This is a continuation of the previous blog entry where only one pt bin was considered. In this analysis, I looped over all the pt bins from the Monte Carlo and converted the variables into the TTree format. I trained the regression with 10000 events and tested with 54438 events.

• Where to find the code for this -> /star/u/rsalinas/
• To produce this output for this analysis, run the following:
  • ./star/u/rsalinas/ALL.sh
    • Produces the files tree_fileD.root & tree_fileP.root 
  • scp -p username@rftpexp.rhic.bnl.gov:path/on/RCF path/to/my/machine -> to copy to your machine
  • root -l TMVARegpract.C\(\"MLP,MLP0,DNN_CPU\"\) -> do this on your machine with the attached TMVA regression .C file
    • MLP
      • Activation Function: sigmoid
      • Hidden layers: 10,10
      • Learning rate: .2
    • MLP0
      • Activation Function: tanh
      • Hidden layers: 15
      • Learning rate: .001
    • The regression outputs produced and in TMVARegprac_embedding.root when you run the TMVA regression

Variables

Input Variables

• dJetPt

• dJetEta

• pJetPt (particle) 

  

 

Regression Results

DNN_CPU

MLP0

        
 

MLP

            

Convergence

Comparisons

When comparing the regression outputs, you can run the following:

• root -l plot_regression.C  (on your machine) 

• From the following plots, you can see the training regression of the different methods:

A lot of the methods undershot the true pt level but that could be due to insufficient events in the training, the learning rate, hidden layers, etc... The goal now is to figure out this "general" unfolding, compare it to the RooUnfold method, and then proceed to do this for hadrons-in-jets. Supplementary, how different input variables affect the regression. Exciting!