PDEFoamMultiTarget.cxx

Go to the documentation of this file.

// @(#)root/tmva $Id$
// Author: Tancredi Carli, Dominik Dannheim, Alexander Voigt

/**********************************************************************************
 * Project: TMVA - a Root-integrated toolkit for multivariate data analysis       *
 * Package: TMVA                                                                  *
 * Classes: PDEFoamMultiTarget                                                    *
 * Web    : http://tmva.sourceforge.net                                           *
 *                                                                                *
 * Description:                                                                   *
 *      Implementation.                                                           *
 *                                                                                *
 * Authors (alphabetical):                                                        *
 *      Tancredi Carli   - CERN, Switzerland                                      *
 *      Dominik Dannheim - CERN, Switzerland                                      *
 *      S. Jadach        - Institute of Nuclear Physics, Cracow, Poland           *
 *      Alexander Voigt  - TU Dresden, Germany                                    *
 *      Peter Speckmayer - CERN, Switzerland                                      *
 *                                                                                *
 * Copyright (c) 2008, 2010:                                                      *
 *      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::PDEFoamMultiTarget
\ingroup TMVA
This PDEFoam variant is used to estimate multiple targets by
creating an event density foam (PDEFoamEvent), which has dimension:

    dimension = number of variables + number targets

This PDEFoam variant stores in every cell the sum of event weights
and the sum of the squared event weights.  During evaluation for a
given event, which has only variables and no targets (number of
event variables is smaller than the foam dimension), the targets
are estimated by finding all cells, which correspond to this event
and calculate the Mean (or Mpv, depending on the ETargetSelection)
cell center weighted by the event density in the cell.

This PDEFoam variant should be booked together with the
PDEFoamEventDensity density estimator, which returns the event
weight density at a given phase space point during the foam
build-up.
*/

#include "TMVA/PDEFoamMultiTarget.h"

#include "TMVA/MsgLogger.h"
#include "TMVA/PDEFoam.h"
#include "TMVA/PDEFoamEvent.h"
#include "TMVA/Types.h"

#include "Rtypes.h"

#include <map>

class TString;

ClassImp(TMVA::PDEFoamMultiTarget);

////////////////////////////////////////////////////////////////////////////////
/// Default constructor for streamer, user should not use it.

TMVA::PDEFoamMultiTarget::PDEFoamMultiTarget()
: PDEFoamEvent()
   , fTargetSelection(kMean)
{
}

////////////////////////////////////////////////////////////////////////////////
/// User constructor
///
/// Parameters:
///
/// - name - name of PDEFoam object
///
/// - ts - target selection method used in
///   GetCellValue(const std::map<Int_t, Float_t>& xvec, ECellValue)
///   Candidates are: TMVA::kMean, TMVA::kMpv
///
///   - TMVA::kMean - The function GetCellValue() finds all cells
///     which contain a given event vector 'xvec' and returns the
///     mean target (for every target variable in the foam).
///
///   - TMVA::kMpv - The function GetCellValue() finds all cells
///     which contain a given event vector 'xvec' and returns the
///     most probable target (for every target variable in the
///     foam), that is the target value which corresponds to the
///     cell with the largest event density.

TMVA::PDEFoamMultiTarget::PDEFoamMultiTarget(const TString& name, ETargetSelection ts)
   : PDEFoamEvent(name)
   , fTargetSelection(ts)
{
}

////////////////////////////////////////////////////////////////////////////////
/// Copy Constructor  NOT IMPLEMENTED (NEVER USED)

TMVA::PDEFoamMultiTarget::PDEFoamMultiTarget(const PDEFoamMultiTarget &from)
   : PDEFoamEvent(from)
   , fTargetSelection(from.fTargetSelection)
{
   Log() << kFATAL << "COPY CONSTRUCTOR NOT IMPLEMENTED" << Endl;
}

////////////////////////////////////////////////////////////////////////////////
/// This function is overridden from PDFEFoam.  It returns all
/// regression targets (in order), given an untransformed event
/// vector 'xvec'.  The key of 'xvec' is the dimension and the value
/// (Float_t) is the coordinate.
///
/// Note: number of foam dimensions = number of variables + number
/// of targets
///
/// Parameters:
/// - xvec - map of event variables (no targets!)
/// - cv - cell value to return (ignored!)
///
/// Return:
/// Targets, ordered by missing dimensions in 'xvec'.
/// The size of the returned vector = foam dimension - size of xvec.

std::vector<Float_t> TMVA::PDEFoamMultiTarget::GetCellValue(const std::map<Int_t, Float_t>& xvec, ECellValue /*cv*/)
{
   // transform event vector
   std::map<Int_t, Float_t> txvec; // transformed event vector
   for (std::map<Int_t, Float_t>::const_iterator it = xvec.begin();
        it != xvec.end(); ++it) {
      Float_t coordinate = it->second; // event coordinate
      Int_t dim = it->first;           // dimension
      // checkt whether coordinate is within foam borders. if not,
      // push event coordinate into foam
      if (coordinate <= fXmin[dim])
         coordinate = fXmin[dim] + std::numeric_limits<float>::epsilon();
      else if (coordinate >= fXmax[dim])
         coordinate = fXmax[dim] - std::numeric_limits<float>::epsilon();
      // transform event
      txvec.insert(std::pair<Int_t, Float_t>(dim, VarTransform(dim, coordinate)));
   }

   // map of targets
   std::map<Int_t, Float_t> target;

   // find cells, which fit txvec
   std::vector<PDEFoamCell*> cells = FindCells(txvec);
   if (cells.empty()) {
      // return empty target vector (size = dimension of foam -
      // number of variables)
      return std::vector<Float_t>(GetTotDim() - xvec.size(), 0);
   }

   // initialize the target map
   for (Int_t idim = 0; idim < GetTotDim(); ++idim) {
      // is idim a target dimension, i.e. is idim missing in txvec?
      if (txvec.find(idim) == txvec.end())
         target.insert(std::pair<Int_t, Float_t>(idim, 0));
   }

   switch (fTargetSelection) {
   case kMean:
      CalculateMean(target, cells);
      break;
   case kMpv:
      CalculateMpv(target, cells);
      break;
   default:
      Log() << "<PDEFoamMultiTarget::GetCellValue>: "
            << "unknown target selection type!" << Endl;
      break;
   }

   // copy targets to result vector
   std::vector<Float_t> result;
   result.reserve(target.size());
   for (std::map<Int_t, Float_t>::const_iterator it = target.begin();
        it != target.end(); ++it)
      result.push_back(it->second);

   return result;
}

////////////////////////////////////////////////////////////////////////////////
/// This function calculates the most probable target value from a
/// given number of cells.  The most probable target is defined to
/// be the coordinates of the cell which has the biggest event
/// density.
///
/// Parameters:
///
/// - target - map of targets, where the key is the dimension and
///   the value is the target value.  It is assumed that this map is
///   initialized such that there is a map entry for every target.
///
/// - cells - vector of PDEFoam cells to pick the most probable
///   target from

void TMVA::PDEFoamMultiTarget::CalculateMpv(std::map<Int_t, Float_t>& target, const std::vector<PDEFoamCell*>& cells)
{
   Double_t max_dens = 0.0;            // maximum cell density

   // loop over all cells and find cell with maximum event density
   for (std::vector<PDEFoamCell*>::const_iterator cell_it = cells.begin();
        cell_it != cells.end(); ++cell_it) {

      // get event density of cell
      const Double_t cell_density = GetCellValue(*cell_it, kValueDensity);

      // has this cell a larger event density?
      if (cell_density > max_dens) {
         // get cell position and size
         PDEFoamVect  cellPosi(GetTotDim()), cellSize(GetTotDim());
         (*cell_it)->GetHcub(cellPosi, cellSize);

         // save new maximum density
         max_dens = cell_density;

         // calculate new target values
         for (std::map<Int_t, Float_t>::iterator target_it = target.begin();
              target_it != target.end(); ++target_it) {
            const Int_t dim = target_it->first; // target dimension
            target_it->second =
               VarTransformInvers(dim, cellPosi[dim] + 0.5 * cellSize[dim]);
         }
      }
   }
}

////////////////////////////////////////////////////////////////////////////////
/// This function calculates the mean target value from a given
/// number of cells.  As weight the event density of the cell is
/// used.
///
/// Parameters:
///
/// - target - map of targets, where the key is the dimension and
///   the value is the target value.  It is assumed that this map is
///   initialized such that there is a map entry for every target
///   with all target values set to zero.
///
/// - cells - vector of PDEFoam cells to pick the most probable
///   target from

void TMVA::PDEFoamMultiTarget::CalculateMean(std::map<Int_t, Float_t>& target, const std::vector<PDEFoamCell*>& cells)
{
   // normalization
   std::map<Int_t, Float_t> norm;

   // loop over all cells and find cell with maximum event density
   for (std::vector<PDEFoamCell*>::const_iterator cell_it = cells.begin();
        cell_it != cells.end(); ++cell_it) {

      // get event density of cell
      const Double_t cell_density = GetCellValue(*cell_it, kValueDensity);

      // get cell position and size
      PDEFoamVect  cellPosi(GetTotDim()), cellSize(GetTotDim());
      (*cell_it)->GetHcub(cellPosi, cellSize);

      // accumulate weighted target values
      for (std::map<Int_t, Float_t>::iterator target_it = target.begin();
           target_it != target.end(); ++target_it) {
         const Int_t dim = target_it->first; // target dimension
         target_it->second += cell_density *
            VarTransformInvers(dim, cellPosi[dim] + 0.5 * cellSize[dim]);
         norm[dim] += cell_density;
      }
   }

   // normalize the targets
   for (std::map<Int_t, Float_t>::iterator target_it = target.begin();
        target_it != target.end(); ++target_it) {

      // get target dimension
      const Int_t dim = target_it->first;

      // normalize target in dimension 'dim'
      if (norm[dim] > std::numeric_limits<Float_t>::epsilon()) {
         target[dim] /= norm[dim];
      } else {
         // normalisation factor is too small -> return approximate
         // target value
         target[dim] = (fXmax[dim] - fXmin[dim]) / 2.;
      }
   }
}