ConvergenceTest.cxx

Go to the documentation of this file.

// @(#)root/tmva $Id$
// Author: Andreas Hoecker, Peter Speckmayer, Joerg Stelzer, Helge Voss

/**********************************************************************************
 * Project: TMVA - a Root-integrated toolkit for multivariate data analysis       *
 * Package: TMVA                                                                  *
 * Class  : ConvergenceTest                                                             *
 * Web    : http://tmva.sourceforge.net                                           *
 *                                                                                *
 * Description:                                                                   *
 *      Implementation (see header for description)                               *
 *                                                                                *
 * Authors (alphabetical):                                                        *
 *      Andreas Hoecker <Andreas.Hocker@cern.ch> - CERN, Switzerland              *
 *      Peter Speckmayer <Peter.Speckmayer@cern.ch>  - CERN, Switzerland          *
 *      Joerg Stelzer   <Joerg.Stelzer@cern.ch>  - CERN, Switzerland              *
 *      Helge Voss      <Helge.Voss@cern.ch>     - MPI-K Heidelberg, Germany      *
 *                                                                                *
 * Copyright (c) 2006:                                                            *
 *      CERN, Switzerland                                                         *
 *      MPI-K Heidelberg, Germany                                                 *
 *                                                                                *
 * Redistribution and use in source and binary forms, with or without             *
 * modification, are permitted according to the terms listed in LICENSE           *
 * (http://tmva.sourceforge.net/LICENSE)                                          *
 **********************************************************************************/

/*! \class TMVA::ConvergenceTest
\ingroup TMVA

Check for convergence.

*/
#include "TMVA/ConvergenceTest.h"
#include "TMath.h"

////////////////////////////////////////////////////////////////////////////////
/// constructor

TMVA::ConvergenceTest::ConvergenceTest()
   : fCurrentValue( 0 ),
     fImprovement( 0 ),
     fSteps( 0 ),
     fCounter( -1 ),
     fConvValue( FLT_MAX ),
     fMaxCounter( 0 ),
     fBestResult( FLT_MAX ),
     fLastResult( FLT_MAX )
{
}

////////////////////////////////////////////////////////////////////////////////
/// destructor

TMVA::ConvergenceTest::~ConvergenceTest()
{
}

////////////////////////////////////////////////////////////////////////////////
/// gives back true if the last "steps" steps have lead to an improvement of the
/// "fitness" of the "individuals" of at least "improvement"
///
/// this gives a simple measure of if the estimator of the MLP is
/// converging and no major improvement is to be expected.

Bool_t TMVA::ConvergenceTest::HasConverged( Bool_t withinConvergenceBand )
{
   if( fSteps < 0 || fImprovement < 0 ) return kFALSE;

   if (fCounter < 0) {
      fConvValue = fCurrentValue;
   }
   Float_t improvement = 0;
   if( withinConvergenceBand )
      improvement = TMath::Abs(fCurrentValue - fConvValue);
   else
      improvement = fConvValue - fCurrentValue;
   if ( improvement <= fImprovement || fSteps<0) {
      fCounter ++;
   } else {
      fCounter = 0;
      fConvValue = fCurrentValue;
   }
   if (fCounter < fSteps) return kFALSE;
   return kTRUE;
}

////////////////////////////////////////////////////////////////////////////////
/// returns a float from 0 (just started) to 1 (finished)

Float_t TMVA::ConvergenceTest::Progress()
{
   if( fCounter > fMaxCounter )
      fMaxCounter = fCounter;
   return Float_t(fMaxCounter)/Float_t(fSteps);
}


////////////////////////////////////////////////////////////////////////////////
/// this function provides the ability to change the learning rate according to
/// the success of the last generations.
///
/// Parameters:
///
///  - int ofSteps :  = if OF the number of STEPS given in this variable (ofSteps) the
///                       rate of improvement has to be calculated
///
/// using this function one can increase the stepSize of the mutation when we have
/// good success (to pass fast through the easy phase-space) and reduce the learning rate
/// if we are in a difficult "territory" of the phase-space.

Float_t TMVA::ConvergenceTest::SpeedControl( UInt_t ofSteps )
{
   // < is valid for "less" comparison (for minimization)
   if ( fBestResult > fLastResult || fSuccessList.size() <=0 ) {
      fLastResult = fBestResult;
      fSuccessList.push_front( 1 ); // it got better
   } else {
      fSuccessList.push_front( 0 ); // it stayed the same
   }
   while( ofSteps <= fSuccessList.size() ) // remove the older entries in the success-list
      fSuccessList.erase( fSuccessList.begin() );
   Int_t n = 0;
   Int_t sum = 0;
   std::deque<Short_t>::iterator vec = fSuccessList.begin();
   for (; vec != fSuccessList.end() ; ++vec) {
      sum += *vec;
      n++;
   }

   return  n ? sum/Float_t(n) : 0;
}