Tmva Instructions
tmva_instructions
User Manual:
Open the PDF directly: View PDF 
.
Page Count: 10
TMVA package
●The TMVA package performs a multivariate analysis on a number of input
variables and returns a single MVA variable
●TMVA is included in ROOT versions 5.30 and above, the development version
can be found at http://sourceforge.net/projects/tmva
●All source files and examples can be found under $ROOTSYS/tmva
●The typical TMVA analysis is divided into:
–TMVAClassification: training and testing of a MVA method with included data set
($ROOTSYS/tmva/test/TMVAClassification.cxx)
–TMVAClassificationApplication: applying MVA cut to a data set
($ROOTSYS/tmva/test/TMVAClassificationApplication.cxx)
●User guide:
http://tmva.sourceforge.net/docu/TMVAUsersGuide.pdf
Preparing for MVA
●Example MVA:
http://sabotin.ung.si/~gkukec/gkukec/tmva/tmva_example.tar.gz
●$TMVAEX will from now on designate the folder where the above example is
extracted to
●The above example trains and tests a neural network method on a data set
($TMVAEX/example.root) → check the contents of the file by opening a
TBrowser in ROOT
●The data set should be saved in a ROOT file, where a tree holds input
variables:
TreeS1
Variable1
Variable2
...
TreeS2
Variable1
Variable2
...
...
Program structure ($TMVAEX/tmva_simple.cpp)
●At the beginning, we open two root files – one for reading input variables, the
other to save the MVA output:
Tfile *ifile = Tfile::Open(“<input.root>”, “READ”);
Tfile *ofile = Tfile::Open(“<output.root>”, “RECREATE”);
●Then we create a Factory class object that will take care of training and
testing (optionally, we can designate the MVA weights folder):
TMVA::Factory *factory = new TMVA::Factory(“<JobName>”, ofile,
“<options>”);
(TMVA::gConfig().GetIONames()).fWeightFileDir = “./weights”;
●Each variable that will be used in the MVA is then added (the last argument is
the variable type F for float, I for integer):
factory->AddVariable(“<variable name”, ‘<variable type>’);
Program structure ($TMVAEX/tmva_simple.cpp)
●Each tree that will be used needs to be designated as signal or background –
in the MVA, all signal trees will be combined together:
TTree *signalTree = (TTree*)ifile->Get(“<TreeName>”);
TTree *backgroundTree = (TTree*)ifile->Get(“<TreeName>”);
factory->AddSignalTree(signalTree, <weight>);
factory->AddBackgroundTree(backgroundTree, <weight>);
●Both trees are then prepared for training and testing:
factory->PrepareTrainingAndTestTree(“<sigSelectionCuts>”,
“<backSelectionCuts>”, “<options>”);
●If preselection cuts are not needed, first two arguments are left empty
●Some options:
–nTrain_Signal, nTrain_Background, nTest_Signal, nTest_Background:
Number of signal and background events used for training and testing – if set to 0,
half of the events will be used for training and half for testing
–SplitMode: Selection of events for training/testing (Random, Alternate, Block)
Program structure ($TMVAEX/tmva_simple.cpp)
●Then we select the MVA method that will be used for the analysis – more
information on each can be found:
http://tmva.sourceforge.net/optionRef.html
factory->BookMethod(<TMVA type>, “<TMVA name>”, “<options>”);
●Some method types (with names):
–Likelihood: TMVA::Types::kLikelihood (Likelihood, LikelihoodD,
LikelihoodPCA,...)
–Function discrimination analysis: TMVA::Types::kFDA (FDA_GA, FDA_SA,
FDA_MC,...)
–Artificial neural networks: TMVA::Types::kMLP (MLP, MLPBNN,...),
TMVA::Types::kCFMlpANN (CFMlpANN),...
–Boosted decision trees: TMVA::Types::kBDT (BDT, BDTG, BDTB, BDTD,...)
●See $TMVAEX/def_methods.cpp for default options for a collection of
different methods
●For classification, more than one method can be used (with additional
BookMethod definitions)
Program structure ($TMVAEX/tmva_simple.cpp)
●Train, test and evaluate the methods, then close both files:
factory->TrainAllMethods();
factory->TestAllMethods();
factory->EvaluateAllMethods();
ifile->Close();
delete factory;
ofile->Close();
●The output file can now be opened using a TMVA GUI with:
–root -l ‘TMVAGui.C(“<output.root>”)’
–./tmvagui <output.root>
●Now we continue with the classification application part of the program
●Wish to apply a MVA variable cut onto the data set – we create a Reader
class object and variables we will read values to:
TMVA::Reader *reader = new TMVA::Reader(“<option>”);
float obsvars[nrvars];
Program structure ($TMVAEX/tmva_simple.cpp)
●We add all variables we wish to apply the MVA cut to and prepare the method
we used in the classification:
reader->AddVariable(“<variable name>”, &obsvars[0]);
...
reader->BookMVA(“<method name>”, “<path to weights XML
file>”);
●We now open the file that holds the trees and variables for our case and set
the MVA cut (can be determined from optimal cut from GUI):
TFile *ifile = TFile::Open(“<input.root>”, “READ”);
double cut = <cut value>;
●Making a loop through all the trees, we set the address each variable (the
names here must be the same as variables we gave to the reader):
for(...) {
TTree *treeCur = (TTree*)ifile->Get(“<current tree name>”);
treeCur->SetBranchAddress(“<variable name>”, &obsvars[0]);
...
}
Program structure ($TMVAEX/tmva_simple.cpp)
●Inside the existing loop, we now do another loop through all events in the tree
and check, if event is considered as signal or background:
for(...) {
TTree *treeCur = (TTree*)ifile->Get(“<current tree name>”);
treeCur->SetBranchAddress(“<variable name>”, &obsvars[0]);
...
for(int ievt = 0; ievt < treeCur->GetEntries(); ievt++) {
treeCur->GetEntry(ievt);
if(reader->EvaluateMVA(“<method name>”) >= cut)
// event is signal
else
// event is background
}
}
●Make sure that the used MVA method has the classifier that above the cut is
signal and below the cut is background (some could have it reversed)
Plots from example
