Russian and other "iron curtain"
journals.
Table V
The Flow of Information along the
East- West Political Axis
Phys.
Rev.
Western Jnls. 96.9
Eastern Jnls.
1. 4
All Others
1. 6

Geographic Distribution of References
in the Italian Nuovo Cimento (Jan. - June 1958)

'fo of
Total

Physica
Other Netherlands
Russian
Physical Review
Other American
British
European
Others

Similar counts were made on Nuovo
Cimento and Physica, Italian and Dutch journals
of physics. The results are shown in Tables
III and IV.
Table III

No. of
References

Nuovo
Cimento

Physica JETP
86.4
3.0
10.4

90. 1
3.9
6.2

50.7
45.9
3.3

To extend this picture somewhat beyond
pure physics, a count was made on the Journal
of Applied Physics (JAP) and the Proceedings
of the Institute of Radio Engineers (IRE). In
the latter case, we picked January, June, and
September 1957 as representative issues. A
special issue of the IRE devoted to a symposium
on transistor technology was treated separately.
The results are shown in Table VI. The Phys.
Rev. data are reproduced for comparison.
Table VI
Geographic Distribution of References
in Phys. Rev., JAP, and IRE
Proc. IRE
Phys. In. App. Proc. Transistor
Rev.
Phys.
IRE
Issue
ences to
U.S.A.
British
European
Russian
Others

75.0
9.5
12.4
1.4
1.6

70.5
12.0
11.7
2.9
2.9

77.0
11. 6
8.7
1.2
1.4

78.0
5.3
8.6
2.4
1.9

249
6.2
The data suggest the following conclu-

Table VII

sions.
1.
The Physical Review is truly a definitive journal for physicists. It commands overwhelming dominance over all other journals as a
carrier of information between physicists of all
lands.
2. American physics is the chief source
of information, not only for other Americans but
for the international community of physicists.
3. American workers in fields of applied
physics (as typified by authors in JAP and IRE)
find their literature needs overwhelmingly satisfied by American journals.
4.
European physicists draw heavily on
American literature. Their coupling to the Russian literature is not significantly greater than
that of American physicists.
5.
The Western world is virtually selisufficient with regard to physics. The Russian
cultural sphere on the other hand draws heavily
on the We st for its information.
A close comparison of the various tables
suggests that these conclusions are valid even if
we take into account the language barrier between
East and West.

Journal Distribution of References
in the Transistor Issue of IRE
Reference to

The June 1958 is sue of the Proceedings of
the IRE was chosen for study. This issue is a
symposium of 22 papers (353 pages) devoted to
transistor technology. Of the fifty authors,
forty-one indicated affiliation with industry,
eight with universities, and one with the government. Transistor technology is of great interest
to industrial and defense workers and yet it is
new enough to have rather simple and short routes
to the underlying sciences. For these reasons it
was thought that a detailed analysis of the transistor issue would be instructive. Table VII
shows the distribution of references among
journals.

0/0 of
Total

230
129
69
52
48
61
36
22
94

31.0
17 .4
9.3
7.0
6.5
8.2
4.9
3.0
12.6

Physical Review
Proc. IRE
J. Applied Physics
Bell System Tech. Jnl.
British
German
Other Foreign
Rus sian, etc.
Miscel. Jnls. (American)

Table s VI and VII sugge st that the transistor issue of the IRE shows the typical American pattern of bibliographic distribution and doe s
not differ much from a random is sue of the IRE.
This invariance applies only if we consider the
geographic distribution of references. Both
cases follow the American pattern; over 750/0 of
the references are to 'American and over 900/0 to
Western journals. But when we analyze in detail the American journals for the two cases an
entirely different picture emerges. Table VIII
is a break-down of the references to American
journals in the two samples of the IRE.
Table VIII

Flow of Information to an Applied Field
A significant special case of information
flow and retrieval concerns communication across
disciplinary boundaries. We suspect that a retrieval or communications scheme designed to
process physics literature for the physicist is not
the same as what is needed to proce s s physic s
literature for the chemist, the engineer, or the
biologist. A related problem, particularly important in the dynamics of applied research and
development, is the flow of information from
basic research scientists to production engineers.
Ii the coupling is tight and information flows
freely, one may expect a low lag time between
basic discovery and application. A study of certain reference statistics indicates that definite
patterns exist in this information circuit.

No. of
References

Detailed Analysis of IRE References
to American Journals
IRE
Physical Review
JAP
Proc. IRE
Other American

7.3
2.7
23.0
43.0

IRE
Transistor
Issue
25.2
7.5
14. 1
26.3

Whereas the average issue of IRE refers
the ~hysical Review 7.30/0, the special translStor issue has 25.20/0 of its references to the
Physical Review. Thus we see that the bibliographic count is sensitive enough to measure the
degree of coupling between science and technology.
t~

The authors who contribute generally to
the ~RE have a lesser coupling to the Physical
ReVIew than those who are preoccupied with the
new and rapidly developing field of transistors.
T~is analysis of information coupling between
SCIence and technology can be continued another
step. We see from Table VIII that 250/0 of the
references in the transistor issue of the IRE are
to authors of papers in the Physical Review. We
now ask who are these authors? Do they differ
as a. class from the usual authors in the Physical
RevIew? In other words, do the contributors to

250

6.2
the IRE transistor issue make contact with a.
representative group of Physical Review authors,
or is there some transitional group of physicists
who serve as a bridge between the general population of physicists and the industrial group?
Table IX presents the institutional origins
of authors who publish in the Physical Review,
Journal Applied Physics, and IRE. In all cases
column A refers to all authors in a random issue
of the journal. Column B refers to those authors
in the journal who were cited as references in the
transistor is sue of IRE.
Table IX
Institutional Distribution of Authors
in Phys. Rev., JAP, and IRE

acquaintance with the literature of science would
extend the curve farther into the past. Both
hypotheses may be true, namely, workers in
fast-growing, competitive fields may have no
time to search the literature and thus confine
their sources to current material. It is at any
rate clear that this phenomenon has to be understood and absorbed into a serious retrieval
scheme.
The Detailed Bibliographic Structure of a Single
Journal
As a final example of bibliographic statistics, we mention a very detailed study that we
made of the references in the Physical Review.
The study was made for other purposes and involved a complete recording on IBM cards of all
the references in 26 volumes of the Physical
Review.
Some of the results are mentioned
here because they relate to the subject of this
paper. Excluding the unpublished and nonperiodic literature, the authors in the volumes
made 74,599 references to journal articles. Of
this number 45,592 or 60% were to articles in
the Physical Review. The next five most frequently used journals contributed another 13.8%
of the references and the next twelve in order of
frequency contributed 12.2%. Thus 18 journals
accounted for 86% of all the references. The remaining 14% of the references were distributed
among some 650 journals of which 240 were mentioned not more than once in all the 26 volumes
and 420 were mentioned four times or less.

*

JAP

Phx:s. Rev.
Affiliation
University
Industry
Govern:ment
Foreign
Others

A

66
9
13
9
3

B

16.4
74.0
6.3
3.4

A

B

40.3 31.0
31.4 61.0
5.4
17.5
2.7
9.0
1.8

IRE
A

B

25.8 14. 1
44.0 76.8
10.0 0.6
20.5 8.5

The data show that ordinarily 660/0 of
Physical Review authors have university connections and only 9% are from industry. The
sub-group referred to by IRE authors more than
reverses this picture; only 16.4% are university
affiliated and 74% are from industry. Similar
trends are apparent in the other columns of
Table IX. On the basis of the se limited data,
it seems reasonable to assume that the coupling
between basic science and its applied technology
is tighter for the newer technologies and that the
coupling is made through an intermediate group
of scientists who form an intellectual bridge between the university and industrial community.
Flow of Information from the Past
At any given time scientists draw heavily
on the accumulated experience of the past. Indeed, one of the greatest as sets of the journalarticle mode of communication is that it conserves
the record orthe past in an orderly and chronological manner. This coupling to the literature
of the past was studied by plotting the number of
references as a function of .time into the past.
Thus a distribution curve of references was obtained with age as the independent variable.
Graphs 1, 2, 3, and 4 show such distributions
for four journals. The integrated curves are
shown in Fig~re 5. Although the curve s are
orderly, their statistical base is rather limited
and one should be careful with conclusions. One
may speculate that vigorous and fast-growing
fields will show a sharp early rise and level off
rather soon. Another hypothesis, however, may
suggest that a sharp early rise indicates a superficial scholastic approach and that a more basic

Another by-product of this study that is
relevant to our discussion is the following. In
spite of the strong definitive position of the
Physical Review, it is nevertheless true that
any given paper in the Physical Review has a very
low probability of ever being used as a reference.
Indeed, the largest single class of papers never
appears in the reference literature at all. It is
hard to assume that this large group of papers
are never cited because they are worthless.
Other reasons must be sought.
Discussion
The over-all impression left by the data
may be summarized as follows:
a) If we consider the scientific paper as
a message unit and the journal the mes sage carrier, if we accept that inclusion of a paper in a
published list of reference s indicate s that the
message found a relevant receiver, and if we
regard the population of Physical Review author s
as a representative group of physicists, then
there is a massive asymmetry and an overwhelming inhomogeneity in the capacity of the many

*

This work was done in collaboration with Mr.
Frank Heart of the Lincoln Laboratory and will
be published elsewhere.

251
6.2
hundreds of journals to serve as carriers of the
scientific message. The imbalance may be due to
language barriers, cultural and political isolation,
reputation of the journal we have examined and
its availability or to a combination of the se. The
inhomogeneity exists. whatever the reason, and
in its most extreme form gives rise to the definitive journal, such as the Physical Review. The
data raise important questions about the design
philosophy of retrieval systems. In view of this
inhomogeneity, should the retrieval process and
the flow channels be the same for all carriers
or should they take account of the carrier's
capacity? If account is to be taken of the carrier's capacity, should the system be designed
to further reinforce the strong and efficient carriers at the expense of the less efficient, thus
reducing the noise. or should we on the contrary
take the attit1,lde that the efficient carriers need
Ie s s attention than the weak, anQ. therefore concentrate on the latter and raise their efficiency?
Is the definitive journal a desirable phenomenon
or does it in the long run inhibit the communication process?
How should we approach the complex problem of injecting the results of Russian research
into the main stream of American physic s. Such
an injection is certainly desirable, but it is not
clear that a mas sive translation program and
wide distribution of the translated material is
the best way of doing it. Another method might
be to encourage a small number of practicing
American physicists to learn the Russian language and depend on them as the instrument of
injection. Considering that the mOI).ey and effort
available for this purpose are limited, one should
not blandly accept either method.
A retrieval and communication system.
unless its contribution is trivial, will have sufficient feedback to strengthen or weaken the
various elements of the communication process
that now exists. It is therefore important to
understand these elements and to design our
system with them in mind.
b) The previous section concerned communication within the relatively homogeneou&
group of physicists who are in the habit of publishing in the Physical .Review. The problem of
communication across field boundaries or between scientists and engineers is a somewhat different matter. The limited amount of data that
we have collected must be considered as a
sample that only indicates the complexity of the
problem. It would seem that in this case there
is no definitive journal. Furthermore. meaningful communication seems to involve a chain of
intermediaries that form a bridge between pure
science and applied technology. Should a system
be designed to encourage traffic along this chain
of bridges, or should it attempt to short circuit
the chain? That this is not a purely academic
problem is clear from the experience of the
Chemical Abstracts, a major communicative
link between physic s and chemistry. The

Chemical Abstracts attempts to short circuit any
bridging mechanism and bring physics directly to
each chemist by abstracting practically the entire physics literature. This, together with
other examples of short circuitry, has so overloaded its own channel that Chemical Abstracts
is becoming increasingly awkward and bulky.
They could have chosen not to include the physics
literature in its abstracts and to depend on the
various hyphenated journals, such as the In. of
Physical-Chemistry, Chemical-Physics, Colloid
Chemistry, etc. etc., to act as injectors of
physics into the main stream of the chemical
literature. This would appear to loosen the
coupling between physics and chemistry and increase the time necessary for information transfer. But the decrease in volume of traffic could
well compensate for the looser coupling and produce a more efficient system. The phenomenon
of coupling between fields is at any rate important enough to be considered in the design of a
retrieval and communication system.
c) The flow of scientific information in
time. past to pre sent, or the useful half life of
a scientific message unit is another important
element of our problem. Our numbers indicate
a useful half life of some five to ten years. A
rigorous definition and measure of this phenomenon is not easy to come by. But it is obvious
that no system of retrieval and communication
can long survive without some method of purging
its message population from time to time. This
is not just a matter of eliminating poor material
in favor of newer and better results. It is a
curious fact that even the masterpieces of scientific literature will in time become worth1e s s
except for historical reasons. This is a basic
difference between the scientific and belletristic
literature. It is inconceivable for a serious student of English literature, for example, not to
have read Shakespeare, Milton and Scott. A
serious student of physics, on the other hand,
can safely ignore the original writings of Newton,
Faraday and Maxwell. The removal of a scientific paper from the retrieval system should not
depend on a value judgment. The correct criterion should be based on the degree to which the
paper's information has been metabolized into the
flow stream of science. We could perhaps decide that once a paper has appeared in the citation literature a given number of times, it need
no longer be carried as an independent message
unit. This too needs more study.
At the other extreme we have the phenomenon of a large group of papers, perhaps the largest single group, that is never quoted in other
peop1e's bibliographies. If these papers do not
enter the bibliographic literature within, say,
five years of publication, they may become effectively lost to the literature. In view of the
careful editing process, it is hard to believe
that this large group of papers is useless or
redundant. If they are useless or redundant, the
publishers and editors could well afford to review
their processing criteria. On the other hand, a

252
6.2
continuing review of the citation literature over a
travelipg five-year period may reveal a group of
papers that warrant more detailed channeling.
Concluding Remarks
Science and technology are approaching a
in communication. It would be a mistake
to define this crisis entirely in terms of retrieval
problems. Indeed, the break-down in communication between scientists must itself be evaluated
in terms of the maturing sociology of science.
Communication, after all, is not the only aspect
of science that is in crisis. There are problems
of technical manpower shortages, the increasing
cost of scientific research, the growing imbalance between basic and applied re search, and
many others that relate to the emergence of science as a major instrument of national favor,
stature and propaganda. But even if we confine
our attention to the, communication problems only,
we must still remember that no single channel or
mode is likely to solve all our needs. The components must be evaluated in terms of their contribution to the over-all system performance.
Many systems problems must be studied if a working solution is ever to be achieved.
CrlS1S

a) No system can be designed in an
economic and social vacuum. Even the vital
functions of national defense are subject to restraints. We must estimate the social and economic limits that will govern our system and
optimize its function within the se limits. How can
limited resources best be apportioned between
the various segments of the scientific community?
What part of the re source s shall be as signed to
retrieval and other question-answer functions as
opposed to directing the flow of information on
the basis of generalized need-to-know criteria?
What are the information needs within welldefined fields like physics as opposed to the flow
of information across field boundaries?
b) Ii there is a critical failure of communication now, how will we know when improvement has taken place? What test procedures and
criteria of performance can we use to evaluate
the system? Unless a figure of merit can be assigned to the system as a whole, the contribution
of any given component is always in doubt. One
cannot test the performance of a system by
measuring each component separately in terms of
its own parameters. Experience with large multicomponent systems indicates that one cannot
generally arrive at a figure of merit analytically.
It is usually necessary to build a model and evaluate new components in terms of their effect on
the model.
c) To what extent should the system
operate only when interrogated and to what extent should it operate on the initiation of its
built-in logic?

d) Should the system be local, regional,
or national? To what extent can communication
media other than the journal article be exploited
(radio, television, newspapers, remote printers,
etc. )?
e) What is the probable cost of various
systems, both initial capital investment and
operating cost? Shall it be financially selisupporting or shall it be subsidized? Stability of
financing is of particular significance because
the system must have long-range continuity in
order to be at all effective.
f) Finally, we wish to stress that the
problems of science communication are not primarily equipment and hardware problems. Nor
are they primarily problems in indexing, abstracting, or retrieval. The significant problems are
in the area of systems design and systems logic.
At this time there is no single organization whose
professional competence and involvement embraces
the entire spectrum of problems. Such an organization is needed if a reasonably successful solution is expected.

Reference
1.

See for example the series of papers in
"Area I - The Collected Papers of the
International Conference on Scientific
Information," Washington, D. C. (1958)

\3-40-4658\
0.20

en

w
u

Z 0.16
w

a::
w

LL

w
a:: 0.12
--I
--I

«

~ 0.08
I-

Z

W

u 0.04

a::
w

0..

o

4

12

8

16

20

24

28

32

36

40

AGE OF REFERENCES (years)

Figure 1

Distribution of References by Age
Physical Review 1957
en N
(J1

N

eN

O"lN

en

Not>.

\3-40-4659]
0.20

en
lJJ

U

~ 0.16

a:

lJJ

1.1..
lJJ

a: 0.12
..J
..J

«.
~ 0.08
~

z

lJJ

u 0.04

a:
a..

lJJ

o

4

8

12

16

20

24

28

AGE OF REFERENCES (years)
Figure 2

Distribution of Reference s by Age
Russian JETP, February, March, April 1957

32

36

40

13-40-4660j
0.20
CJ)

w

u

w 0.16
a::
w

Z

IJ..
W

0:::

0.12

...J
...J

«

~0.08
IZ
UJ

u 0.04
0:::

w

0...

o

4

8

12

24
28
AGE OF REFERENCES (years)

Figure 3

16

20

32

36

40

Distribution of References by Age
Physica 1957
en
tv

N
(J1
(J1

O')N

•

U1

NO')

13-40-46611
0.20
en

w
u

~ 0.16

0:::

w
LL

W
0:::

0.1 2

...J
...J

«

~ 0.08

.....
z
w

U

0:::
W

a.

o

4

12

24
28
AGE OF REFERENCES (years)

8

Figure 4

16

20

Distribution of References by Age
J. Acoust. Soc. Atn., Jan., Feb., March 1957

32

36

40

13 - 40 - 46621
100

~

w

(.!)

i:!Zwen

~

1

~-----~

I

80

wU

uZ

o::w
wO::
a.. w
w~

PHYSICAL REVIEW

60

JETP

>0::

~-'

-' -' 40

::::>«

~lL..

:EO
::::>

u

20

o

4

8

12

16

20

24

28

AGE OF REFERENCES (years)

Figure 5

Cumulative Distribution of References by Age

0')

N

N
CJ1
"'-l

259
6.3

A SCREENING METHOD FOR
LARGE INFORMATION RETRIEVAL SYSTEMS
Robert T. Moore
Data Processing Systems Division
National Bureau of Standards
Washington, D.C.
Summary
This paper is addressed primarily to describing a method for reducing the excessive processing
times sometimes encountered in large information
retrieval systems. Two tools are suggested:
(1) a screening system to allm., multi-level
processing of material and (2) pre-processing
of the files to allow block rejection of documents
not answering a retrieval re~uest.

encountered among "descriptions" of different
documents in the file, which may be applied to
compreSSing and structuring a large file of
retrieval information. Each of these ideas is
abstracted and discussed in a more general
context before being applied to the specific
system under consideration here.
Screening and the Use of Descriptors

A screen using descriptors to constitute the
first level of screening is described in detail,
and a pre-processing techni~ue, based upon
clustering of descriptors, is developed. Two
methods of implementation on high-speed data
processors are described based on serial access
type storage and random access "dynamic storage,"
respectively.
This screening techni~ue is compared for
efficiency against two other familiar methods of
employing descriptors: the ordinary serial
descriptor list method and the inverted file
techni~ue.

Introduction
The primary purpose of this paper is to
describe a method of constructing an initial
screening stage for a multiple-component information retrieval system. The screening system (or
sub-system) utilizes document descriptions
comprised of independent single-word index terms,
herea.fter referred to as "descriptors ll which are
thus to be interpreted in their most restricted
connotation, and applies a method of file
organization and compression which has several
special advantages. Two different methods of
implementation will be discussed, the choice of
method being dependent upon the type of data
proceSSing e~uipment to be used.
The deSigner of a large retrieval system is
faced with two separate but related problems. He
must minimize the expensive machine time re~uired
to process a request for information. It is also
desirable to minimize the total storage, both
internal and peripheral, required for the index
file. The system described here represents an
attack primarily upon the processing time but
some of the ideas introduced show promise of
reducing the file size as well.
In the process of discussing this specific
screening system, several ideas of greater
generality are utilized, among them (1) the
"screening approach 11 , as applied to storage and
retrieval systems, and (2) the redundancy

Advantages of the Screening Approach
It is useful to distinguish between two
alternative approaches to the problems of
information storage and retrieval. The importance of this distinction, it is hoped, will be
adequately illustrated in the discussion to
follow.
One approach is to look upon an information
retrieval system as a IImechanized librarian".
The user explains his problem or question to the
librarian, and is given suggestions as to where
in the library he may find the answer. If his
~uestion turns out to have elicited no helpful
suggestions, he rephrases it and tries again.
The function of the librarian (mechanical or
human) is to compare the re~uest with "stored"
information·about the library and, after
considering various "aspects" of the question,
to make determinations of "relevance", thus
directing the user to the right documents. This
might be described as the searching approach,.
since it involves the librarian's actively
seeking the relevant material.
Alternatively, we may deSign a retrieval
system which behaves in a manner that would be
considered obtuse, at best, in a human librarian.
It can single out all those references which
could not possibly be of any use and remove
these from consideration. It serves as an
information garbage disposal unit, which discards
the definitely useless and presents the
remainder to the user. Since this renresents a
screening out of the undesirable doc~ents (to
change metaphors), it may be called a screening
approach.
Of course, for perfect systems this
distinction is uninteresting. Any document that
"passes through a perfect screen 11 will be "found
by a perfect search" and vice versa. What is of
interest is the way in which an imperfect
searching system fails, as contrasted with the
comparable problems in an imperfect screening

260
6.3
system. An adequate searching system will only
rarely "find" documents which are irrelevant, but
it may very well fail to locate documents which
a human reader of all the documents would classify
as pertinent to a request. An imperfect screen,
on the other hand, will not block good documents
nearly as often as it will allow documents to pass
which a human reader would know were irrelevant.
(These may be considered as the defining properties of the two types of systems. However, they
also reflect two different philosophies of system
design. )
In all of the discussion to follow, we
presume that the indexing is such that the systems
~ discussed are imperfect, since no one seems to
have produced a system which cannot be confounded
by the vagaries of "semantics". However, the
problem of human or mechanical error is not to be
considered.

In the National Bureau of Standards-Patent
Office HAYSTAQ project1 , we are finding it
necessary to adopt the screening approach as a
supplement to the searching approach in order to
solve the particular set of problems with which
we are faced.
In the HAYSTAQ program, our more
elaborate and sophisticated programs (topological
compari~on of chemical structure diagrams,
comparison of logical expreSSions, etc.) are too
detailed and time consuming to be run against
every document in files of the size we will
eventually have to handle. Thus the needs of the
HAYSTAQ system point directly to the desirability
of a simple economical screen to reduce the
volume of input to the more elaborate routines.
Such a screen would also be a logical starting
pOint for the construction of new systems. The
main subject of discussion in the remainder of
this paper is a system of that type.

because of the "peripheral" vagueness in the
meaning of such termso An index using these
terms will also tend to become obsolete as
science and technology progress. Ideas and
techniques developed in one field often find
application in seemingly unrelated fields.
One can easily get a list of candidates for
the vocabulary by listing all objects, processes,
and relations that occur in a reasonable-sized
random sample of the documents in the library to
be indexed. The problem is then to choose that
small set of terms which has the greatest "power
of discrimination" among documents in the library.
In calculating power of discrimination, one must
also consider how often terms will be used in
framing requests. (See discussion of measure of
merit in the Appendix for an example.) It is not
our purpose to discuss solutions to this difficult
problem. Suffice it to say that it is closely
related to the problem of character recognition:
What is the cheapest way to distinguish between
centered binary images of the letters of the
alphabet? What bits in the images carry the most
information?
A descriptor set for a document will then be
a simple list of the descriptors applying to the
document (objects, processes and relations
discussed in the document) chosen from the
selected vocabulary. Then when a given request
is being processed, a document will be
screened out, if and only if there is one or more
descriptors in the request that is not in the list
applying to that document. If the screen is to
function in a fail-safe manner, the user must
keep this fact in mind when he sets up the
descriptor list of each request.
File Preparations

Descriptors for Screening

General Method

Although mutually independent descriptors
or index terms have been widely used since the
early years of information retrieval research,
more recently many workers in the field have
begun to question their value for projects
requiring indexing in depth. 2 They argue that
scientific discourse is too complex to allow
adequate indexing by means of unconnected oneword comments on a document. 3 While it is
demonstrable that Simple descriptor methods are
inadequate as a single solution to problems of
deep and exhaustive indexing~ this by no means
implies that they should be rejected entirely.
In fact, their very simplicity adapts them very
well for use in initial screening systemso

Rather than use a file which is encoded in
the language used by the human document analysts
and arranged in the order in which the documents
are prepared for machine storage, it is
altogether reasonable to transform it into a "prescreened" or "pre-searched" file. Such a file
would be organized so that documents could be
processed partially in parallel against a request.
Furthermore, many of the manipulations and
housekeeping operations which are independent of
the specific content of a request can be performed
during the file preparation process in advance of
processing of requests. (An example is the
"inverted file" investigated by Nolan and Firth4
and used in several Patent Office exper1ments. 5 )
It should be noted that such prepared files are
of great utility in either searching or screening
systems, although our attention here will be
confined solely to the latter application.

The key to setting up a nearly fail-safe
descriptor screen naturally lies in the type of '
vocabulary selected. The vocabulary should be
a list of objects, processes and relations
discussed in documents in the file. A
vocabulary which classifies documents ("physics",
"mechanical translation", etc.) or describes
them in general terms cannot be fail-Safe

When an information storage and retrieval
operation is to be carried out, one chooses an
appropriate symbolic It:jIl.guage, which can be
manipulated by computer (English text is one

261
6.3
possible example). Document descriptions are
then encoded in this language. In many cases
there will exist several equivalent symbol
sequences which could describe a given document.
( "Equivalent" should be taken to mean that there
are rules within the formal language which allow
one to transform one expression into another and
the converse.)
From this point on, the symbolic language
will be regarded as fixed. The sequence of
symbols representing each document will be one of
a class of equivalent expressions, i.e., only
transformations of the symbol string from its
original form to an equivalent form are allowed.
One might think of the transformation
P v Q ~ Q v P in the propositional calculus as
an example of this type of transformation. In
this section, no further reference will be made
to the meaning of the symbol strings involved;
only syntactic or formal properties will be
employed.
We then consider possible redundancy within
a file. .It is difficult to give an exact
definition of what is meant by "redundancy" in so
general a context. Reference to one of the
specific symbolic languages is needed in order to
be precise. Basically what is intended is that
if a certain configuration of specific symbols
occurs repeatedly in many document descriptions
in the file, it carries less information per
symbol (in the information theoretic sense) than
does a comparable configuration which occurs only
rarely.
If a "met as ymb ol " or "mOlecular symbol"
were asSigned to each commonly occurring" .
"
configuration (which is composed of the obJect

or "atomic" symbols in the original language),
the file could be in some sense recoded in terms
of these metasymbols. The recoded file would
potentially require substantially fewer bits to
be stored in memory and might be organized to
place documents with metasymbols in common near
one another to facilitate group inspection.
It is helpful to visualize a document
description as a set or as a circle on a Venn
diagram. The metasymbols are applied to intersections or "overlaps" of these sets. (See
Figure 1.) In at least some of the possible
symbolic languages (descriptors, propositional
calculus, predicate calculus, and topological
structure codes), it is possible to develop this
line of reasoning beyond the heuristic level by
imposing an appropriate Boolean algebra on the
symbol strings. The use of Boolean algebra
provides no startingly new information about the
metasymbol techniques, but it serves as a
convenient and compact language in which to
discuss the topic. It permits us to abstract
only those aspects of the situation which are
useful, so that we do not become bogged down in
irrelevant detail.
In files in which there is an equivalence
class of expressions for each document in the
library, the size and frequency of occurrence of
overlaps are strongly dependent upon which
representation is chosen for each document. It
then becomes necessary to discover an algorithm
for finding an optimal canonical represent~tion
for the file, i.e., one which maximizes the
overlaps and selects a unique representative for
each document. Fortunately, in the descriptor
case there exists a unique representation for

c

Regions I - VII correspond to metasymbols
Figure 1

262

6.3
each document, so the problem is entirely avoided
here.

will now describe a recoded file using
descriptors and see what kind of information
such questions elicit in this particular case.

utility of Metasymbols in Screening
A File structure for Descriptor Screens
A set of metasymbols for an index file (a
file of symbol strings for a library) is of some
intrinsic interest, since it is likely to
correspond in many cases to clusters of ideas
that are discussed together in the literature.
Such "co-occurrence" phenomena may lead to some
interesting information about the structure and
interrelationships of various scientific
disciplines represented in the basic library.
Such possibilities rest upon a number of
debatable epistemological assumptions about the
connection between the content of a library and
the syntactic relations between expressions used
to encode that content. The teeth of any such
arguments must be sharpened on a plentiful
supply of data before they can be expected to cut
into the problems of information retrieval.
Fortunately, the "conceptual significance"
of such metasymbols need not concern us here. So
long as they exist at all (which is tantamount to
saying "the file is not a completely random set
of symbol strings II) we can put them to work in
screening our files. Their utility lies in the
possibility of "block screening". Every metasymbol represents the cornmon overlap of a set of
documents and may be said to apply to each
document in that set. Suppose that when a
request (sequence of symbols) is directed to the
file, comparison of this request symbol string
indicates that no document to which the metasymbol applies could satisfy the request. Then
every document to which the metasymbol applies
can be rejected outright. No further information
about these documents, e.g., which other metasymbols ~pply, what are their identification
numbers, etc. need be examined. In other words,
documents can be rejected in large blocks if
metasymbols are used.
Hence, after discovering a fruitful way of
recoding a file in terms of metasymbols, we
concern ourselves with finding methods of
exploiting this "block rejection" in as efficient
a manner as possible.
In fact, one of the best measures of the
value of meta-recoding a file will lie precisely
in the degree to vrhich it can be applied to the
idea of block rejection of irrelevant documents.
If a document will eventually be rejected in
response to a given request, how early in the
process (on the average) can it be rejected? Is
rejection information used as soon as it is
obtained, or must rejected documents be carried
as excess baggage for a time? If a document is
to be accepted, how much extra manipulation is
necessitated by the recoding? How much extra
IIbookkeeping ll information is introduced by the
recoding?
Questions such as these must be answered
before any proposed system can be evalu8.ted. We

Since our discussion is limited to the,
problem of designing a file structure for use in
descriptor screening, we can develop the set
analogy explicitly and demonstrate its utility,
as well as give a concrete example of a symbolic
language. It will become clear that the
simplicity of descriptors and the application of
the screening approach both contribute to the
workability of the structure.
Initially, the documents are indexed
according to which of the N descriptors apply to
them. However, it is more convenient for
screening to represent them in terms of their
II re jectors II or descriptors-that-do-not-apply.
Also, a binary number notation is often used in
connection with descriptor methods.
The N-word vocabulary is then organized in
some convenient order, e.g., alphabetical, and
an N-digit binary number or Boolean N-vector is
constructed for each document by putting a 1 in
the ith position if the ith descriptor applies
or a-O in that position If it does not. Note
that this is the usual "fixed field" method of
handling descriptors, particularly in punched
card applications. This technique is of great
use here, although the logical complement of the
usual Boolean vector will be used (this is the
vector for the rejector list of the document).
This complemented number will be called the
rejector vector for the document. Since the settheory operations of intersect~on(A A B), union
(A V B), and complementation (A) correspond
directly to the Boolean or logical II and " , "or",
and "complement" as defined on the rejector
vectors, any manipulation of rejector vectors
can be pictured in a suitable Venn diagram and
vice versa.
Using these tools, we are equipped to talk
about the appropriate metasymbols. Giving up
the logician's prefix IImeta" in favor of the
more suggestive chemist I s term "mOlecular II,
henceforth, in this paper, we shall call the
metasymbols "molecular rejectors". (Note that
these behave more like radicals than molecules,
so the chemical analogy cannot be pushed too
far.) A molecular rejector will be the intersection or "and" of some collection of rejector
vectors, that is, the set of l-bits common to
all the rejector vectors in that collection.

Oux purpose here is to describe an
algorithm for selecting a set of molecular
rejectors for the file. In the interest of
efficiency as previously discussed, this
particular scheme defines molecular rejectors
which in many cases "make sense" only for a
subsection of the total file. The molecular
vocabulary then has a highly variegated "local"
aspect, and the different molecular rejectors

263
6.3
are not necessarily disjoint. Nevertheless each
document will have a unique molecular
characterization.
The guiding principle is that of maximizing
the average size of blocks rejected and of
carrying out this rejection as early and as
completely as possible. It is possible to
develop a rigorous theory of the redundancy of a
molecular rejector S, but this is an involved
process. Such an analysis is important in
choosing the method of storing atomic definitions
of molecular rejectors that minimizes storage
requirements. This will not be taken up here,
however.
The quantity of interest is the average
number of documents in a given subfile G rejected
by a given molecular rejector--a measure of merit.
This measure Il (S) will be the product of the
number of documents, t, which would be rejected by
S, t(S,G), and the probability peS) that the
request will overlap with S(l,L(S) = t(S,G)p(S)] .See
Appendix for a discussion of possible approximate
functions for Il(S),
To find a structure for the file
proceed as follows:

F we

Generate all 2N_l possible molecular
rejectors for the file F.(Ingenuity may make it
possible to reduce significantly the number to be
generated. )
1.

2. By appropriate manipulations, compute
Il(S) for each molecular rejector and select the
one of highest Il as Sl' If there is more than
one S of the same maxTImum Il, life becomes
complicated. We could pick one of these at
random for Sl' (See the Appendix for additional
discussion. )

3. Decompose the file F into two disjoint

subfiles, Fl and F2:
a.

Fl contains all documents which
Sl rejects -- all of the t(SlF)
documents having at least the
atomic rejectors, or I-bits, of Sl
in their rejector vectors.

b.

F2 contains all documents of F not
in Fl -- those which lack at least
one l-bit of Sl'

,

4. Repeat the procedure of steps 1 through
3 for F2 to generate S2' (F might be called Fi
for consistent notation.) We then have F2
which S2 rejects and F~ which it does not.
This procedure may be repeated, always operating
on a "primed" file to produce an "unprimed"
file with the same subscript, a molecular
rejector with the same subscript, and a new
smaller "primed" file with its rightmost subscript
one greater than that of the subfile from which
it is derived. It is possible, when attempting
to apply this process, that the primed subfile
will turn out to contain a number of "rugged

individualist" rejector vectors which do not
group well with one another, i.e., all Il(S) in
this subfile are very small, less than 1, for
example. In this case, each rejector vector in
the subfile is given its own molecular rejector
vector, which completes its definition and
produces no new primed ~ unprimed subfile.

5. Apply the following procedure for F
(note that it is slightly different than that
applied for F2 ): Since all Fl rejector vectors
have Sl in common, it is necessary to disregard
these nits in any further processing. This is
accomplished by "subtracting II Sl from all the
rejector vectors, AiJ of Fl (Ai 1\ Sl) before
proceeding. Then steps 1 through 3 may be
applied to the modified Fl generating Sll' Fll
and Fi2' In general, this process, when applied
to an unprimed subfile, generates a new smaller
unprimed subfile with the new subscript q to the
right of those subscripts applying to the old
file (F.fJo
p ~ F-IJo
p),
a new molecular
q
..L,.

•••

..L.

•••

rejector with the same subscript (S-IJO
p) and a
....... CJ.
primed file with its rightmost subscript
increased by 1 (F'ij ••• (p+l»)'
In some cases of application of step 5, the
subtraction process will leave some of the
rejector vectors in the modified file (modified
Fl in the case explicitly considered) completely
empty of I-bits. This means that these have
been completely characterized by the existing
list of molecular rejectors, so the corresponding document numbers, with complete molecular
rejector list, can be set aside for later
insertion into some storage unit. The process is
then continued on that part of the mOdified subfile containing non-empty rejector vectors.
Another possibility is that the newly
generated subfile contains either only one
rejector vector or a set of identical vectors
(applying, of course, to different documents).
In both cases, the remaining rejector vectors
are assigned a single final molecular rejector
and classified completely. Hence no unprimed
file is generated in this case.
The procedure just outlined, when developed
in detail, leads to a complete characterization
of the rejector vector of every document in
terms of molecular rejectors. If Al is the
original rejector vector and Sj' Sjk' ••• Sjk ••• p
are the molecular rejectors assigned to it,
then ~ = So v Sjk v ••• v Sjk ••• p'
Figure 2 isJa rough flow chart for one possible
algorithm for accomplishing thiS, and Figure 3
sketches the "generation tree ll for some file.
At this point, one might set up a
dictionary of a sort, defining each molecular
rejector in terms of the I-bits it contains, and
then store the file described in terms of these
molecular rejectors. Actually the subscripts of
the rejector with the most subscripts which
applies to a document suffices to specify all its

264

6.3
molecular rejectors, as study will indicate.
Such a procedure would lead to no particular
advantage, however. The full potentialities of
block rejection can only be realized if a somewhat more elaborate storage method is used. Two
such methods will be described in the next
section.
The File Growth Problem
In the majority of applications, the library
will not be a static collection of documents, but
one which is continually growing. Hence it is
neoessary to examine the possibility of incorporating the rejector vectors for new documents into the structured index file. The file described
above can accommodate considerable expansion
easily, particularly if new material has the same
basic statistical properties as the original file,
i.e., all of the t,(S,G) in the augmented file are
nearly proportional to those of the original.

It is not difficult to insert a new document
into an organized file of the sort described.
What is required is an algorithm which will give
a unique molecular characterization of the
rejector vector of the new document. If AM+l is
the new rejector vector, AM+l is compared with
Sl,S2. • until the Sj of lowest index which
applied to it (AM+l /\ S.1 = Sj) is located. If
no such S· exists, all of ~l becomes a new
single in~ex molecular rejector. I f an Sj does
exist, then it is applied to the document and
subtracted from the rejector vector. The remaining rejector vector is then compared with
SJ1,Sj2, ••• until again either one of lowest
r1ght index is found which applies or it is
found that none applies. Both cases are treated
substantially as before, and in general the
process is repeated until, in one manner or
another, the new rejector vector is fully
described by a set of molecular rejectors (at
most one of these will be a newly generated one).
In most cases a reasonable number of pre-existing
molecular rejectors will apply, and only a few
atomic rejectors will remain to be lumped into a
new molecular rejector.
Clearly, however, it will eventually become
profitable to reprocess the file and generate a
complete new set of molecular rejectors, because
new accessions will have rendered the old
statistics invalid, and a less efficient file
will have been built up. The file structuring
method outlined here does not really come into
its own until the file is large enough so that
the percentage increase (per year, say) is small,
and the statistics are really representative of
the type of information being inserted. (In
scientific and technical literature, progress and
changes in the areas attracting interest will
tend to cause systematic drifts in file
statistics. )

Storage and Screening Schemata
Conventional Techniques
The usual method of approaching the storage
and search problem consists of storing the index
file on some permanent medium (punched cards,
tapes, drums, discs, etc.), often storing
auxiliary tables such as dictionaries, and
providing an internally stored "search program"
for a high-speed digital data processor. In our
application, this program takes requests as input
and delivers as output the identification numbers
of those documents which'pass screening.
The storage-and-screening method to be
described here is especially well adapted for
tape storage or other storage in which serial
access is necessary. It minimizes storage
requirements by USing tape position to store
much of the information. All molecular
rejectors are defined exactly once on the tape,
and every document identification number is
listed exactly once. Relative position on the
tape determines which molecular rejectors reject
which documents.
In the process, various
marker symbols are used. These may, in practice,
be fields of a few bits in words used primarily
for molecular rejector definition, but in the
discussion to follow they are regarded as
independent entities.
In this storage, one begins with the marker
symbol Me and then gives the atomic definition of
the molecular rejector Sl. It is helpful to
visualize this as the list of atomic rejectors
for Sl' stored as a rejector vector. However,
in general this would be a comparatively
inefficient use of storage space, since Sl
itself is composed primarily of O-bits. The Sl
definition is followed by Mll Sll' ~, S111, M3 ,
etc., until the complete molecular sequence for
one or more documents has been completely
defined. Then a T (terminal) symbol is written,
followed by a list of the identification
number{s) of the{se) document{s).

At this point two different situations may
arise:
1. All of the molecular rejectors which
reject the first document{s) (suppose there are k
of these) may reject more documents, but not
constitute a complete characterization of the
rejector vectors for these documents. In this
case, Mk is written after the identification
numbers of the documents, followed by
Sll ••• ll (k + 1 "ones" in subscript), Mk+l , etc.
until another document is completely described.
2. The molecular rejector Sll ••• l (k "ones ll
in subscript) may apply only to the first
document rejector vector, but the other k-l
molecular rejectors apply to more documents

265

6.3
(in a more general case, this will be
k-p, 1 ~ P ~k-l). Then Mk- l (Mk _p ) is written
after the document identification numbers
followed by Sll ••• 12 (k subscripts in all), Mk,
and so on, until another document rejector vector
is completely described.
This will suffice to suggest the general
method of listing document numbers and defining
molecular rejectors. This method is ite~ative.
Figure 4 gives a diagrammatic example of a
possible tape configuration.
The file is screened in a particularly simple
way. One takes the request descriptor vector and
begins at Mo on the tape. The request is compared
with each mOlecular rejector in turn until one of
the follOWing conditions occurs: (1) a T symbol,
along with a list of one or more identification
numbers, is reached; or (2) following ~, the
molecular rejector definition contains l-bits
where the request also contains l-bits.
In the first case, none of, the mOlecular
rejectors characterizing the rejector vector of
the listed documents has any bits in common with
the request; hence the documents in question must
possess at least those descriptors given in the
request, and the documents have passed screening.
Therefore, the identification material may either
be printed out or passed on to a more powerful
routine which calls in a more detailed request
and compares it with a more detailed index file.
In the second case, the ''bit overlap"
between request and molecular rejector means that
none of the documents to which the rejector
applies can possibly satisfy the request, since
they all lack at least one descriptor that is
asked for in the request. Therefore, all further
molecular rejectors pertinent to these documents,
as well as their identification numbers, can be
ignored, and the tape can be advanced immediately
to the next occurrence of an Mj with j < i, and
screening can be begun again at this point. (See
Figure 4 for an example of this process.)
Clearly this procedure avoids the necessity
of examining many of the molecular descriptors in
the file. The full rationale for the file
structuring procedure now becomes clear, since it
assigns molecular rejectors in such a way as to
maximize the number of documents screened out in
a single operation and also to increase the
probability that rejector-request overlaps will
occur early in the screening process.
Serial storage has the one drawback that it
is costly to add new rejector vectors (with
document identification numbers), since these
will in general belong somewhere in the middle
of the file, and recopying a part of the file tape
will be necessary to accomplish this. It is also
input-output limited, since there is a bare
minimum of "comput at ion II to be done. The best use
of this screen would be as the input-controlling
subroutine of a larger routine (as suggested
above), preferably one with much greater internal

processing. (It may be used in just this way in
connection with the HAYSTAQ chemical structure
search program.)
Random Access Integrated Program-File
If very large random-access addressable
storage is available constituting, for example,
more than half of the total required storage,
another method becomes feasible. The screening
program and the file may be integrated into one
unit, through which, in a manner of speaking,
control IIpropagates" until it reaches an output
point where print-out occurs.
In this method, the same basic organization
is used as was discussed under Conventional
Techniques, but the molecular rejector and its
marker are replaced by a subroutine which
examines the request and, depending upon what
bits occur in the request, passes control
directly on to the next subroutine or diverts it
to the subroutine which replaces the Mj (Q.~~i).
These subroutines might actually have the
rejector vectors of their molecular rejectors
stored with them, which are compared with the
request, but other more efficient alternatives
are available. (See Figure 5 for a sample flow
chart. )
This sort of file-program has the probable
drawback of requiring more storage than the
procedure outlined earlier (experimental data are
needed here), but it is faster in that it avoids
any necessity of passing over rejected material.
It proceeds directly from one comparison to the
next, with great consequent gains in speed.
A compromise between these two methods would
be to keep part of the screening routine distinct
from the file, replacing the marker Mi with an
instruction which loads an appropriate counter
.,ith the address of the molecular rejector which
is associated with Mj in the prior procedure.
This counter would control the "work area" of the
routine and would normally advance by one, unless
modified by an "M instruction".
Such methods could be most efficiently
applied in systems with very limited control
units, having only a few indexing, logical, and
conditional transfer instructions in their
repertory but very large address fields, and a
very ,large bank of addressable memory, for
example, disc storage. The random-access feature
would make it quite Simple to enter new data into
the file, since appropriate adjustment of
transfer (or indexing) instructions can allow
"logically contiguous" material to be stored in
widely separated sections of storage. (The
author would like to acknowledge the relationship
between this and the comparable techniques
employed in the various list processing
routines. 8 It may be possible to use list
processing techniques directly here without loss
of efficiency, but this seems unlikely, at least
as long as elaborate compilers and interpretative

266
6.3
routines are necessary to the use of list
processing.)
Evaluation: Comparison with
Ordinary and Inverted F~
The Files from a List Viewpoint
The best way to evaluate the worth of the
nelf file structure described here is to compare
it with the more familiar ordinary and inverted
files. Then the questions posed in the section
on Utility of Metasymbols in Screening can be
anslfered in a comparative way, and the new file
put into some sort of perspective. In order to
compare the files and the methods by which
requests are handled, Ife need to describe them in
a common language. The terminology of list
processing (used loosely) is well adapted to this
description. We will find it convenient to
describe some files in terms of descriptors,
others in terms of rejectors.
The ordinary file. In this type of file,
the primary list is the list of M document
numbers. Each document number then serves as
the point of attachment for an index sublist.
These appended sublists may be either descriptor
lists or rejector lists, although the latter may
be cheaper to process. In the processing, a
request descriptor sublist is compared with the
sublists for each document in turn to see whether
all requested descriptors are present (or that no
request descriptor occurs as a rejector), and
the document number is added to an "accepted"
list if it passes.
The inverted file. A file is normally
inverted with respect to its descriptors. Then
the primary list is made up of the N descriptors.
To each descriptor is attached a sublist of those
documents to which it applies. A request
descriptor list can be processed by use of a
"candidate list ". This list is initially the
document number sublist for the first descriptor
that appears i.n the request. Each additional
request descriptor is processed by obtaining from
the file the document number sublist attached to
it and comparing this sublist with the candidate
list. All candidate numbers which do not appear
on the descriptor sublist are removed. That is,
only those that appear on both the descriptor
sublist and the candidate list are retained
(thus taking the "and" or intersection of the two
lists). If the file were to be inverted with
respect to its rejectors, one would combine the
sublists for all those rejectors appearing as
descriptors in the request. The result would be
a list of rejected documents, which would have to
be "subtracted" from the list of all documents.
The new file. In the case of the file described in earlier sections of this paper, a more
elaborate and unconventional list structure is
required. The molecular symbols Sij __ •• p
participate in the structure, as well as document
numbers and rejectors. The primary list is a
list of single-subscript molecular symbols.

Each has attached to it two sublists:
1. A rejector sublist vThich defines the
molecular rejector in terms of its atomic
rejectors;

2. A sublist of all two-subscript molecular
symbols which have the subscript of their primary
molecular symbol as the first subscript. In the
case of a "rugged individualist" rejector vector
vThich has been~made a single index molecular
rejector, this sublist will contain the document
number or numbers which the rejector vector
would sincle out for rejection.
In turn, each two-index molecular symbol will
have attached its two sublists, and so on,
creating a highly branched tree of lists. The
logic of the screening procedure for this set of
lists is essentially that previously described
under Storage and Screening Schemata. We
observe that, in practice, the "list" method
would be impractical here, since the molecular
symbols are excess baggage.

Before leaving this topic, it is interesting
to examine the relationship of the new type file
structure to the ordinary and inverted file
types. The ordinary type, it turns out, is
nearly a special case of the new type and the
inverted file is a close relative of another
special case.
If every rejector vector of the new file is
treated as a "rugged individualist", then there
are no molecular symbols with more than one
subscript. Hence, the file becomes a Single list
of molecular symbols, each "defined" by a
rejector list and paired with one (or more)
document numbers. With the exception of the
superfluous molecular symbols, this is a refined
version of the ordinary file. However, unless
all M documents have different rejector vectors
(or lists), there will not be M entries in this
primary list, since all documents with any given
vector are combined in one entry. This is
clearly a sensible innovation in any case.
On the other hand, if each of the molecular
rejectors contains exactly one atomic rejector,
the file structure that results becomes a multilevel file inverted with respect to rejectors.
If the file were inverted with respect to
rejectors, and each of the document sublists
inverted again with respect to the remaining
rejectors, and so on repeatedly, a closely
related file structure would be obtained. In
this case, however, every document appears b
times in the file, if it has b rejectors.
the one-rejector case it appears only once.
Basically, the relational diagram of the
multiply-inverted file would contain Figure 3
as a subdiagram.

In a sense, then, the new file uses the
"inversion" idea in a more generalized form,
but applies it in a more effective way in order

267
6.3

to avoid repeated storage of the same
information.
Detailed Intercomparison of System Efficiencies
Obviously a conclusive comparison of the
relative efficiencies of the three file types is
a matter for experimental test on large real files
by using real requests. No such tests have been
performed to date. However, several possible
tests are under discussion at the National Bureau
of Standards at the present time.
One interesting test could be made on
chemical structures. Descriptors can be chosen
either from the chemical elements or from the
functional groups which can occur in the
structures. Since the occurrence or absence of
these in structures is unambiguous, the worst
problems of finding a IIfail-safe ll vocabulary are
avoided. The "libraryll indexed by these
descriptors might have as "documents" single
structures, sets of structures, and/or information
about properties, depending upon the application
envisioned.
In the absence of empirical knmdedge we are
reduced to the procedure of examining degenerate
cases, which may be quite informative. The cases
of interest are (a) a random sample of questions
which elicit nearly the whole file as answers,
and (b) a random sample of requests which reject
nearly the entire file or the whole file. The
details of the comparison depend, of course, upon
the exact way in which each of the procedures is
programmed. What one \vould like to know is
how many instruction executions on some computer
are required to process the average request. The
best we can do here is to count "manipulations",
but this provides a good indication.
In the case in which nearly the whole file
answers the request, the two more conventional
files are probably superior, but for different
reasons. The descriptor inverted file will be
the quickest for the simple reason that, given a
sensible vocabulary, a request can have fe,,, if
any descriptors in it if it is going to get
anywhere near 100%
response. Granted that
only a few descriptors can occur in a high percentage request, at most two or so lists need be
intercompared. For one, the job is done
immediately; for two-or three, by any efficient
algorithm, probably fewer than M document
identification-number comparisons are required.

more than 2N different rejector vectors pOSSible,
and the actual number in the file will often be
somewhat smaller. If M > 2N by any great
amount, it is possible that there are fewer than
M-2N "super-structure" molecular rejectors. In
any event, every reql1est satisfied by nearly
every document will have to be compared with
nearly every molecular rejector's rejector list,
and if M ~ 2N this will be more expensive than
the ordinary file. (Note that if this is really
done in list-processing form, rather than in
"vector" form, time will be saved by the fact
that there are fewer rejectors per molecular
rejector than there are per individual rejector
vector. In fact, as far as number of descriptorrejector comparisons are concerned, the values
,{ill be comparable.)
Shifting to the opposite extreme now, we see
the new type file come into its own. Here
nearly every document is rejected. We observe
that the single-subscript rejectors may be nearly
all that need be examined. In the majority of
requests having 100 percent rejection this w.ill
be true, assuming the sticky problem of finding
a good measure of merit ~(S) has been solved
satisfactorily. Unfortunately it is difficult
to estimate the probable number of these singlesubscript rejectors. It ,viII obviously be
between 0 and 2N, and considerably less than the
latter. If the degenerate "one atomic rejector
per molecular rejector" case is conSidered, there
would be N (or fewer); the number may lie near N,
in any case. At least, it will be much less than

M.
In some near-total rejection cases a request
may filter i'airly far down the "rejector tree"
before being blocked, but because of the use of
~, this will be in the minority of cases.
Hence
the average number of molecular rejector M
request comparisons will be small, generally
much smaller than M, particularly if M > 2N.
Here, as in the total acceptance case, the
ordinary file still requires M request listdocument list comparisons and will lose out so
far as efficiency is concerned.
The inverted file is a bit more complicated.
Suffice it to say that if a given document has k
out of the r descriptors requested, it is
rejected k tImes, in effect. Until a request
descriptor is reached that does not apply to that
document, the document appears in the "candidate
list
Even after it is stricken from that list,
other lists must be manipulated in which it is
carried as excess baggage, and for each of these
it must be classified as "not in the candidate
list II and discarded. Since the molecular
rejector file rejects each document only once
(in a sense less than once since the document in
question is likely to be rejected along with
others), it is more efficient in this respect.
11 •

In the case of the ordinary file, M
descriptor or rejector lists must be compared
with the request. If vectors are used, this
amounts to M 'and' -and-test sequences (multiple
precision if necessary).
In the case of the new file, we observe that
every set of identification numbers (with the
same rejector vectors) has a "last" S.ik ••• p
associated with it alone. But there is then a
superstructure of molecular rejectors with fewer
subscripts erected over this base. There are no

To summarize our results, then, the new
file is likely to be superior to the ordinary
file if the screen is powerful enough to average

268

6.3
a very hi8h percentage of rejection. If it is
not this powerful, then it is hardly worth using
in any case. If M > 2N , it is better in all
cases. We also conclude that the molecular
rejector file is very probably better than the
inverted file in the high-rejection cases, but
there is a much stronger need for data in this
case.

Documentation, Vol. X, No.1, pp 20-26,
Jan. 1959.
For topological network features: Herbert
R. Koller, Ethel Marden, and Harold Pfeffer,
"The HAYSTAQ System: Past, Present and
Future, II Preprints of Papers for the
International Conference on Scientific
Nov. 16-21,
195 •

A Final Efficiency Problem
Even granting that a case can be made for
the contention that the new type of file is
superior to the other two if heavy screening is
pOSSible, a more difficult question remains. l\n
estimate is needed of the savings which will
result from use of the pre-processed file, as
contrasted with the expense of writing, debugging,
and then using the file-organizing routine. This
Cluestion appears to be unanswerable without data,
but it does suggest the desirability for a
potential user to consider the "reCluest-traffic ll
of his file. A file with light traffic will
naturally take longer to pay for itself than one
with heavy use. Because of the possibility of
periodic reprocessing of the file because of new
docu~ents (as discussed before), the use must be
heavy enough to -pay for this, at least.

2.

"Depth" refers to the degree to which the
complete content of a document is
represented by the descriptor set (or other
representation) aSSigned to it. Deep
indexing, then, leaves out very little of
the content of any document in the file.

D. D. iilldrews and Simon M. Newman,
"Activities and Objectives of the Office of
Research and Development in the U.S.Patent
Office, " Journal of the Patent Office
SOCiety, Vol. 40, No.2, Feb. 1958, pp.7985.

4.

It is hoped that actual machine testing of
some of the theories described here \OTill be
possible in the near future.

F.E. Firth, An Experiment in Literature
Searching with the IBM 305 RAMAC, San
Jose, California: IBM, November 17, 1958;
J. J. Nolan, Principles of Information
Storage and Retrieval Using a Large Scale
Random Access Memory, San Jose, California:
IBM, November 17, 1958.

Conclusion
Two basic ideas helpful in constructing
information retrieval systems have been discussed.
~irst, the advantages of the screenin8 approach
to the fi".:..e orcanization in the develo)ment of
systems were argued. Secondly, the desirability
of pre-processing files to take advantage of
block-rejection was suggested. These two ideas
were then applied to the ~roblem of setting up
a r..early fail-safe screen using descriptors or,
more accurately, rejectors. An algorithm was
sketched for generating descriptions of documents
in terms of "mOlecular rejectors ll • Another
algorithm was outlined for storing on tape a file
of the type described. Also discussed was an
alternative "stora8ell system, using the idea of
"dynamic storage" in ,vhich information is conveyed
by branch points in a program.
The new type of file was then compared with
the better.-&~own ordinary and inverted files.
The conclusion of the theoretical reasoning was
that the system sho\vs enough promise of increased
efficiency to warrant detailed testing of real
files.

Jacob Leibowitz, Julius Frome, and Don D.
Andrews, Variable 3cope Patent Searching by
an Inverted File Technique, Patent Office
Research and Development Reports ••• No. 14,
U. S. Department of Commerce, Washington,
~.C., Nov. 17, 1958.
Jacob Leibowitz, Julius Frome, and F. D.
Hamilton, "Chemical Language Coding for
Machine' Searching,1I Abstracts of Papers,
l35th Meeting, i:l,merican Chemical Society,
Boston, 5-10 April 1959, p. 3G.

6.

Victor H. Yngve, liThe Feasibility of
Machine Searching of English Texts," Preprints of Papers for the International
Conference on Scientific Information
Washington, D.C., Nov. 16-21, 1958. Area 5,
pp. 161-169)

t

, In Defense of English, Preprint of
Paper presented at An International
Conference for standards on a Common
Language for Machine Searching and Translation, Sept. 6-12, 1960, Cleveland,

8 p.
References
1.

For the logical (propositional calculus)
features: Harold Pfeffer, Herbert R. KOller
and Ethel C. Marden, "A First Approach to
Patent Searching Procedures on Standards
Electronic Automatic Computer," American

If tape is used, it is clearly most
efficient to replace Mi with the number of
words between Mi and the appropriate Mj, so
that the tape can be advanced immediately
without having to examine each word until
Mj is found. For large files (many tapes),

2.69

6.3

8.

tape number and tape position could be
specified ,in place of Mi.

of the predicted average may be used as the
"quali ty" of the file.

J. C. Shaw, A. Newell, H. A. Simon and T. 0.
Ellis, "A Command Structure for Complex
Information Processing", Proceedings of the
Western Joint Comuuter Conference - Contrasts
in Computers, presented at Los Angeles,
California, May 6-8, 1958, pp. 119-128.

However, another competing criterion is
involved, that of the complexity of the algorithm
needed to apply the prediction of averages to the
choice of a vocabulary. Such an algorithm -when
programmed, must be able to find a "best"
vocabulary in some reasonable running time.
Unfortunately, in order to get manageable
routines, we have to cut corners and make shaky
approximations in various places. Ultimately,
we must balance time saved in screening against
time consumed in processins, and stop at some
(at this time) ill-defined "break' even" point.

John McCarthy, "Recursive F\mctions of
Symbolic Expressions and Their Computation
by Machine, Part I," Communications of the
ACM, Vol. 3, No.4, April 1960, pp. 184-195.
Victor H. Yngve, "The COHrr System for
Mechanical Translation", Proceedings of the
International Conference on Information
Processing, UNESCO, Paris, France, June 1520, 1959, pp. 183-187.
Appendix
Details of Selecting an Optimal
File Structure
A.

Introduction

Here we address ourselves to the details of
the choice of an optimal molecular rejector
language for the file. The desirability of this
molecular description will be assumed.
The concept of an optimal language needs
considerable amplification. He wish to minimize
the average number of molecular rejector-request
comparisons needed to "screen" a given request,
assuming that the amount of machine time used in
screening one request is proportional to the
number of comparisons performed. This assumption
will be strictly true for fixed field proceSSing
in which N is less than or equal to the number of
bits in a word, and on the average true when fixed
field "vectors" must be processed with multiple
precision. It is not really valid if variable
field (list-type) processing is used. In the
latter case a different analysis is needed.
The notion of an average is necessary Since,
as the section on "Evaluation" pOints up, the
number of comparisons for a given request is
strongly dependent upon the nature of the request.
But then an unpleasant question asserts itself:
"Over what do you average?" The answer one would
like to give is "We want the average over all the
requests which will ever be presented to the file."
If this average is minimized, we have the best
file, by definition. Unfortunately this is an
unhelpful answer to the question since it cannot
be found until all requests have been processed.
Then it is a bit late to be of any use.
Hence we must resort to imperfect predictors
of what this average would be for each molecular
vocabulary in order to choose the best one. Each
prediction method can then be used in a
vocabulary generation algorithm. The reciprocal

B.

Development of the Prediction Method

One can break down the prediction methods
into two types: external sample methods and file
sample methods. In the external sample methods,
a sample of requests is gathered from potential
users, and these are used as predictors of the
kind of requests likely to be directed to the
file. On the other hand, to the extent to which
the content of the file is representative of the
distribution of interest of people using the file
statistics about occurrence of molectuar
'
descriptors in the file can be used to predict
requests without external data f'rom potential
users. Our tentative hypotheSis is that this
latter method is probably quite adequate. It is
obviously cheaper, since it reduces the amount of
error-free machinable data that must be prepared.
One quite serious problem in the choice of
the best molecular vocabulary arises from the
question of how many alternative complete
vocabularies must be eenerated. If viC have a
method of aSSigning each voca1ulary a value,
clearly we could generate all possible
vocabularies, assign each a value, and then
choose the best. However, the nu~ber of
oper ations involved grOloTs as some "fierce II
exponential of the number of documents in the
file and the vocabulary size.
The most desirable method vlould fol101oT a
direct path from molecular rejector to mOlectuar
rejector, in every case choosing exactly that one
which will lead to the best final file. \'Jhat
this requires is that at every point we be able
to assign to each candidate-rejector S a
"measure of merit" Il(S) vlhich predicts the
ultimate quality of the file which will result
if it is chosen. It is by no me~~s clear that
such a 11 can be found "Thieh avoids effectively
"trying out" nearly every possible resultant file
(another exponential factor).
,~t this point,
too, we Ifill find it necessary to make imoerfect
predictions in order to find a manaGeable-process.
Our procedure might be described as
peSSimistic, or cautious. Ithen choosin~ a molecular rejector at any pOint, we consider the
~ possibility for number of com"!;>arisons in a

270
6.3

file determined by the choice, and then choose
the one that gives the IIstrongest guarantee
against the worst". In other words, at each
point of choosing a molecular rejector, we are
in a position to assert, "The maximum possible
number of comparisons to which I might obligate
myself if I choose this rejector is k.
Then we
choose the rejector with minimum k. Every
rejector will, in fact, do better than this
estimate when the whole vocabulary is chosen (this
will be clarified later), and it is possible that
one of the apparent underdogs would come out best
in the end, but to check thiS, one would have to"try it and see", which means another case of
exponential growth in number.
We reason as follows:

ascribed to them, further justifying the fact
that 4 and 5 are pessimistic estimates.

9.

It is clear that we minimize the number of
comparisons that the (sub)file F contributes to
the total by maximizing fiPeS.) where fi is the
number of documents in Fl' TEerefore, we choose
as the measure of merit u(Si) the number ~(Si) =
fiP(Si) •
10. Suppose ~(Si)
~(Sj)' We would like to
find a further, criterion for chOOSing Si or Sj
as "most likely to succeed". Suf'fice it to say
that the one of largest P(S) is probably best.
If P(Si) = p(Sj), a random choice is the best we
can do.

C.

Estimates of the Rejection Probability pCS)

1.
For each molecular rejector S there will be
a probability peS) that a IIrandom ll reCluest will
have one or more bits in common with S. This we
will call the "rejection probability" of S, the
probability that S will be able to reject all
the documents to which it applies. How one
computes peS) will be taken up later.

Clearly the tally t (SbF) can be obtained
simply by counting and presents no important
problems. The rejection probability P(Si) is
more difficult, however. A number of methods of
evaluating or estimating P(Si) will be presented
here, in order of decreasing sop~stication
(and complexity, in general).

2.
For a (sub)file F with R different rejector
vectors (documents with like vectors are already
grouped), the number t(Si,F) of vectors to which
a given Si applies is calculated. The subfile to
which it applies is Fi'

1.
Outside Sample; Complete Request Method.
Here we assume that information is available
from an outside sample. Every possible set of
descriptors which might make up a request
Q (all 2N_l of them) is tested against the
sample, and a percentage p(Q) is computed.
p(Q) is the percent of sample requests which are
exactly the request Q, in other words, the
probability that a random reauest will be Q. A
table of the P(Qi) for all 2 -1 of the Qi is
stored.

3. We then note that if a reCluest Q is directed
to the (sub)file F, it has a probability P(Si) of
being rejected by Si' and hence a probability
l-P(Si) of passing on to the file Fi'
4.
Being pessimistic, 1tre assume that after Si
is removed, each of the Fi rejector vectors has
to be given its own molecular rejector. Hence
if Q "passes through" S1, t(Si,F) further
comparisons must be made. But only
(1-P(S1»xl02 percent of the reCluests are passed
through. This leads to an average of
(l-P (Si»·t(Si,F) comparisons.

5. In addition, Q must be directed at the
R-t(Si,F) other vectors. Assume all of these
must be treated as rugged individualists, at one
comparison per vector.
6.

Totalling the number of comparisons
(including that with Si itself), we ~et
K=l+(l-P (Si»t (Si,F);j-R-t(Si,F)=R+l-t (Si,F) P(Si)'

7.
We observe that if t(Si,F)' P(Si)< 1, then
K < R. Since R comparisons would be involved if
each vector were processed separately, in this
case use of the molecular rejector Si will be an
improvement over "ordinary II processing. It
t(Si,F). P(Si) < 1, all the vectors should
probably be stored individually.

For each Si, all those requests overlapping
with Si,identified as R~j' are generated. Then
we have P(Si) = ~(Rij)'
(This is the usual
probability that at least one of a list of
independent events will occur. In this case the
"event" is the occurrence of exactly the
particular Rij as a request and hence is
independent of that for any other R~k' klj, even
if Rij A Rik has many I-bits in it.)
2.
File Sample; Complete Request Method. The
procedure is exactly the same as above, except
that the file itself is used as the sample.
This is not a trivial process, since the file is
encoded in terms of its rejectors rather than
descriptors. It is best done for each Q~ by
looking for exactly Qi (complement of Qi) in the
file, counting the number of these to get N(Qi),
and then dividing by M to get P(Qi) = N(Qi)'
M

It the file is uncomfortably large, one might

select a manageable random subfile for this
purpose.

8.

Furthermore, we note that if this reasoning
is applied to the subfiles of Fi and F-Fi,
respectively, it is clear that each of these will
probably do better than the t(Si~F)and R-t(S~)

3.
Atomic Request Method: Either Sample.
for each atomic descriptor Di{ we simply
evaluate the probability P(Di) that this

Here,

271

6.3
descriptor will be found in a given request. ~:re
then estimate the number P(Qj) of the above
analysis by listing the set {Dk} that occur in
Qj (symbolically, Qj = VDk). We say
k

p(Q) ~ P(Dl)·P(D2) ••• P(Dk) = n 1'(Dk) (iterated
proauct). Then the formula for P(8i) is
computed as in C.l. This computation is based
upon the well known formula for the probability
that a number of independent events will occur
together. We know that, in fact, the occurrence
of descriptors in a request will not be
independent; they will tend to "cluster". C.l
-and C.2 take this into account, but it is thrown
away here. Clearly a considerable computation is
aVOided however.

4. The No-Sample Complete-Request Method. Here
we assume that all requests are equally probable.
Thus we need only to evaluate the percentage of
all possible requests which overlap 8 i to get
P(Si) to within some multiplicative constant
(which can be ignored). There are i:N-l
requests. If 8i has bi bits, there are N-bi bits
that do not "intersect" 8i and 2N- bL l requests
made up only of those bits. Hence there are
(2N-l) - (2 N- b i-l) = 2N_2N- b i requests which do
overlap with 8i. This P(Si) equals
-

The important quantity is
1-

-1- , which clearly increases

2bi
directly with bio (This can be found neatly
on a computer with one-bit right shifts as one
counts the number of bits in 8i and then
subtracts. )

Clearly, all of these methods are
programmable and not too forbidding when
considered separately. The problem arises from
the fact that the ~(S) subroutine will need to
be used an astronomical number of times in
generating a file structure and, in fact, is
likely (it appears to the author) to account for
a high percentage of the running time of the
file-organization routine.

272

6.3

C

start)

~
Load F-store ',rith
entire File, Clear Rstore, enter entire
vocabulary into V-store

of

Yes

No
F-store empty?

f

;Reload V-store i'Tith
full vocabulary, remove
all terms appearing in
R-store molecular
rejectors

Generate all molecular
rejectors containing
terms in the V-store

t
Remove all descriptors
occurring in S from
the V-store

t

Read in from Prime Store:
(a ) Prime d urnp
F-store
(b) R-dUIllP
R-store

Dlliap each doc. IU,
new molecular rejector,
plus contents of R-store,
on output

,

Make each rejector
vector in F-store into
new molecular rejector

Compute ~(S) for each
of these molecular
rejectors

Select the S with
maximwa ~(S) (or one
such S)

Max J..I.:

Remove all doc. ID's
with empty rejector
vectors from F-storej
dump with copy of Rstore on output

"Subtract" S from
all rejector vectors
in F-store

1
Remove all entries
to which S does not
apply from F-store

Select from F-store all
entries whose rejector
vectors do not contain S;
dump these as prime dump
on prime store

Insert S into
R-store

tI
Copy R-store on Prime
store as R-dump

I

Note: The Prime store is a storage area for primed subfiles. Each primed
sub file is stored along with all those molecular rejectors applying to every
document in the subfile. The left branch corresponds to Step 4 in text;
the right to Step 5.

Figure 2.

Flow Chart of File Organization Routine

•

273

6.3

Figure 3.
a.

file Generation Tree

Tape Configuration

c

A

b.

Generation Tree for File

E

0

c.

F

Documents in File

Document

Molecular Description of
Rejector Vector

A

81 V 811

B

81 V 811 V 8111

C

81 V 812 V 8J.2l

D

81 V 8
V8
122
12

E

81 V 8
13

F

8 2 V 8 21

d.
Request
hence by c.

has bits in common with 8111 8122 , and 82' but with no other molecular rejectors,
should reject A, B, D, F. C and E should be accepted.
Q

Processing starts at

0 ' proceeds word by word to ® where overlap is

immediately to (]) , proceeds word by word to
occurs.

Tape is advanced to

occurs.

At

®

@ ,

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 18:31:55-08:00
Create Date                     : 2008:11:16 18:31:55-08:00
Metadata Date                   : 2008:11:16 18:31:55-08:00
Format                          : application/pdf
Document ID                     : uuid:b080bf44-3d6e-442c-ae9f-1a9b9f621122
Instance ID                     : uuid:a300b07a-fc9e-4c9f-9725-fd28d0adc665
Page Layout                     : SinglePage
Page Mode                       : UseOutlines
Page Count                      : 668