TMVAGAexample.C

Go to the documentation of this file.

/// \file
/// \ingroup tutorial_ml
/// \notebook -nodraw
/// This executable gives an example of a very simple use of the genetic algorithm
/// of TMVA
/// - Project   : TMVA - a Root-integrated toolkit for multivariate data analysis
/// - Package   : TMVA
/// - Executable: TMVAGAexample
///
/// \macro_code
/// \macro_output
/// \author Andreas Hoecker
 
#include <iostream> // Stream declarations
#include <vector>
 
#include "TMVA/GeneticAlgorithm.h"
#include "TMVA/GeneticFitter.h"
#include "TMVA/IFitterTarget.h"
 
using std::vector;
 
using namespace TMVA;
 
class MyFitness : public IFitterTarget {
    public:
       MyFitness() : IFitterTarget() {
       }
 
       // the fitness-function goes here
       // the factors are optimized such that the return-value of this function is minimized
       // take care!! the fitness-function must never fail, .. means: you have to prevent
       // the function from reaching undefined values (such as x=0 for 1/x or so)
       //
       // HINT: to use INTEGER variables, it is sufficient to cast the "factor" in the fitness-function
       // to (int). In this case the variable-range has to be chosen +1 ( to get 0..5, take Interval(0,6) )
       // since the introduction of "Interval" ranges can be defined with a third parameter
       // which gives the number of bins within the interval. With that technique discrete values
       // can be achieved easier. The random selection out of this discrete numbers is completely uniform.
       //
       Double_t EstimatorFunction( std::vector<Double_t> & factors ){
           //return (10.- (int)factors.at(0) *factors.at(1) + (int)factors.at(2));
           return (10.- factors.at(0) *factors.at(1) + factors.at(2));
 
           //return 100.- (10 + factors.at(1)) *factors.at(2)* TMath::Abs( TMath::Sin(factors.at(0)) );
       }
};
 
 
class MyGA2nd : public GeneticAlgorithm {
    public:
       MyGA2nd( IFitterTarget& target, Int_t size, vector<Interval*>& ranges ) : GeneticAlgorithm(target,
       size, ranges ){
       }
 
 
       // this method has to be activated if one wants to change the behaviour of the evolution
       // works only with the head version
       //void Evolution(){
       //    fSexual = true;
       //    if (fSexual) {
       //       fPopulation.MakeCopies( 5 );
       //       fPopulation.MakeChildren();
       //       fPopulation.NextGeneration();
 
       //       fPopulation.Mutate( 10, 3, kTRUE, fSpread, fMirror );
       //       fPopulation.Mutate( 40, fPopulation.GetPopulationSize()*3/4 );
       //    } else {
       //       fPopulation.MakeCopies( 3 );
       //       fPopulation.MakeMutants(100,true, 0.1, true);
       //       fPopulation.NextGeneration();
       //    }
      // }
};
 
 
 
void TMVAGAexample() {
 
   std::cout << "Start Test TMVAGAexample" << std::endl
             << "========================" << std::endl
             << "\nEXAMPLE"              << std::endl;
   // define all the parameters by their minimum and maximum value
   // in this example 3 parameters are defined.
   vector<Interval*> ranges;
   ranges.push_back( new Interval(0,15,30) );
   ranges.push_back( new Interval(0,13) );
   ranges.push_back( new Interval(0,5,3) );
 
   for( std::vector<Interval*>::iterator it = ranges.begin(); it != ranges.end(); it++ ){
      std::cout << " range: " << (*it)->GetMin() << "   " << (*it)->GetMax() << std::endl;
   }
 
   IFitterTarget* myFitness = new MyFitness();
 
   // prepare the genetic algorithm with an initial population size of 20
   // mind: big population sizes will help in searching the domain space of the solution
   // but you have to weight this out to the number of generations
   // the extreme case of 1 generation and populationsize n is equal to
   // a Monte Carlo calculation with n tries
 
   MyGA2nd mg( *myFitness, 100, ranges );
   // mg.SetParameters( 4, 30, 200, 10,5, 0.95, 0.001 );
 
   #define CONVSTEPS 20
   #define CONVCRIT 0.0001
   #define SCSTEPS 10
   #define SCRATE 5
   #define SCFACTOR 0.95
 
   do {
      // prepares the new generation and does evolution
      mg.Init();
 
      // assess the quality of the individuals
      mg.CalculateFitness();
 
      mg.GetGeneticPopulation().Print(0);
      std::cout << "---" << std::endl;
 
      // reduce the population size to the initially defined one
      mg.GetGeneticPopulation().TrimPopulation();
 
      // tricky thing: control the speed of how fast the "solution space" is searched through
      // this function basically influences the sigma of a gaussian around the actual value
      // of the parameter where the new value will be randomly thrown.
      // when the number of improvements within the last SCSTEPS
      // A) smaller than SCRATE: divide the preset sigma by SCFACTOR
      // B) equal to SCRATE: do nothing
      // C) greater than SCRATE: multiply the preset sigma by SCFACTOR
      // if you don't know what to do, leave it unchanged or even delete this function call
      mg.SpreadControl( SCSTEPS, SCRATE, SCFACTOR );
 
   } while (!mg.HasConverged( CONVSTEPS, CONVCRIT ));  // converged if: fitness-improvement < CONVCRIT within the last CONVSTEPS loops
 
   GeneticGenes* genes = mg.GetGeneticPopulation().GetGenes( 0 );
   std::vector<Double_t> gvec;
   gvec = genes->GetFactors();
   int n = 0;
   for( std::vector<Double_t>::iterator it = gvec.begin(); it<gvec.end(); it++ ){
      std::cout << "FACTOR " << n << " : " << (*it) << std::endl;
      n++;
   }
}
 
 
int main( int argc, char** argv )
{
   TMVAGAexample();
}