nsga2.cpp

/********************************************************************************
 *  FARSA Genetic Algorithm Library                                             *
 *  Copyright (C) 2007-2011 Gianluca Massera <emmegian@yahoo.it>                *
 *                                                                              *
 *  This program is free software; you can redistribute it and/or modify        *
 *  it under the terms of the GNU General Public License as published by        *
 *  the Free Software Foundation; either version 2 of the License, or           *
 *  (at your option) any later version.                                         *
 *                                                                              *
 *  This program is distributed in the hope that it will be useful,             *
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of              *
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the               *
 *  GNU General Public License for more details.                                *
 *                                                                              *
 *  You should have received a copy of the GNU General Public License           *
 *  along with this program; if not, write to the Free Software                 *
 *  Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA  *
 ********************************************************************************/

#include "factory.h"
#include "gas/nsga2.h"
#include "core/reproduction.h"
#include "core/genome.h"
#include "core/evaluation.h"
#include "configurationhelper.h"
#include <cmath>
#include <cfloat>
#include <QThreadPool>
#include <QtConcurrentMap>
using namespace QtConcurrent;

namespace farsa {

NSGA2::NSGA2()
    : GeneticAlgo(), lastPareto() {
    fitfunc = 0;
    reprod = 0;
    numGens = 0;
    currPhase = initEvaluation;
    numThreadv = 1;
    isInitialized = false;
    isFinalized = true;
}

NSGA2::~NSGA2() {
    foreach( NSGA2::evaluationThread* e, evalThreads ) {
        delete e;
    }
    delete fitfunc;
}

void NSGA2::configure( ConfigurationParameters& params, QString prefix ) {
    setGenome( params.getObjectFromGroup<Genome>( prefix + QString( "GENOME" ) ) );
    setEvaluation( params.getObjectFromGroup<Evaluation>( prefix + QString( "EVALUATION" ) ) );
    fitfunc->setGenome( genome() );
    setReproduction( params.getObjectFromGroup<Reproduction>( prefix + QString( "REPRODUCTION" ) ) );
    setNumGenerations( ConfigurationHelper::getInt( params, prefix + QString( "ngenerations" ), 1000 ) );
    setNumThreads( ConfigurationHelper::getInt( params, prefix + QString( "numThreads" ), 1 ) );
}

void NSGA2::save( ConfigurationParameters& params, QString prefix ) {
    params.createParameter( prefix, QString("type"), "NSGA2" );
    params.createParameter( prefix, QString("numThreads"), QString("%1").arg( numThreads() ) );
    params.createParameter( prefix, QString("ngenerations"), QString("%1").arg( numGenerations() ) );
    //--- EVALUATION
    fitfunc->save( params, params.createSubGroup( prefix, "EVALUATION" ) );
    //--- REPRODUCTION
    reproduction()->save( params, params.createSubGroup( prefix, "REPRODUCTION" ) );
    //--- GENOME
    genome()->save( params, params.createSubGroup( prefix, "GENOME" ) );
}

void NSGA2::describe( QString type ) {
    Descriptor d = addTypeDescription( type, "Non-dominated Sorting Genetic Algorithm version 2" );
    d.describeInt( "numThreads" ).limits( 1, 32 ).def(1).help( "Number of threads to parallelize the evaluation of individuals" ).runtime(&NSGA2::setNumThreads, &NSGA2::numThreads);;
    d.describeInt( "ngenerations" ).limits( 1, INT_MAX ).def( 1000 ).help( "Number of the generations of the evolutionary process" );
    d.describeSubgroup( "EVALUATION" ).type( "Evaluation" ).props( IsMandatory ).help( "Object that calculate the fitness", "Create a subclass of Evalution and code your custom fitness function" );
    d.describeSubgroup( "REPRODUCTION" ).type( "Reproduction" ).props( IsMandatory ).help( "Object that generate the new generations" );
    d.describeSubgroup( "GENOME" ).type( "Genome" ).props( IsMandatory ).help( "Object containing the individuals under evolution" );
}

void NSGA2::setNumThreads( int numThreads ) {
    if ( numThreads < 1 ) {
        qWarning( "The number of Threads must be greater than one!!" );
    }
    Q_ASSERT_X( !isInitialized && isFinalized ,
            "NSGA2::setNumThreads",
            "This method can only called before initialize of NSGA2" );
    Q_ASSERT_X( fitfunc != 0 ,
            "NSGA2::setNumThreads",
            "This method must be called after an Evaluation object has been setted by NSGA2::setEvaluation" );
    numThreadv = numThreads;
    return;
}

int NSGA2::numThreads() const {
    return numThreadv;
}

void NSGA2::setEvaluation( Evaluation* fitfunc ) {
    this->fitfunc = fitfunc;
    this->fitfunc->setGA( this );
}

Evaluation* NSGA2::evaluationPrototype()
{
    return fitfunc;
}

QVector<Evaluation*> NSGA2::evaluationPool() {
    QVector<Evaluation*> ev;
    foreach( NSGA2::evaluationThread* e, evalThreads ) {
        ev.append( e->eval );
    }
    return ev;
}

void NSGA2::setReproduction( Reproduction* reprod ) {
    this->reprod = reprod;
    this->reprod->setGA( this );
}

Reproduction* NSGA2::reproduction() {
    return reprod;
}

void NSGA2::initialize() {
    if ( isInitialized && !isFinalized ) return;
    Q_ASSERT_X( fitfunc != 0 ,
            "NSGA2::initialize",
            "You have to setup the Evaluation object of NSGA2 (Fitness Function)" );
    Q_ASSERT_X( reprod !=0 ,
            "NSGA2::initialize",
            "You have to setup the Reproduction operator of NSGA2" );

    isInitialized = true;
    isFinalized = false;
    setGeneration( 0 );
    currPhase = initEvaluation;
    evolutionEnd = false;
    evaluationDone = false;
    getIODelegate()->recoverData(this);
    //--- Setting up the evalThreads
    for( int i=0; i<evalThreads.size(); i++ ) {
        delete (evalThreads[i]);
    }
    evalThreads.clear();
    for( int i=0; i<numThreadv; i++ ) {
        evalThreads.append( new evaluationThread( this, fitfunc ) );
        evalThreads.last()->eval->initGeneration( generation() );
    }
    //--- set the number of thread to create
    QThreadPool::globalInstance()->setMaxThreadCount( numThreadv );
}

void NSGA2::gaStep() {
    switch( currPhase ) {
    case initEvaluation:
        for( int i=0; i<numThreadv; i++ ) {
            evalThreads[i]->sequence.clear();
        }
        for( int i=0; i<(int)genome()->size(); i++ ) {
            evalThreads[ i%numThreadv ]->sequence.append( i );
        }
        for( int i=0; i<numThreadv; i++ ) {
            evalThreads[i]->idSeq = 0;
            evalThreads[i]->id = evalThreads[i]->sequence[ evalThreads[i]->idSeq ];
            evalThreads[i]->eval->setGenome( genome() );
            evalThreads[i]->eval->initialize( genome()->at( evalThreads[i]->id ) );
            evalThreads[i]->blocked = false;
        }
        currPhase = evaluating;
        break;
    case evaluating: { /* Multi Thread Block (i.e. Parallel Evaluation of Genotypes */
        nextGeneration = true;
        if ( numThreadv == 1 ) {
            // Don't use Threads if is not necessary
            evalThreads[0]->runStep();
        } else {
            QFuture<void> future = map( evalThreads, NSGA2::runStepWrapper );
            future.waitForFinished();
        }
        if ( nextGeneration ) {
            currPhase = nextGeneration_pass1;
        }
        } /* End of Multi Thread Block */
        break;
    case nextGeneration_pass1: {
        // --- Core of the NSGA-II Algorithm
        //   -1) merge previous front population lastPareto with current genome()
        //--- this merge assure that the best (elite) pareto-fronts are mantained
        nsgaGenome allGenome;
        for( unsigned int i=0; i<lastPareto.size(); i++ ) {
            allGenome.append( new nsgaGenotype( lastPareto.at(i), 0, 0.0 ) );
        }
        for( unsigned int i=0; i<genome()->size(); i++ ) {
            allGenome.append( new nsgaGenotype( genome()->at(i), 0, 0.0 ) );
        }
        //   -2) fastNonDominatedSort( onMergedPopulation )
        //--- genotype grouped by the front rank
        QVector<nsgaGenome> frontsByRank = fastNonDominatedSort( allGenome );
        //   -3) create new population of size genome().size() using the fronts
        //       calculated and use the crowdingDistanceAssignment
        unsigned int currentGenotype = 0;
        int numOfFronts = frontsByRank.size();
        int numObjs = allGenome[0]->genotype->numOfObjectives();
        for( int front=0; front<numOfFronts; front++ ) {
            //--- sort it on crowding distance and add to the new Genome
            crowdingDistanceAssignment( frontsByRank[front] );
            qStableSort( frontsByRank[front].begin(), frontsByRank[front].end(), NSGA2::crowdingDistanceGreaterThan );
            foreach( nsgaGenotype* gen, frontsByRank[front] ) {
                //--- assign the rank
                gen->genotype->setRank( 2*(numOfFronts - front) + (gen->distance/numObjs) );
                genome()->set( currentGenotype, gen->genotype );
                currentGenotype++;
                if ( currentGenotype == genome()->size() ) break;
            }
            if ( currentGenotype == genome()->size() ) break;
        }
        //   -4) lastPareto <- genome()
        //--- this pass correspond to elite the pareto-fronts
        lastPareto = *(genome());
        //   -5) clean up memory
        for( int i=0; i<allGenome.size(); i++ ) {
            delete (allGenome[i]);
        }
        updateStats();
        evaluationDone = true;
        for( int i=0; i<numThreadv; i++ ) {
            evalThreads[i]->eval->endGeneration( generation() );
        }
        notifyEndGeneration();
        getIODelegate()->saveData(this);
        if ( generation() < numGens ) {
            currPhase = nextGeneration_pass2;
        } else {
            currPhase = endEvolution;
        }
        }
        break;
    case nextGeneration_pass2: {
        //--- this additional pass is for avoid to modify the
        //--- genotypes contained in the last generation of the evolution
        //--- and to allow to save the genotypes after each generation
        Genome* old = genome();
        setGenome( reprod->reproduction( old ) );
        delete old;
        fitfunc->setGenome( genome() );
        evaluationDone = false;
        setGeneration( generation()+1 );
        for( int i=0; i<numThreadv; i++ ) {
            evalThreads[i]->eval->initGeneration( generation() );
        }
        currPhase = initEvaluation;
        }
        break;
    case endEvolution:
        finalize();
        break;
    default:
        qFatal( "Default switch in NSGA2::gaStep" );
        break;
    }
}

void NSGA2::skipEvaluation() {
    // Set evaluation done, and check which phase to go to
    evaluationDone = true;
    if ( generation() < numGens ) {
        currPhase = nextGeneration_pass2;
    } else {
        currPhase = endEvolution;
    }
}

void NSGA2::finalize() {
    if ( isFinalized && !isInitialized ) return;

    isInitialized = false;
    isFinalized = true;
    evolutionEnd = true;
}

QVector<NSGA2::nsgaGenome> NSGA2::fastNonDominatedSort( nsgaGenome& pareto ) {
    QMap<nsgaGenotype*, nsgaGenome> dominatedBy;
    QVector<nsgaGenome> frontsByRank;
    frontsByRank.resize(1);
    //--- create the first front containing the top solutions
    for( int p=0; p<pareto.size(); p++ ) {
        //--- domination counter of genP reset to zero
        pareto[p]->dominationCounter = 0;
        for( int q=0; q<pareto.size(); q++ ) {
            if ( p==q ) continue;
            if ( pareto[q]->genotype->dominatedBy( pareto[p]->genotype ) ) {
                // OPTIMIZE: it seems that most of the time is spent on accessing the QMap
                dominatedBy[ pareto[p] ].append( pareto[q] );
            } else if ( pareto[p]->genotype->dominatedBy( pareto[q]->genotype ) ) {
                pareto[p]->dominationCounter++;
            }
        }
        //--- if nP == 0 means that genP is a solution belongs to the top pareto-front
        if ( pareto[p]->dominationCounter == 0 ) {
            pareto[p]->rank = 0;
            frontsByRank[0].append( pareto[p] );
        }
    }
    //--- create all the other fronts
    bool done = false;
    int currentFront = 0;
    while( !done ) {
        nsgaGenome newFront;
        for( int i=0; i<frontsByRank[currentFront].size(); i++ ) {
            nsgaGenotype* p = frontsByRank[currentFront][i];
            nsgaGenome pDominate = dominatedBy[p];
            for( int q=0; q<pDominate.size(); q++ ) {
                pDominate[q]->dominationCounter--;
                if ( pDominate[q]->dominationCounter == 0 ) {
                    pDominate[q]->rank = currentFront+1;
                    newFront.append( pDominate[q] );
                }
            }
        }
        currentFront++;
        if ( newFront.isEmpty() ) {
            done = true;
        } else {
            frontsByRank.append( newFront );
        }
    }
    int total = 0;
    for( int i=0; i<frontsByRank.size(); i++ ) {
        total += frontsByRank[i].size();
    }
    return frontsByRank;
}

void NSGA2::crowdingDistanceAssignment( nsgaGenome& genome ) {
    int dimGenome = genome.size();
    if ( dimGenome == 0 ) return;
    int numObjs = genome[0]->genotype->numOfObjectives();
    //--- vectors containing the max and min values of objectives in this genome
    QVector<double> fmax;
    QVector<double> fmin;
    fmax.resize( numObjs );
    fmin.resize( numObjs );
    for( int i=0; i<numObjs; i++ ) {
        fmax[i] = genome[0]->genotype->objective( i );
        fmin[i] = genome[0]->genotype->objective( i );
    }
    //--- initialize distance and calculate fmax and fmin values
    for( int i=0; i<dimGenome; i++ ) {
        genome[i]->distance = 0;
        for( int m=0; m<numObjs; m++ ) {
            fmax[m] = qMax( fmax[m], genome[i]->genotype->objective(m) );
            fmin[m] = qMin( fmin[m], genome[i]->genotype->objective(m) );
        }
    }
    //--- calculate the distance
    nObjectiveGreaterThan objCompare;
    for( int m=0; m<numObjs; m++ ) {
        // currentObjective is used by nObjectiveGreaterThan for sorting
        objCompare.currentObjective = m;
        qStableSort( genome.begin(), genome.end(), objCompare );
        // the maximum value is numObj, setting to numObj assure that this two
        // genotypes are always the top in the current front
        genome[0]->distance = numObjs; //DBL_MAX;
        genome.last()->distance = numObjs; //DBL_MAX;
        for( int i=1; i<dimGenome-1; i++ ) {
            double m1 = genome[i+1]->genotype->objective(m);
            double m2 = genome[i-1]->genotype->objective(m);
            genome[i]->distance += fabs(m1-m2)/(fmax[m]-fmin[m]);
            // if the value is nan, then it will setted to zero (worst distance)
            if ( genome[i]->distance != genome[i]->distance ) {
                genome[i]->distance = 0.0;
            }
        }
    }
}

NSGA2::evaluationThread::evaluationThread( NSGA2* p, Evaluation* eProto )
    : parent(p), id(0), blocked(false) {
    eval = eProto->clone();
    eval->setGenome( p->genome() );
    eval->setGA( p );
}

NSGA2::evaluationThread::~evaluationThread() {
    delete eval;
    sequence.clear();
}

void NSGA2::evaluationThread::runStep() {
    if ( blocked ) {
        return;
    }
    
    parent->nextGeneration = false;
    eval->evaluateStep();
    if ( eval->isEvaluationDone() ) {
        int nextIdSeq = idSeq + 1;
        eval->finalize();
        if ( nextIdSeq >= sequence.size() ) {
            blocked = true;
            return;
        }
        idSeq = nextIdSeq;
        int nextId = sequence[ idSeq ];
        id = nextId;
        eval->initialize( parent->genome()->at( id ) );
    }
}

void NSGA2::runStepWrapper( NSGA2::evaluationThread* e ) {
    e->runStep();
}

} // end namespace farsa