README for TMVATrainer/TMVATester
---------------------------------
This README file explains how a TMVA training can be performed on a TTree containing signal and background
events with a simple set of two macros. The TMVATester macro shows the performance as well as writes
a file containing a TNtuple with the network output, which can be added as a friend to the original TTree.


################
- Relevant files:
  - makeTMVADemoData.C (produces demodata.root)
  - TMVATrainer.C (ROOTVER < 6.08),  TMVATrainer_608.C (ROOTVER >= 6.08)
  - TMVATester.C 
################


When using some root-file 'input.root' with a ttree 'ntp' containing signal and background, 
you can train a BDT (boosted decision tree) TMVA with

root [0] .x TMVATrainer.C+("input.root", "ntp",
                           <signal cut>,
                           <variables>,
                           [method],
                           [precut]) 

where <signal cut> is the cut separating good from bad candidates, and <variables> is a blank 
separated list of the variables which should be used for the training. Also an opional precut can be applied, and
the TMVA-methods available are "BDT" (boosted decision tree), "MLP" (multi layer perceptron) and "Likelihood"
(conventional likelihood method).

To test and write out a friend tree 'tf' to a file 'input_tmvaout.root' with the network output, execute:

root [0] .x TMVATester.C+("input.root", "ntp",
                           <signal cut>,
                           <weights file>,
                           [precut]) 

Here, <weights file> is the output of the training process, in our example in the new created subfolder weights/ 
named input_ntp_METHODE.weights.xml, where METHOD = BDT, MLP or Likelihood. If a precut was used for training, 
the same should be applied here. The tester shows the network output for signal and background and a ROC curve.

It can be used interactively by this commands:

root [0] TFile f("input.root")
root [1] ntp->AddFriend("tf","input_tmvaout.root")
root [2] ntp->Draw(<your variable>,"tmvaout>0");


------------------

EXAMPLE:

With help of the macro 'makeTMVADemoData.C' you can create some dataset containing artifical signal and background together
with 5 gaussian distributed variables (v1, v2, ... , v5) as input for the TMVATrainer. The branches are

m        : the (artificial mass; signal is gaussian around 1.864, background is flat)
ev       : event number
signal   : flag for signal (=1.0) and background (=0.0)
v1,..,v5 : the variables to distinguish the signal from background



Run the full example with: 

# Create demo data
root -l -b -q makeTMVADemoData.C

# Training for older ROOT versions #
root -l -b -q 'TMVATrainer.C("demodata.root","ntp","signal>0","v1 v2 v3 v4 v5")'

# Training for ROOT version >= 6.08 #
root -l -b -q 'TMVATrainer_608.C("demodata.root","ntp","signal>0","v1 v2 v3 v4 v5")'

# Test the MVAs
root -l 'TMVATester.C("demodata.root","ntp","signal>0","weights/demodata_ntp_BDT.weights.xml")'


This prints the TMVA output and ROC curve. To attach the TMVA output to the n-Tuple, use (see above)

# Using the TMVA output for your input NTuple

root [0] TFile f("demodata.root")
root [1] ntp->AddFriend("tf","demodata_tmvaout.root")
root [2] ntp->Draw("m","tmvaout>0");