Home » Groups » STAR Protected

Using TMVA to distinguish photons

Updated on Tue, 2012-01-10 13:32. Originally created by sgliske on 2012-01-10 13:32.

I tried using ROOT's TMVA to test several multivariate tools to distinguish photons from non-photons, based on the preshower energies.

The variables used for the MVA were the preshower1 energy divided by the tower energy, and the preshower2 energy divided by the tower energy. Note: the TMVA was setup as

TMVA::Factory *factory = new TMVA::Factory( "TMVAnalysis", outputFile, Form("V:!Silent:Color", gROOT->IsBatch()?"!":"") );
factory->AddVariable( "ePre1/eTow", 'F' );
factory->AddVariable( "ePre2/eTow", 'F' );
factory->PrepareTrainingAndTestTree( TCut(""), "SplitMode=Random:NormMode=EqualNumEvents:!V" );
factory->BookMethod( TMVA::Types::kBDT, "BDT", "!H:!V:NTrees=100:BoostType=AdaBoost:SeparationType=GiniIndex:nCuts=20:PruneMethod=CostComplexity:PruneStrength=1.5" );
factory->BookMethod( TMVA::Types::kMLP, "MLP", "!H:V:!Normalise:NeuronType=tanh:NCycles=200:HiddenLayers=N+1,N:TestRate=5" );

While other methods were tried in the TMVA, the MLP and BDT seemed to have the best results.

The various options considered in the following table are

  1. MLP: a multi-layer perceptron (neural network)
  2. BDT: a boosted decision tree
  3. Cut A: cut on "ePre1/eTow*1000"
  4. Cut B: cut on "(ePre1-ePre2)/(ePre1+ePre2)"
  5. Cut C: cut on "(ePre2-ePre1)/eTow < thres"
  6. Cut D: combine cuts A, B, and C

Cuts were adjusted to either accept only 30% of the background or 10% of the background.  In both cases, the percentage photons accepted are reported in the following table

Method Sig. Eff @ 10% Bkg Sig. Eff. @ 30% Bkg
MLP 73.8% 88.3%
BDT 73.5% 86.8%
Cut A 73.5% 87.3%
Cut B 62.7% 76.7%
Cut C 40.5% 67.5%
Cut D 47.1% 71.5%

 

The data used was the 8 set 1 files, i.e. eta = 1.5 and phi = 0, and at 8 different pT values, generated via the "Micky Mouse" generator.  See this email for more details.

Conclusions

It seems a simple cut, i.e. cut A, is comparable to the fancy MVA results.  Thus, the simpler method is preferable.  Note, the preshower 1 energy is the most useful in distinguishing between photons and other particles.

 

Groups: