evoga.cpp

/********************************************************************************
 *  FARSA Experiments Library                                                   *
 *  Copyright (C) 2007-2012                                                     *
 *  Stefano Nolfi <stefano.nolfi@istc.cnr.it>                                   *
 *  Onofrio Gigliotta <onofrio.gigliotta@istc.cnr.it>                           *
 *  Gianluca Massera <emmegian@yahoo.it>                                        *
 *  Tomassino Ferrauto <tomassino.ferrauto@istc.cnr.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 "evoga.h"
#include "evodataviewer.h"
#include "logger.h"
#include "randomgenerator.h"
#include <configurationhelper.h>
#include <QVector>
#include <QThreadPool>
#include <QtConcurrentMap>
#include <QtAlgorithms>
#include <QTime>

// All the suff below is to avoid warnings on Windows about the use of unsafe
// functions. This should be only a temporary workaround, the solution is stop
// using C string and file functions...
#if defined(_MSC_VER)
    #pragma warning(push)
    #pragma warning(disable:4996)
#endif

namespace farsa {

class EvaluatorThreadForEvoga
{
public:
    EvaluatorThreadForEvoga(Evoga *ga, EvoRobotExperiment *exp) :
        m_ga(ga),
        m_exp(exp)
    {
    }

    ~EvaluatorThreadForEvoga()
    {
        delete m_exp;
    }

    void setGenotype(int id)
    {
        m_id = id;
        m_exp->setNetParameters(m_ga->getGenes(id));
    }

    int getGenotypeId() const
    {
        return m_id;
    }

    void run()
    {
        m_exp->doAllTrialsForIndividual(m_id);
        m_fitness = m_exp->getFitness();
    }

    double getFitness() const
    {
        return m_fitness;
    }

    EvoRobotExperiment* getExperiment()
    {
        return m_exp;
    }

private:
    Evoga *const m_ga;

    EvoRobotExperiment *const m_exp;

    int m_id;

    double m_fitness;
};

void runEvaluatorThreadForEvoga(EvaluatorThreadForEvoga* e)
{
    e->run();
}

struct FitnessAndId
{
    float fitness;

    int id;
};

bool operator<(FitnessAndId first, FitnessAndId second)
{
    return (first.fitness < second.fitness);
}

int Evoga::mrand(int i)
{
    int r;

    r = rand();
    r = r % (int) i;
    return r;

}

Evoga::Evoga()
{
    exp = NULL;
    mutations = NULL;
    tfitness = NULL;
    statfit = NULL;
    ntfitness = NULL; //used with super ag implementation
    loadedIndividuals = 1;
    numThreads = 1;
    savePopulationEachNGenerations = 0;
    stopEvolution = false;
    averageIndividualFitnessOverGenerations = true;
    isStepByStep = false;
}

Evoga::~Evoga()
{
    delete exp;
    delete[] tfitness;
    if (statfit != NULL) {
        for(int i = 0; i < nogenerations; i++) {
            delete[] statfit[i];
        }
    }
    delete[] statfit;
    delete[] ntfitness;
    delete[] mutations;
}

void Evoga::setSeed(int s)
{
    srand(s);
    globalRNG->setSeed( s );
    currentSeed = s;
}

//return a random value between 0-1
double Evoga::drand()
{
    return (double)rand()/(double)RAND_MAX;
}

double Evoga::getNoise(double minn, double maxn)
{
    double nrange;
    if(maxn>minn)
        nrange=maxn-minn;
    else
        nrange=minn-maxn;

    return drand()*nrange+minn;
}

int Evoga::mutate(int w, double mut)
{
    int b[8];
    int val;
    int ii;

    val = w;
    for(ii=0;ii < 8;ii++) {
        b[ii] = val % 2;
        val  = val / 2;
    }
    for(ii=0;ii < 8;ii++) {
        //if (mrand(100) < percmut)
        if(drand()<mut) {
            b[ii] =(b[ii]+1)%2; // con questa modifica il bit switcha //mrand(2);
        }
    }
    w = 0;
    w += b[0] * 1;
    w += b[1] * 2;
    w += b[2] * 4;
    w += b[3] * 8;
    w += b[4] * 16;
    w += b[5] * 32;
    w += b[6] * 64;
    w += b[7] * 128;
 
    return(w);
}

void Evoga::putGenome(int fromgenome, int tobestgenome)
{
    if (tobestgenome < this->nreproducing) {
        for (int i = 0; i < this->glen; i++) {
            bestgenome[tobestgenome][i] = genome[fromgenome][i];
        }
    } else {
        Logger::error("putGenomeError!");
    }
}

void Evoga::getGenome(int frombestgenome, int togenome, int mut)
{
    for(int i = 0; i < this->glen; i++) {
        if (mut == 0) {
            genome[togenome][i] = bestgenome[frombestgenome][i];
        } else {
            if (mutations[i] == Evonet::DEFAULT_VALUE) {    //standard mutation
                genome[togenome][i] = mutate(bestgenome[frombestgenome][i], mutation);
            } else {                //specific mutation
                genome[togenome][i] = mutate(bestgenome[frombestgenome][i], mutations[i]);
            }
        }
    }
}

void Evoga::copyGenes(int from, int to, int mut)
{
    for(int i = 0; i < this->glen; i++) {
        if (mut == 0) {
            genome[to][i] = genome[from][i];
        } else {
            if (mutations[i] == Evonet::DEFAULT_VALUE) {    //standard mutation
                genome[to][i] = mutate(genome[from][i], mutation);
            } else {                //specific mutation
                genome[to][i] = mutate(genome[from][i], mutations[i]);
            }
        }
    }
}

//Reproduce individuals with higher ranking
void Evoga::reproduce()
{
    //to do
    //selecting best fathers
    int i;
    int bi,bx;
    double bn;

    char sbuffer[64];
    FILE *fp;

    //first of all we compute fitness stat
    this->computeFStat();


    for(bi=0;bi<this->nreproducing;bi++) {
        bn=-9999.0;
        bx=-1; //individual to be copied
        for(i=0;i< this->popSize;i++) {
            if(tfitness[i]>bn) {
                bn=tfitness[i];
                bx=i;
            }
        }

        this->putGenome(bx,bi);

        //here we save best genome
        if ((bi+1)<=this->savebest && cgen< this->nogenerations) {
            sprintf(sbuffer,"B%dS%d.gen",bi+1,this->currentSeed);
            if (cgen==0) {
                if ((fp=fopen(sbuffer, "w")) == NULL) {
                    Logger::error(QString("I cannot open file B%1S%2.gen").arg(bi+1).arg(this->currentSeed));
                }
            } else {
                if ((fp=fopen(sbuffer, "a")) == NULL) {
                    Logger::error(QString("I cannot open file B%1S%2.gen").arg(bi+1).arg(this->currentSeed));
                }
            }

            fprintf(fp,"**NET : s%d_%d.wts\n",cgen,bx);
            saveagenotype(fp,bx);
            fflush(fp);
            fclose(fp);
        }
        tfitness[bx]=-9999.0;
    }

    //reproducing best
    bx=0;
    for(bi=0;bi<this->nreproducing;bi++)
        for(i =0;i< this->noffspring;i++) {
            if(this->elitism && bi==0)
                this->getGenome(bi,bx,0);
            else
                this->getGenome(bi,bx,1);// 1 mette mutazione

            bx++;
        }

    //resetting fitness
    for (i=0;i<this->popSize;i++) tfitness[i]=0.0;
    this->saveFStat();
    cgen++;
}

void Evoga::saveBestInd()
{
    //to do
    //selecting best fathers
    int i;
    int bi;
    double bn, ffit;
    bn=-999999.0;// it was 0 possible source of bug in case of negsative fitness
    bi=-1;

    char sbuffer[64];
    FILE *fp;

    sprintf(sbuffer,"B%dS%d.gen",bi+1,this->currentSeed);
    if (cgen==0) {
        if ((fp=fopen(sbuffer, "w")) == NULL) {
            Logger::error(QString("I cannot open file B%1S%2.gen").arg(bi+1).arg(this->currentSeed));
        }
    } else {
        if ((fp=fopen(sbuffer, "a")) == NULL) {
            Logger::error(QString("I cannot open file B%1S%2.gen").arg(bi+1).arg(this->currentSeed));
        }
    }

    //finding the best simply the best, one individual
    for(i=0;i<this->popSize;i++) {
        ffit=this->tfitness[i]/this->ntfitness[i];
        if(ffit>bn) {
            bn=ffit;
            bi=i;
        }
    }

    //now saving
    fprintf(fp,"**NET : s%d_%d.wts\n",cgen,bi);
    saveagenotype(fp,bi);
    fflush(fp);
    fclose(fp);
}

//Reproduce individuals with higher ranking
void Evoga::mreproduce()
{
    //to do
    //selecting best fathers
    int i;
    int bi,bx;
    double bn;

    char sbuffer[64];
    FILE *fp;

    //first of all we compute fitness stat
    this->computeFStat();


    for(bi=0;bi<this->nreproducing;bi++) {
        bn=9999.0;
        bx=-1; //individual to be copied
        for(i=0;i< this->popSize;i++) {
            if(tfitness[i]<bn) {
                bn=tfitness[i];
                bx=i;
            }
        }
    
        this->putGenome(bx,bi);
    
        //here we save best genome
        if ((bi+1)<=this->savebest && cgen< this->nogenerations) {
            sprintf(sbuffer,"B%dS%d.gen",bi+1,this->currentSeed);
            if (cgen==0) {
                if ((fp=fopen(sbuffer, "w")) == NULL) {
                    Logger::error(QString("I cannot open file B%1S%2.gen").arg(bi+1).arg(this->currentSeed));
                }
            } else {
                if ((fp=fopen(sbuffer, "a")) == NULL) {
                    Logger::error(QString("I cannot open file B%1S%2.gen").arg(bi+1).arg(this->currentSeed));
                }
            }

            fprintf(fp,"**NET : s%d_%d.wts\n",cgen,bx);
            saveagenotype(fp,bx);
            fflush(fp);
            fclose(fp);
        }
        tfitness[bx]=9999.0;
    }

    //reproducing best
    bx=0;
    for(bi=0;bi<this->nreproducing;bi++)
        for(i =0;i< this->noffspring;i++) {
            if(this->elitism && bi==0)
                this->getGenome(bi,bx,0);
            else
                this->getGenome(bi,bx,1);

            bx++;
        }

    //resetting fitness
    for (i=0;i<this->popSize;i++) tfitness[i]=9999.0;
    this->saveFStat();
    cgen++;
}


void Evoga::printPop()
{
    for(int i = 0; i < this->popSize; i++) {
        QString output = QString("Fit %1 | ").arg(tfitness[i]);
        for(int l = 0; l < this->glen; l++) {
            output += QString("%1 ").arg(this->genome[i][l]);
        }
        Logger::info(output);
    }
}

void Evoga::printBest()
{
    for(int i = 0; i < bestgenome.size(); i++) {
        QString output = QString("Best %d | ").arg(i);
        for (int s = 0; s < this->glen; s++) {
            output += QString("%1 ").arg(this->bestgenome[i][s]);
        }
        Logger::info(output);
    }
}

void Evoga::computeFStat()
{
    int i;
    double min, max, av;

    min=max=tfitness[0];
    av=0.0;

    for(i=0;i<this->popSize;i++) {
        if(tfitness[i]<min) min=tfitness[i];
        if(tfitness[i]>max) max=tfitness[i];
        av+=tfitness[i];
    }
    this->faverage=av/(double)this->popSize;
    this->fmax=max;
    this->fmin=min;
    this->statfit[this->cgen][0]=this->fmax;
    this->statfit[this->cgen][1]=this->faverage;
    this->statfit[this->cgen][2]=this->fmin;
}

void Evoga::computeFStat2()
{
    int i;
    double min, max, av;

    //min=max=tfitness[0]/ntfitness[0];
    //try to fix a problem
    min=9999.00;
    max=-9999.00;
    av=0.0;

    for(i=0;i<this->popSize;i++) {
        if((tfitness[i]/ntfitness[i])<min) {
            min=tfitness[i]/ntfitness[i];
        }
        if((tfitness[i]/ntfitness[i])>max) max=tfitness[i]/ntfitness[i];
        av+=(tfitness[i]/ntfitness[i]);
    }
    this->faverage=av/(double)this->popSize;
    this->fmax=max;
    this->fmin=min;
    this->statfit[this->cgen][0]=this->fmax;
    this->statfit[this->cgen][1]=this->faverage;
    this->statfit[this->cgen][2]=this->fmin;
}

void Evoga::saveagenotype(FILE *fp, int ind)
{
    int j;
    fprintf(fp, "DYNAMICAL NN\n");
    for (j=0; j < this->glen; j++)
        fprintf(fp, "%d\n", this->genome[ind][j]);
    fprintf(fp, "END\n");
}

//save all current generation
void Evoga::saveallg()
{
    FILE *fp;
    char filename[64];
    int i;

    sprintf(filename,"G%dS%d.gen",cgen,currentSeed);
    if ((fp=fopen(filename, "w+")) == NULL) {
        Logger::error(QString("Cannot open file %1").arg(filename));
    } else {
        //we save
        for(i=0;i<this->popSize;i++) {
            fprintf(fp,"**NET : %d_%d_%d.wts\n",cgen,0,i);
            this->saveagenotype(fp,i);
        }
        fclose( fp );
    }
}

void Evoga::saveFStat()
{
    FILE *fp;
    char sbuffer[128];
    sprintf(sbuffer,"statS%d.fit",currentSeed);
        if (cgen == 0)
        fp=fopen(sbuffer , "w");
    else
        fp=fopen(sbuffer , "a");

    if (fp != NULL) {
        fprintf(fp,"%.3f %.3f %.3f\n",fmax,faverage,fmin);
        fclose(fp);
    } else
        Logger::error("unable to save statistics on a file");
}

void Evoga::getLastFStat( double &min, double &max, double &average ) {
    min = fmin;
    max = fmax;
    average = faverage;
}

void Evoga::loadgenotype(FILE *fp, int ind)
{
    int j;
    int v;

    fscanf(fp, "DYNAMICAL NN\n");
    for (j=0; j <this->glen; j++) {
        fscanf(fp,"%d\n",&v);//this->genome[ind][j]);
        this->genome[ind][j]=v;
    }
    fscanf(fp, "END\n");
}

int Evoga::loadallg(int gen, const char *filew)
{
    FILE *fp;
    char filename[512];//[64];
    char message[512];//[128];
    char flag[512];//[64];

    if (gen >= 0) {
        sprintf(filename, "G%dS%d.gen", gen, seed);
    } else {
        sprintf(filename, "%s", filew);//in case of B%P$S.gen
    }

    if ((fp = fopen(filename, "r")) != NULL) {
        genome.clear();
        while (true) {
            flag[0] = '\0'; //sprintf(flag,""); //flag = NULL;
            fscanf(fp, "%s : %s\n", flag, message);
            if (strcmp(flag, "**NET") == 0) {
                loadgenotype(fp, genome.addOne());
            } else {
                break;
            }
        }
        Logger::info(QString("Loaded ind: %1").arg(genome.size()));
        fclose(fp);
    } else {
        Logger::error(QString("File %1 could not be opened").arg(filename));
    }

    loadedIndividuals = genome.size();

    return genome.size();
}

/*
 * load a .fit file (return the number loaded individuals, 0 if the file does not exists)
 */
int Evoga::loadStatistics(char *filename)
{
    FILE *fp;
    int loaded=0;
    int i=0;
    float max,av,min;
    max=min=av=-1;
    if ((fp=fopen(filename, "r")) != NULL) {
        while(fscanf(fp,"%f %f %f\n",&max,&av,&min)!=EOF) {
            this->statfit[i][0]=max;
            this->statfit[i][1]=av;
            this->statfit[i][2]=min;
            i++;
        }
        loaded=i; 
        fflush(fp);
        fclose(fp);
        return(loaded);
    } else {
        return(0);
    }
}

void Evoga::randomizePop()
{
    for (int i = 0; i < genome.size(); i++) {
        for(int g = 0; g < glen; g++) {
            genome[i][g] = mrand(256);
        }
    }
}

void Evoga::setInitialPopulation(int* ge)
{
    int i,g;

    //fill genome with .phe parameters
    for(i=0; i<genome.size(); i++)
        for(g=0; g<glen; g++)
        {
            if(ge[g] == Evonet::DEFAULT_VALUE)
                genome[i][g] = mrand(256);
            else
                genome[i][g] = ge[g];
        }
}

void Evoga::setMutations(float* mut)
{
    //fill mutation vector
    for(int i=0; i<glen; i++)
        mutations[i] = mut[i];
}

int* Evoga::getGenes(int ind)
{
    return this->genome[ind];
}

int* Evoga::getBestGenes(int ind)
{
    return this->bestgenome[ind];
}

void Evoga::resetGenerationCounter()
{
    this->cgen=0;
}

void Evoga::getPheParametersAndMutationsFromEvonet()
{
    ResourcesLocker locker(this);

    Evonet* evonet = getResource<Evonet>( "evonet" );
    if (evonet->pheFileLoaded()) {
        //temporary mutations vector
        Evonet* evonet = getResource<Evonet>( "evonet" );
        float* muts = new float[evonet->freeParameters()];
        int* pheParams = new int[evonet->freeParameters()];

        //setting the mutation vector
        evonet->getMutations(muts); //taking it from the net
        setMutations(muts);     //pushing it inside GA

        //setting initial genome
        evonet->copyPheParameters(pheParams);   //copy *phe parameters
        setInitialPopulation(pheParams);            //put *phe parameters inside genome

        delete[] muts;
        delete[] pheParams;
    }
}

/*
 * Main function of the Genetic Algorithm (Steady State Version)
 */
void Evoga::evolveSteadyState()
{
    int rp; //replication
    int gn; //generation
    int id; //individuals
    double fit;
    double minfit=9999;
    int    minid=-1;
    double mfit;
    float  final_mrate;
    int startGeneration=0;
    char statfile[128];
    char genFile[128];

    // Resizing genome
    genome.resize(popSize * 2);

    final_mrate = mutation;
    Logger::info("EVOLUTION: steady state");
    Logger::info("Number of replications: " + QString::number(nreplications));

    // Creating evaluator objects in case of a multithread simulation. Also setting the actual number of threads used
    QVector<EvaluatorThreadForEvoga*> evaluators(popSize, NULL);
    if (numThreads > 1) {
        for (int i = 0; i < evaluators.size(); i++) {
            EvoRobotExperiment* newExp = savedConfigurationParameters.getObjectFromGroup<EvoRobotExperiment>(savedExperimentPrefix, false);
            newExp->setEvoga(this);
            evaluators[i] = new EvaluatorThreadForEvoga(this, newExp);
        }
        QThreadPool::globalInstance()->setMaxThreadCount(numThreads);
    }

    for(rp=0;rp<nreplications;rp++) {   // replications
        mutation=0.5; // initially mutation rate is set to 50%
        //setSeed(getSeed());
        setSeed(getStartingSeed()+rp);
        Logger::info(QString("Replication %1, seed: %2").arg(rp+1).arg(getStartingSeed()+rp));
        resetGenerationCounter();
        randomizePop();
        // --- section runnable only if there is an Evonet object
        if ( hasResource( "evonet" ) ) {
            getPheParametersAndMutationsFromEvonet();
        }

        emit startingReplication( rp );
        QTime evotimer;
        evotimer.start();
        for (int i=0;i<popSize+1;i++) {
            tfitness[i]=0.0;
            ntfitness[i]=0.0;
        }
        //code to recovery a previous evolution: Experimental
        sprintf(statfile,"statS%d.fit", getStartingSeed()+rp);
        //now check if the file exists
        DataChunk statTest(QString("stattest"),Qt::blue,2000,false);
        if (statTest.loadRawData(QString(statfile),0)) {
            startGeneration=statTest.getIndex();
            sprintf(genFile,"G%dS%d.gen",startGeneration,getStartingSeed()+rp);
            Logger::info("Recovering from startGeneration: " + QString::number(startGeneration));
            Logger::info(QString("Loading file: ") + genFile);
            loadallg(startGeneration,genFile);
            cgen=startGeneration;
            mutation=mutation-startGeneration*mutationdecay;
            if (mutation<final_mrate) mutation=final_mrate;
            emit recoveredInterruptedEvolution( QString(statfile) );
        } //end evolution recovery code

        for(gn=startGeneration;gn<nogenerations;gn++) { // generations
            evotimer.restart();
            Logger::info(" Generation " + QString::number(gn+1));
            // Here we do this to avoid too many complications: if we have to run no threads, we use the old
            // code, otherwise we go for the multithread code below
            if (numThreads <= 1) {
                exp->initGeneration(gn);
                if ( commitStep() ) { return; }
                // Not running with multiple threads, using the old code
                for(id=0;id<popSize;id++) { //individuals
                    fit=0.0;
                    exp->setNetParameters(getGenes(id)); // get the free parameters from the genotype
                    exp->doAllTrialsForIndividual(id);
                    fit = exp->getFitness();
                    if (averageIndividualFitnessOverGenerations) {
                        tfitness[id] += fit;
                        ntfitness[id]++;
                    } else {
                        tfitness[id] = fit;
                        ntfitness[id] = 1;
                    }
                    if (isStopped()) { // stop evolution
                        return;
                    }
                    copyGenes(id, popSize,1); //generate a variation by duplicating and mutating
                    tfitness[popSize]=0;
                    ntfitness[popSize]=0;

                    exp->setNetParameters(getGenes(popSize)); // get the free parameters from the genotype
                    exp->doAllTrialsForIndividual(popSize + id);
                    fit = exp->getFitness();
                    if (averageIndividualFitnessOverGenerations) {
                        tfitness[popSize] += fit;
                        ntfitness[popSize]++;
                    } else {
                        tfitness[popSize] = fit;
                        ntfitness[popSize] = 1;
                    }
                    if (isStopped()) { // stop evolution
                        return;
                    }

                    // find the worst individual
                    minfit=999;
                    minid=-1;
                    for(int wi=0;wi<popSize;wi++) {
                        if (ntfitness[wi] == 0.0) {
                            continue;
                        }
                        mfit=tfitness[wi]/ntfitness[wi];
                        if(mfit<minfit) {
                            minfit=mfit;
                            minid=wi;
                        }
                    }
                    // eventually swap
                    if ((tfitness[popSize]/ntfitness[popSize])>minfit) {
                        copyGenes(popSize, minid,0);
                        tfitness[minid]=tfitness[popSize];
                        ntfitness[minid]=ntfitness[popSize];
                    }
                }
                exp->endGeneration(gn);
                if ( commitStep() ) { return; }
            } else {
                // Multithread code

                // Calling initGeneration on all evaluator
                for (int i = 0; i < popSize; i++) {
                    evaluators[i]->getExperiment()->initGeneration(gn);
                }
                if (commitStep()) return; // stop the evolution process

                // We first evaluate all parents, so setting genotypes of parents (we have as many evaluators as individuals)
                for (int i = 0; i < popSize; i++) {
                    evaluators[i]->setGenotype(i);
                }
                if (commitStep()) return; // stop the evolution process

                // Now starting parallel evaluation of parents and wating for it to finish
                QFuture<void> evaluationFuture = QtConcurrent::map(evaluators, runEvaluatorThreadForEvoga);
                evaluationFuture.waitForFinished();
                if (commitStep()) return; // stop the evolution process

                // We have finished evaluating parents, updating the fitness vectors
                for (int i = 0; i < popSize; i++) {
                    if (averageIndividualFitnessOverGenerations) {
                        tfitness[evaluators[i]->getGenotypeId()] += evaluators[i]->getFitness();
                        ntfitness[evaluators[i]->getGenotypeId()]++;
                    } else {
                        tfitness[evaluators[i]->getGenotypeId()] = evaluators[i]->getFitness();
                        ntfitness[evaluators[i]->getGenotypeId()] = 1;
                    }
                }
                if (commitStep()) return; // stop the evolution process

                // Now we can generate all children for all individuals and set them in the evaluators
                for (int i = 0; i < popSize; i++) {
                    copyGenes(i, popSize + i, 1); //generate a variation by duplicating and mutating
                    tfitness[popSize + i] = 0;
                    ntfitness[popSize + i] = 0;
                    evaluators[i]->setGenotype(popSize + i);
                }
                if (commitStep()) return; // stop the evolution process

                // Now starting parallel evaluation of children and wating for it to finish
                evaluationFuture = QtConcurrent::map(evaluators, runEvaluatorThreadForEvoga);
                evaluationFuture.waitForFinished();
                if (commitStep()) return; // stop the evolution process

                // We have finished evaluating parents, updating the fitness vectors
                for (int i = 0; i < popSize; i++) {
                    if (averageIndividualFitnessOverGenerations) {
                        tfitness[evaluators[i]->getGenotypeId()] += evaluators[i]->getFitness();
                        ntfitness[evaluators[i]->getGenotypeId()]++;
                    } else {
                        tfitness[evaluators[i]->getGenotypeId()] = evaluators[i]->getFitness();
                        ntfitness[evaluators[i]->getGenotypeId()] = 1;
                    }
                }
                if (commitStep()) return; // stop the evolution process

                // Calling endGeneration on all evaluator
                for (int i = 0; i < popSize; i++) {
                    evaluators[i]->getExperiment()->endGeneration(gn);
                }
                if (commitStep()) return; // stop the evolution process

                // Finally, we look for the worst individuals (parents) and substitute them with best children.
                // What we do is: we order both the parents and the children in descending order, then we take the
                // popSize best individuals. This is not the same as what is done in the sequential version of
                // the algorithm, but should be similar. We overwrite the worst parents with the best children
                QVector<FitnessAndId> parents(popSize);
                for (int i = 0; i < popSize; i++) {
                    parents[i].fitness = tfitness[i] / ntfitness[i];
                    parents[i].id = i;
                }
                QVector<FitnessAndId> children(popSize);
                for (int i = 0; i < popSize; i++) {
                    children[i].fitness = tfitness[popSize + i] / ntfitness[popSize + i];
                    children[i].id = popSize + i;
                }
                // Sorting both parents and children. They are sorted in ascending order but we need the best
                // individuals (those with the highest fitness) first
                qSort(parents);
                qSort(children);
                int p = popSize - 1;
                int c = popSize - 1;
                for (int i = 0; i < popSize; i++) {
                    if (parents[p].fitness > children[c].fitness) {
                        // No need to swap, parents are already in the population vector
                        p--;
                    } else {
                        // Swapping with one of the worst parents (we know for sure that p + c = popSize)
                        copyGenes(children[c].id, parents[popSize - 1 - c].id, 0);
                        tfitness[parents[popSize - 1 - c].id] = tfitness[children[c].id];
                        ntfitness[parents[popSize - 1 - c].id] = ntfitness[children[c].id];
                        c--;
                    }
                }
            }

            saveBestInd();
            computeFStat2();
            saveFStat();
            emit endGeneration( cgen, fmax, faverage, fmin );
            if (commitStep()) {
                return; // stop the evolution process
            }

            cgen++;
            if (mutation > final_mrate) {
                mutation -= mutationdecay;
            } else {
                mutation = final_mrate;
            }

            if ((savePopulationEachNGenerations != 0) && (gn % savePopulationEachNGenerations == 0)) {
                saveallg();
            }

            Logger::info(QString("Generation %1 took %2 minutes").arg(gn+1).arg((double)evotimer.elapsed()/60000.0, 0, 'f', 2));
            fflush(stdout);
        }
        saveallg();
    }

    // Deleting all evaluators
    for (int i = 0; i < evaluators.size(); i++) {
        delete evaluators[i];
    }
}

/*
 * Main function of the Genetic Algorithm (generational version with truncation selection)
 */
void Evoga::evolveGenerational()
{
    int rp; //replication
    int gn; //generation
    int id; //individuals
    double fit;
    int startGeneration=0;
    char statfile[128];
    char genFile[128];

    // Resizing genome
    genome.resize(popSize);

    for(rp=0;rp<nreplications;rp++) {// replications
        setSeed(getStartingSeed()+rp);
        Logger::info(QString("Replication %1 seed: %2").arg(rp+1).arg(getStartingSeed()));
        resetGenerationCounter();
        randomizePop();
        // --- section runnable only if there is an Evonet object
        if ( hasResource( "evonet" ) ) {
            getPheParametersAndMutationsFromEvonet();
        }

        emit startingReplication( rp );
        QTime evotimer;
        evotimer.start();

        //code to recovery a previous evolution: Experimental
        sprintf(statfile,"statS%d.fit", getStartingSeed()+rp);
        //now check if the file exists
        DataChunk statTest(QString("stattest"),Qt::blue,2000,false);
        if (statTest.loadRawData(QString(statfile),0)) {
            startGeneration=statTest.getIndex();
            sprintf(genFile,"G%dS%d.gen",startGeneration,getStartingSeed()+rp);
            Logger::info("Recovering from startGeneration: " + QString::number(startGeneration));
            Logger::info(QString("Loading file: ") + genFile);
            loadallg(startGeneration,genFile);
            emit recoveredInterruptedEvolution( QString(statfile) );
        } //end evolution recovery code

        for(gn=startGeneration;gn<nogenerations;gn++) { // generations
            evotimer.restart();
            Logger::info(" Generation " + QString::number(gn+1));
            exp->initGeneration( gn );
            for(id=0;id<popSize;id++) { //individuals
                exp->setNetParameters(getGenes(id)); // get the free parameters from the genotype
                exp->doAllTrialsForIndividual(id);
                fit = exp->getFitness();
                tfitness[id]=fit;
                if (commitStep()) { // stop evolution
                    return;
                }
            }
            reproduce();

            emit endGeneration( gn, fmax, faverage, fmin );
            exp->endGeneration( gn );

            if(savePopulationEachNGenerations!=0 && gn%savePopulationEachNGenerations == 0)
                saveallg();

            Logger::info(QString("Generation %1 took %2 minutes").arg(gn+1).arg((double)evotimer.elapsed()/60000.0, 0, 'f', 2));
            fflush(stdout);
        }

        saveallg();
    }
}


// this function generate a random seed for initialize the random number generator
unsigned int generateRandomSeed() {
    // this number is always different amongs processes because this function
    // is on the stack of the process
    unsigned long int stackMem = (unsigned long int)( generateRandomSeed );
    // time on which the process has been started
#ifdef FARSA_WIN
    unsigned long int startTime = GetTickCount();
#else
    unsigned long int startTime = time(NULL);
#endif
    // the seed is generated mixing the values above
    unsigned long int randSeed = 0;
    
    stackMem=stackMem-startTime;  stackMem=stackMem-randSeed;  stackMem=stackMem^(randSeed >> 13);
    startTime=startTime-randSeed;  startTime=startTime-stackMem;  startTime=startTime^(stackMem << 8); 
    randSeed=randSeed-stackMem;  randSeed=randSeed-startTime;  randSeed=randSeed^(startTime >> 13);
    stackMem=stackMem-startTime;  stackMem=stackMem-randSeed;  stackMem=stackMem^(randSeed >> 12);
    startTime=startTime-randSeed;  startTime=startTime-stackMem;  startTime=startTime^(stackMem << 16);
    randSeed=randSeed-stackMem;  randSeed=randSeed-startTime;  randSeed=randSeed^(startTime >> 5);
    stackMem=stackMem-startTime;  stackMem=stackMem-randSeed;  stackMem=stackMem^(randSeed >> 3);
    startTime=startTime-randSeed;  startTime=startTime-stackMem;  startTime=startTime^(stackMem << 10);
    randSeed=randSeed-stackMem;  randSeed=randSeed-startTime;  randSeed=randSeed^(startTime >> 15);
    
    return randSeed%10000;
}

void Evoga::configure(ConfigurationParameters& params, QString prefix)
{
    genome.clear();
    bestgenome.clear();

    evolutionType = ConfigurationHelper::getString(params, prefix + "evolutionType", "steadyState");
    if ( evolutionType != "steadyState" && evolutionType != "generational" ) {
        Logger::error( "Evoga - evolutionType has been wrongly setted. It can assume only 'steadyState' or 'generational' values" );
        evolutionType = "steadyState";
    }
    nogenerations = ConfigurationHelper::getInt(params, prefix + "ngenerations", 100); //number of generations
    nreplications = ConfigurationHelper::getInt(params, prefix + "nreplications", 10); //number of replications
    nreproducing = ConfigurationHelper::getInt(params, prefix + "nreproducing", 20); //number of father
    noffspring = ConfigurationHelper::getInt(params, prefix + "noffspring", 5); //number of sons
    // starting seed - if not specified (or setted as 'auto') it auto generate a seed
    QString checkSeed = ConfigurationHelper::getString(params, prefix + "seed", "auto");
    if ( checkSeed == "auto" ) {
        seed = generateRandomSeed();
    } else {
        seed = checkSeed.toInt();
    }
    Logger::info( QString("Evoga - Random seed setted to ")+QString::number(seed) );
    //seed = ConfigurationHelper::getInt(params, prefix + "seed", 1234); //starting seed
    savebest = ConfigurationHelper::getInt(params, prefix + "savenbest", 1); //saving best n genotype
    elitism = ConfigurationHelper::getBool(params, prefix + "elitism", false); //activate elitism
    numThreads = ConfigurationHelper::getInt(params, prefix + "numThreads", 1); //number of concurrent threads to use
    savePopulationEachNGenerations = ConfigurationHelper::getInt(params, prefix + "savePopulationEachNGenerations", 0);
    averageIndividualFitnessOverGenerations = ConfigurationHelper::getBool(params, prefix + "averageIndividualFitnessOverGenerations", true);

    //mutation rate can be written both as int or as double
    mutation = ConfigurationHelper::getDouble(params, prefix + "mutation_rate", mutation); //mutation rate
    if(mutation >= 1) {
        mutation /= 100;
    }
    mutationdecay = ConfigurationHelper::getDouble(params, prefix + "mutation_decay", 0.01);    //speed of decay (only valid for steady-state GA)

    exp = params.getObjectFromGroup<EvoRobotExperiment>(prefix + "Experiment", false);
    exp->setEvoga(this);
    Logger::info( "Created EvoRobotExperiment " + params.getValue(prefix+"Experiment/type") + " from group " + prefix + "Experiment" );

    // Copying the ConfigurationParameters object and the prefix: we will need them to create new instances of the experiment for
    // multithreading evolution
    savedConfigurationParameters = params;
    savedExperimentPrefix = prefix + "Experiment";
}

void Evoga::save(ConfigurationParameters&, QString)
{
    Logger::error("NOT IMPLEMENTED (Evoga::save)");
    abort();
}

void Evoga::describe( QString type ) {
    Descriptor d = addTypeDescription( type, "Implements the genetic algorithm developed by Stefano Nolfi" );
    d.describeEnum( "evolutionType" ).def("steadyState").values( QStringList() << "steadyState" << "generational" ).props( IsMandatory ).help("Specify the type of evolution process to execute");
    d.describeInt( "ngenerations" ).def(100).limits(0,MaxInteger).help("Number of generations");
    d.describeInt( "nreplications" ).def(10).limits(1,MaxInteger).help("The number of which the evolution process will be replicated with a different random initial population");
    d.describeInt( "nreproducing" ).def(20).limits(1,MaxInteger).help("The number of individual allowed to produce offsprings; The size of populazion will be nreproducing x noffspring");
    d.describeInt( "noffspring" ).def(5).limits(1,MaxInteger).help("The number of offsprings generated by an individual; The size of populazion will be nreproducing x noffspring");
    d.describeInt( "seed" ).def(1234).limits(0,MaxInteger).help("The number used to initialize the random number generator; when a new replication will start, this value will be incremented by one to guarantee a truly different initial population for the next replica");
    d.describeInt("savenbest").def(1).limits(1,MaxInteger).help("The number of best genotypes to save each generation");
    d.describeBool("elitism").def(false).help("If use elitism or not");
    d.describeInt("numThreads").def(1).limits(1,MaxInteger).help("The number of thread used to parallelize the evaluation of individuals");
    d.describeInt("savePopulationEachNGenerations").def(0).limits(0,MaxInteger).help("If is zero only the population of the last generation are saved into a file; otherwise it saves the population each N generations done");
    d.describeReal("mutation_rate").def(0.05).limits(0,100).help("The rate at a mutation will occur during a genotype copy; a real value below 1 (i.e. 0.12) is considered as a rate (i.e. 0.12 correspond to 12% of mutation); a value egual or above 1 is considered as a percentage of mutation (i.e. 25 correspond to 25% of mutation, or 0.25 rate of mutation)");
    d.describeReal("mutation_decay").def(0.01).limits(0,1).help("At first generation the mutation rate will be always 0.5, and at each generation done the mutation rate will be decreased by this value until it reachs the mutation_rate value");
    d.describeSubgroup( "Experiment" ).props(IsMandatory).type("EvoRobotExperiment").help("The object delegated to simulate and to evaluate the fitness of an individual");
    d.describeBool("averageIndividualFitnessOverGenerations").def(true).help("Whether to average the current fitness with the previous one or not");
}

void Evoga::postConfigureInitialization()
{
    // Sharing resources and declaring which ones we will access
    shareResourcesWith(exp);
    usableResources(QStringList() << "evonet");

    // Allocating memory
    tfitness = new double[MAXINDIVIDUALS];
    statfit = new double*[nogenerations];
    ntfitness = new double[MAXINDIVIDUALS]; //used with super ag implementation
    for(int i = 0; i < nogenerations; i++) {
        statfit[i] = new double[3];
    }

    //allocating memory for the mutation vector
    glen = exp->getGenomeLength();
    mutations = new float[glen];
    for(int mi = 0; mi < glen; mi++) {
        mutations[mi] = Evonet::DEFAULT_VALUE;
    }

    //now we allocate memory for genome and best genome
    cgen = 0;
    popSize = noffspring * nreproducing;

    //dynamic allocation of genome
    genome.setGenomeLength(glen);
    genome.resize(popSize); // An initial set of individuals just to avoid problems...
    for (int i = 0; i < genome.size(); i++) {
        for (int r = 0; r < glen; r++) {
            genome[i][r] = i;//mrand(256);test
        }
    }

    //dynamic allocation of bestgenome
    bestgenome.setGenomeLength(glen);
    bestgenome.resize(nreproducing);

    //resetting fitness
    for (int i = 0; i < MAXINDIVIDUALS; i++) {
        tfitness[i] = 0.0;//
    }

    for (int i = 0; i < nogenerations; i++) {
        statfit[i][0] = 0.0; //Average fitness of population
        statfit[i][1] = 0.0; //max fitness
        statfit[i][2] = 0.0; //min fitness
    }

    Logger::info("Evoga Configured - Number of genes: " + QString::number(glen));
}

void Evoga::evolveAllReplicas()
{
    stopEvolution = false;
    if ( evolutionType == "steadyState" ) {
        evolveSteadyState();
    } else if ( evolutionType == "generational" ) {
        evolveGenerational();
    } else {
        Logger::error( QString("Evoga - request to execute a unrecognized evolution type: %1").arg(evolutionType) );
    }
}

void Evoga::stop() {
    stopEvolution = true;
    waitForNextStep.wakeAll();
}

bool Evoga::commitStep() {
    if ( isStepByStep && !stopEvolution ) {
        // will block waiting the command for going ahead
        mutexStepByStep.lock();
        waitForNextStep.wait( &mutexStepByStep );
        mutexStepByStep.unlock();
    }
    return stopEvolution;
}

bool Evoga::isStopped() {
    return stopEvolution;
}

void Evoga::resetStop() {
    stopEvolution = false;
}

void Evoga::enableStepByStep( bool enable ) {
    isStepByStep = enable;
    // if disable the step-by-step it wake any eventually blocked commitStep
    if ( !enable ) {
        waitForNextStep.wakeAll();
    }
}

bool Evoga::isEnabledStepByStep() {
    return isStepByStep;
}

void Evoga::doNextStep() {
    waitForNextStep.wakeAll();
}

EvoRobotExperiment* Evoga::getEvoRobotExperiment()
{
    return exp;
}

QVector<EvoRobotExperiment*> Evoga::getEvoRobotExperimentPool()
{
    QVector<EvoRobotExperiment*> v;
    v.append(exp);
    return v;
}

unsigned int Evoga::currentGeneration()
{
    return cgen;
}

unsigned int Evoga::getStartingSeed()
{
    return seed;
}

unsigned int Evoga::getCurrentSeed()
{
    return currentSeed;
}

unsigned int Evoga::getNumReplications()
{
    return nreplications;
}

unsigned int Evoga::getNumOfGenerations() {
    return nogenerations;
}

double Evoga::getCurrentMutationRate() {
    return mutation;
}

void Evoga::setCurrentMutationRate( double mutation_rate ) {
    mutation = mutation_rate;
}

unsigned int Evoga::loadGenotypes(QString filename)
{
    return loadallg(-1, filename.toAscii().data());
}

unsigned int Evoga::numLoadedGenotypes() const
{
    return loadedIndividuals;
}

int* Evoga::getGenesForIndividual(unsigned int id)
{
    return getGenes(id);
}

QString Evoga::statisticsFilename(unsigned int seed)
{
    return "statS" + QString::number(seed) + QString(".fit");
}

QString Evoga::bestsFilename(unsigned int seed)
{
    return "B0S" + QString::number(seed) + QString(".gen");
}

QString Evoga::bestsFilename()
{
    return "B0S*.gen";
}

QString Evoga::generationFilename(unsigned int generation, unsigned int seed)
{
    return "G" + QString::number(generation) + "S" + QString::number(seed) + QString(".gen");
}

QString Evoga::generationFilename()
{
    return "G*S*.gen";
}

void Evoga::doNotUseMultipleThreads()
{
    numThreads = 1;
}

} // end namespace farsa

// All the suff below is to restore the warning state on Windows
#if defined(_MSC_VER)
    #pragma warning(pop)
#endif