void cma_es(Swarm &sw){
if(logCMAES){
printf("\n--------------------------------------------------------------\n");
printf("Repository (%d)\n", sw.repository.getActualSize());
for(int i=0;i<sw.repository.getActualSize();i++){
printVector(sw.repository.getSolution(i).decisionVector, decisionNumber);
printf("\n");
}
}
double sigmaPrev=sw.sigma;
myLearn(sw);
// if(fabs(sigmaPrev-sw.sigma) > 10000){
// fprintf(stderr, "\nWARNING!, sigma changed too much: %f -> %f. resetting...\n",sigmaPrev, sw.sigma); //shows warning message
// sw.init=true; //set the init flag, so all the CMA-ES variables are reset
// myLearn(sw); //learn again using the default values as base
// }
if(sw.sigma != sw.sigma || sw.sigma >= MAXDOUBLE){ //check for invalid numbers NaN or inf
fprintf(stderr, "WARNING!, sigma: %f. resetting...\n", sw.sigma); //shows warning message
// exit(1);
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
myLearn(sw); //learn again using the default values as base
}
//four reset criteria as in the paper "injecting cma-es into moea/d"
//NoEffectCoord
for (int iKoo = 0; iKoo < decisionNumber; ++iKoo){
if (sw.mean[iKoo] == sw.mean[iKoo] + 0.2*sw.sigma*sqrt(sw.C[iKoo][iKoo])){
fprintf(stderr, "NoEffectCoordinate: standard deviation 0.2*%7.2e in coordinate %d without effect\n", sw.sigma*sqrt(sw.C[iKoo][iKoo]), iKoo); //shows warning message
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
myLearn(sw); //learn again using the default values as base
break;
}
}
//NoEffectAxis
int iKoo;
for (int iAchse = 0; iAchse < decisionNumber; ++iAchse){
double fac = 0.1 * sw.sigma * sw.D[iAchse];
for (iKoo = 0; iKoo < decisionNumber; ++iKoo){
if (sw.mean[iKoo] != sw.mean[iKoo] + fac * sw.B[iKoo][iAchse])
break;
}
if (iKoo == decisionNumber){
/* t->sigma *= exp(0.2+t->sp.cs/t->sp.damps); */
fprintf(stderr, "NoEffectAxis: standard deviation 0.1*%7.2e in principal axis %d without effect\n", fac/0.1, iAchse);
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
myLearn(sw); //learn again using the default values as base
break;
}
}
//TolXUp
double stopTolUpXFactor=1e3;
double initialStds=0.3;
int i;
for(i=0; i<decisionNumber; ++i) {
if (sw.sigma * sqrt(sw.C[i][i]) > stopTolUpXFactor * initialStds)
break;
}
if (i < decisionNumber) {
fprintf(stderr, "TolUpX: standard deviation increased by more than %7.2e, larger initial standard deviation recommended \n", stopTolUpXFactor);
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
myLearn(sw); //learn again using the default values as base
}
//ConditionCov
double dMaxSignifKond=1e13;
if (rgdouMax(sw.D, decisionNumber) >= rgdouMin(sw.D, decisionNumber) * dMaxSignifKond) {
fprintf(stderr, "ConditionNumber: maximal condition number %7.2e reached. maxEW=%7.2e,minEW=%7.2e\n", dMaxSignifKond, rgdouMax(sw.D, decisionNumber), rgdouMin(sw.D, decisionNumber));
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
myLearn(sw); //learn again using the default values as base
}
//************************************* RESAMPLE SOLUTIONS FROM THE GOOD FRONTS ******************************//
//re-learn and re-sample when errors in the covariance matrix are detected in the sampling phase
bool success=mySample(sw);
while(!success){
myLearn(sw);
success=mySample(sw);
fprintf(stderr,"Resample\n");
}
// Neighbor orderedSwarms[swarmNumber];
// for(int i=0;i<swarmNumber;i++){
// orderedSwarms[i].index=i;
// orderedSwarms[i].distance=swarms[i].modelQuality*-1;//inverted model quality because we will use the preexisting sort, that sorts based on the smallest value (distance)
// }
// std::sort(orderedSwarms, orderedSwarms+swarmNumber, neighborsComparator);
//
// // for(int i=0;i<swarmNumber;i++){
// // printf("swarm: %d has quality %f sz: %d\n", orderedSwarms[i].index, orderedSwarms[i].distance*-1, repGlobal->getActualSize());
// // }
// // printf("\n\n\n");
//
// bool good=false;
// for(int i=0;i<100;i++){
// if(sw.neighborhood[0].index == orderedSwarms[i].index){
// good=true;
// break;
// }
// }
//
// if(good){
// //re-learn and re-sample when errors in the covariance matrix are detected in the sampling phase
// bool success=mySampleReplacement(sw, 10);
// while(!success){
// myLearn(sw);
// success=mySampleReplacement(sw, 10);
// fprintf(stderr,"Resample\n");
// }
// }else{
// for(int i=0;i<sw.getSize();i++)
// sw.particles[i].solution.evalSolution=false;
// // //re-learn and re-sample when errors in the covariance matrix are detected in the sampling phase
// // bool success=mySample(sw);
// // while(!success){
// // myLearn(sw);
// // success=mySample(sw);
// // fprintf(stderr,"Resample\n");
// // }
// }
//************************************* END OF RESAMPLING SOLUTIONS FROM THE GOOD FRONTS ******************************//
if(logCMAES){
printf("\n\npop after \n\n");
for(int i=0; i<sw.getSize();i++){
//printVector(evo.rgrgx[i],N+2);
printVector(sw.particles[i].solution.decisionVector,decisionNumber);
printf("\n");
}
}
}
bool mySample (Swarm &sw){
int N=decisionNumber;
/* calculate eigensystem */
if(logCMAES){
printf("\n----------------------before eigen calculation ---------------------- ");
// printf("\nCov mat: \n");
// for (int i = 0; i < N; ++i){
// printVector(sw.C[i], N);
// printf("\n");
// }
printf("\nB mat: \n");
for (int i = 0; i < N; ++i){
printVector(sw.B[i], N);
printf("\n");
}
printf("\nD mat: \n");
printVector(sw.D, N);
printf("\n");
printf("---------------------end before ---------------------- \n");
printf("\nCov mat: \n");
for (int i = 0; i < N; ++i){
printVector(sw.C[i], N);
printf("\n");
}
}
double rgdTmp[N];
Eigen( N, sw.C, sw.D, sw.B, rgdTmp);
//By using Cholesky decomposition, it turns out that the determinant of a matrix is equal to the product of its eigenvalues
sw.det=sw.D[0];
for(int i=1;i<N;i++)
sw.det*=sw.D[i];
double lg=log(sw.det);
if(lg >= MAXDOUBLE || lg != lg){
fprintf(stderr, "WARNING!, log of the covariance matrix determinant: %f. ", lg); //shows warning message
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
return false;
}
for (int i = 0; i < N; ++i)
sw.D[i] = sqrt(sw.D[i]);
for(int i=0;i<N;i++){
if(sw.D[i] != sw.D[i] || sw.D[i] >= MAXDOUBLE || sw.D[i] < 0){//tests for 1)null 2)inf 3) positive definiteness of the covariance matrix
// if(sw.det != sw.det || sw.det <= MAXDOUBLE*-1 || sw.D[i] != sw.D[i] || sw.D[i] >= MAXDOUBLE || sw.D[i] < 0){//tests for 1) determinant null or invalid 2)null 3)inf 4) positive definiteness of the covariance matrix
fprintf(stderr, "WARNING!, value: %f in eigenValues vector. ", sw.D[i]); //shows warning message
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
return false;
// exit(1);
//myLearn(sw, neighboringSwarms); //learn again using the default values as base
}
}
if(logCMAES){
printf("---------------------after eigen calculation ---------------------- \n");
// printf("\nCov mat: \n");
// for (int i = 0; i < N; ++i){
// printVector(sw.C[i], N);
// printf("\n");
// }
printf("B mat: \n");
for (int i = 0; i < N; ++i){
printVector(sw.B[i], N);
printf("\n");
}
printf("\nD mat: \n");
printVector(sw.D, N);
printf("\n");
printf("---------------------end after ---------------------- \n");
printf("\nsigma: %f (%e)\n",sw.sigma, sw.sigma);
for (int i = 0; i < N; ++i)
rgdTmp[i] = sw.D[i] * 3;
printf("\nD (sample) (* 3) : ");
for (int i = 0; i < N; ++i){
double sum = 0.0;
for (int j = 0; j < N; ++j)
sum += sw.B[i][j] * rgdTmp[j];
if((sw.mean[i] + sw.sigma * sum ) >= 0 )
printf(" ");
printf("%.3f ", /*fabs*/(sw.mean[i] + sw.sigma * sum ) );
}
for (int i = 0; i < N; ++i)
rgdTmp[i] = sw.D[i] * 0;
printf("\nD (sample) (* 0) : ");
for (int i = 0; i < N; ++i){
double sum = 0.0;
for (int j = 0; j < N; ++j)
sum += sw.B[i][j] * rgdTmp[j];
if((sw.mean[i] + sw.sigma * sum ) >= 0 )
printf(" ");
printf("%.3f ", /*fabs*/(sw.mean[i] + sw.sigma * sum ) );
}
for (int i = 0; i < N; ++i)
rgdTmp[i] = sw.D[i] * -3;
printf("\nD (sample) (* -3): ");
for (int i = 0; i < N; ++i){
double sum = 0.0;
for (int j = 0; j < N; ++j)
sum += sw.B[i][j] * rgdTmp[j];
if((sw.mean[i] + sw.sigma * sum ) >= 0 )
printf(" ");
printf("%.3f ", /*fabs*/(sw.mean[i] + sw.sigma * sum ) );
}
}
// // printVector(sw.mean, decisionNumber); //show the average decision vector per iteration
// // printf("\n");
// printf("\nD (sample) (difsum): ");
// for (int i = 0; i < N; ++i){
// for (j = 0, sum = 0.0; j < N; ++j)
// sum += sw.B[i][j] * rgdTmp[j];
// printf("%f ", fabs(sw.mean[i] + sw.sigma * sw.D[i]*-3 )+fabs(sw.mean[i] + sw.sigma * sqrt(sw.D[i])*3 ) );
// }
/*************************/
double previousSol[N];
if(sw.getSize() == 1)
memcpy(previousSol, sw.particles[0].solution.decisionVector, sizeof(double)*N);
// original sampling
for (int iNk = 0; iNk < sw.getSize(); ++iNk){ /* generate scaled cmaes_random vector (D * z) */
// int rescount=0;
// while(true){//resample when invalid
// bool resample=false;
for (int i = 0; i < N; ++i)
rgdTmp[i] = sw.D[i] * cmaes_random_Gauss();
/* add mutation (sigma * B * (D*z)) */
for (int i = 0; i < N; ++i) {
double sum = 0.0;
for (int j = 0; j < N; ++j)
sum += sw.B[i][j] * rgdTmp[j];
double value=sw.mean[i] + sw.sigma * sum;//normalized new solution
// if(value < 0 || value > 1){
// resample=true;
// // printf("\nresample: %f (%d, %d)\n", value, i, rescount);
// }else
sw.particles[iNk].solution.evalSolution=true;
sw.particles[iNk].solution.decisionVector[i] = (value*(superiorPositionLimit[i]-inferiorPositionLimit[i]))+inferiorPositionLimit[i]; //unormalize solution
}
// if(!resample || rescount > 100)
// break;
// else
// rescount++;
// }
// if(rescount >= 100){
// printf("\n GAVE UP \n");
// exit(1);
// }
}
// //ranking for the new surrogate functions
// if(repGlobal->getActualSize() >0 ){
// repGlobal->organize();
// for(int i=0;i<repGlobal->getActualSize();i++)
// repGlobal->getSolutions()[i].crowdingDistance=logLikelihood(repGlobal->getSolution(i).decisionVector, sw);//stores temporarily in the crowding distance field
//
// std::sort(repGlobal->getSolutions(), repGlobal->getSolutions()+repGlobal->getActualSize(), crowdingComparatorSol);
//
//
// double likelihood=logLikelihood(sw.particles[0].solution.decisionVector, sw);
//
// double rank = likelihoodRanking(sw.particles[0].solution, *repGlobal, sw);
//
// if(rank >0 && rank < repGlobal->getActualSize())
// printf("rank: %.1f lik before: %f lik: %f lik after: %f\n", rank, repGlobal->getSolution((int)rank).crowdingDistance,likelihood,repGlobal->getSolution((int)rank+1).crowdingDistance);
// else
// printf("rank: %f\n", rank);
// }
// // // for (int iNk = 0; iNk < sw.getSize(); ++iNk){//using the likelihood as surrogate
// // // // double best[N], bestValue=MAXDOUBLE*-1;//using best as highest likelihood to get convergence
// // // double best[N], bestValue=MAXDOUBLE;//using the best as lowest likelihood to get diversity
// // // for(int t=0;t<10;t++){//ten trials per particle and choose the best
// // // double candidate[N];
// // //
// // // for (int i = 0; i < N; ++i)
// // // rgdTmp[i] = sw.D[i] * cmaes_random_Gauss();
// // // /* add mutation (sigma * B * (D*z)) */
// // // for (int i = 0; i < N; ++i) {
// // // double sum = 0.0;
// // // for (int j = 0; j < N; ++j)
// // // sum += sw.B[i][j] * rgdTmp[j];
// // //
// // // candidate[i]=sw.mean[i] + sw.sigma * sum;//normalized new solution
// // // }
// // // double opinion=neighborhoodOpinion(sw, candidate);
// // // // printf("opinion: %f of: ",opinion);
// // // // printVector(candidate, N);
// // // // printf("\n");
// // // // if(opinion > bestValue){//using best as highest likelihood to get convergence
// // // if(opinion < bestValue){//using the best as lowest likelihood to get diversity
// // // for (int i = 0; i < N; ++i)
// // // best[i]=candidate[i];
// // // bestValue=opinion;
// // // }
// // // }
// // //
// // // for (int i = 0; i < N; ++i)
// // // sw.particles[iNk].solution.decisionVector[i] = (best[i]*(superiorPositionLimit[i]-inferiorPositionLimit[i]))+inferiorPositionLimit[i]; //unormalize solution
// // // }
/* Test if function values are identical, escape flat fitness */ //original comment
//Check if the sampled solutions are equal, or if only one solution is sampled, if it is equal to the previous solution
// if (t->rgFuncValue[t->index[0]] == t->rgFuncValue[t->index[(int)t->sp.lambda/2]]) {
// t->sigma *= exp(0.2+t->sp.cs/t->sp.damps);
// ERRORMESSAGE("Warning: sigma increased due to equal function values\n",
// " Reconsider the formulation of the objective function",0,0);
// }
if(sw.getSize() == 1){
if(isEqual(previousSol, sw.particles[0].solution.decisionVector, N)){
sw.sigma *= exp(0.2+cSigma/dSigma);
fprintf(stderr, "Warning: sigma increased due to equal decision vectors.\n");
return false;
}
// else
// for(int i=0;i<N;i++)
// printf("%.14f ", abs(previousSol[i]-sw.particles[0].solution.decisionVector[i]));
// printf("\n");
}else{
if(isEqual(sw.particles[0].solution.decisionVector, sw.particles[sw.getSize()-1].solution.decisionVector, N)){
sw.sigma *= exp(0.2+cSigma/dSigma);
fprintf(stderr, "Warning: sigma increased due to equal decision vectors.\n");
return false;
}
}
return true;
} /* SamplePopulation() */
bool mySampleReplacement (Swarm &sw, int trials){
int N=decisionNumber;
//-------------------------------------------------------------------------------------------
/* calculate eigensystem */
double rgdTmp[N];
Eigen( N, sw.C, sw.D, sw.B, rgdTmp);
//By using Cholesky decomposition, it turns out that the determinant of a matrix is equal to the product of its eigenvalues
sw.det=sw.D[0];
for(int i=1;i<N;i++)
sw.det*=sw.D[i];
double lg=log(sw.det);
if(lg >= MAXDOUBLE || lg != lg){
fprintf(stderr, "WARNING!, log of the covariance matrix determinant: %f. ", lg); //shows warning message
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
return false;
}
for (int i = 0; i < N; ++i)
sw.D[i] = sqrt(sw.D[i]);
for(int i=0;i<N;i++){
if(sw.D[i] != sw.D[i] || sw.D[i] >= MAXDOUBLE || sw.D[i] < 0){//tests for 1)null 2)inf 3) positive definiteness of the covariance matrix
// if(sw.det != sw.det || sw.det <= MAXDOUBLE*-1 || sw.D[i] != sw.D[i] || sw.D[i] >= MAXDOUBLE || sw.D[i] < 0){//tests for 1) determinant null or invalid 2)null 3)inf 4) positive definiteness of the covariance matrix
fprintf(stderr, "WARNING!, value: %f in eigenValues vector. ", sw.D[i]); //shows warning message
sw.init=true; //set the init flag, so all the CMA-ES variables are reset
return false;
// exit(1);
//myLearn(sw, neighboringSwarms); //learn again using the default values as base
}
}
//-------------------------------------------------------------------------------------------
double previousSol[N];
if(sw.getSize() == 1)
memcpy(previousSol, sw.particles[0].solution.decisionVector, sizeof(double)*N);
// //ranking for the new surrogate functions
// if(repGlobal->getActualSize() >0 ){
// repGlobal->organize();
// for(int i=0;i<repGlobal->getActualSize();i++)
// repGlobal->getSolutions()[i].crowdingDistance=logLikelihood(repGlobal->getSolution(i).decisionVector, sw);//stores temporarily in the crowding distance field
//
// std::sort(repGlobal->getSolutions(), repGlobal->getSolutions()+repGlobal->getActualSize(), crowdingComparatorSol);
//
//
// double likelihood=logLikelihood(sw.particles[0].solution.decisionVector, sw);
//
// double rank = likelihoodRanking(sw.particles[0].solution, *repGlobal, sw);
//
// if(rank >0 && rank < repGlobal->getActualSize())
// printf("rank: %.1f lik before: %f lik: %f lik after: %f\n", rank, repGlobal->getSolution((int)rank).crowdingDistance,likelihood,repGlobal->getSolution((int)rank+1).crowdingDistance);
// else
// printf("rank: %f\n", rank);
// }
for (int iNk = 0; iNk < sw.getSize(); ++iNk){//using the likelihood as surrogate
double best[N], bestValue=MAXDOUBLE*-1;//using best as highest likelihood to get convergence
// double best[N], bestValue=MAXDOUBLE;//using the best as lowest likelihood to get diversity
for(int t=0;t<trials;t++){//trials per particle and choose the best
double candidate[N];
for (int i = 0; i < N; ++i)
rgdTmp[i] = sw.D[i] * cmaes_random_Gauss();
/* add mutation (sigma * B * (D*z)) */
for (int i = 0; i < N; ++i) {
double sum = 0.0;
for (int j = 0; j < N; ++j)
sum += sw.B[i][j] * rgdTmp[j];
candidate[i]=sw.mean[i] + sw.sigma * sum;//normalized new solution
}
double opinion=neighborhoodOpinion(sw, candidate);
// printf("opinion: %f of: ",opinion);
// printVector(candidate, N);
// printf("\n");
if(opinion > bestValue){//using best as highest likelihood to get convergence
// if(opinion < bestValue){//using the best as lowest likelihood to get diversity
for (int i = 0; i < N; ++i)
best[i]=candidate[i];
bestValue=opinion;
}
}
sw.particles[iNk].solution.evalSolution=true;
for (int i = 0; i < N; ++i)
sw.particles[iNk].solution.decisionVector[i] = (best[i]*(superiorPositionLimit[i]-inferiorPositionLimit[i]))+inferiorPositionLimit[i]; //unormalize solution
}
if(sw.getSize() == 1){
if(isEqual(previousSol, sw.particles[0].solution.decisionVector, N)){
sw.sigma *= exp(0.2+cSigma/dSigma);
fprintf(stderr, "Warning: sigma increased due to equal decision vectors.\n");
return false;
}
// else
// for(int i=0;i<N;i++)
// printf("%.14f ", abs(previousSol[i]-sw.particles[0].solution.decisionVector[i]));
// printf("\n");
}else{
if(isEqual(sw.particles[0].solution.decisionVector, sw.particles[sw.getSize()-1].solution.decisionVector, N)){
sw.sigma *= exp(0.2+cSigma/dSigma);
fprintf(stderr, "Warning: sigma increased due to equal decision vectors.\n");
return false;
}
}
return true;
}
void myLearn(Swarm &sw){
int N=decisionNumber, hsig;
double oldMean[N], initialStds[N], BDz[N];
double muEff, cc, muCov, cCov, trace=0.0, chiN;
cSigma=0; dSigma=0;
//double sigma, C[N][N], pC[N], pSigma[N], D[N], mean[N], B[N][N];
//****************** treating the input data ********************//
Repository repTemp;
//gathering the solutions to be used to learn and putting them on repTemp
if(decomposition){
mergeNeighboringRepositories(repTemp, sw);
}else{
repTemp.initialize(archSubSwarms, sw.getSize()+sw.repository.getActualSize());
if(!strcmp(solSet, "population") || !strcmp(solSet, "both"))
for(int i=0;i<sw.getSize();i++)
if(!sw.particles[i].solution.dominated)
repTemp.add(sw.particles[i].solution);
if(!strcmp(solSet, "repository") || repTemp.getActualSize()==0 || !strcmp(solSet, "both"))
repTemp.add(sw.repository);
}
repTemp.organize();
// if(!strcmp(archSubSwarms, "p-ideal") || !strcmp(archSubSwarms, "w-ideal") || !strcmp(archSubSwarms, "r-ideal")){
if(decomposition){
updateWeightedDistances(repTemp.getSolutions(), repTemp.getActualSize(), sw.repository.weight);
std::sort(repTemp.getSolutions(), repTemp.getSolutions()+repTemp.getActualSize(), weightedDistanceComparatorSol);
}else{
if(!strcmp(cmaRanking, "cd")){
updateCrowdingDistances(repTemp.getSolutions(), repTemp.getActualSize());
std::sort(repTemp.getSolutions(), repTemp.getSolutions()+repTemp.getActualSize(), crowdingComparatorSol);
}
if(!strcmp(cmaRanking, "hv")){
updateContributingHypervolume(repTemp); //stores in the crowding distance field, since both are the higher the contribution, the better, there is no problem
std::sort(repTemp.getSolutions(), repTemp.getSolutions()+repTemp.getActualSize(), crowdingComparatorSol);
}
if(!strcmp(cmaRanking, "r2")){
if(refSize==-1){
perror("ERROR ON CMA-ES R2 RANKING! Reference file not set.");
}
updateContributingR2(repTemp); //stores in the crowding distance field, stores in the crowding distance field //if the value of the r2 increase after I take this solution, it is important, so the higher is the contribution, the better
std::sort(repTemp.getSolutions(), repTemp.getSolutions()+repTemp.getActualSize(), crowdingComparatorSol);
}
}
//normalizing non dominated solutions
double nonDominated[repTemp.getActualSize()][N];
int nonDom=0;
// for(int i=0;i<repTemp.getActualSize()/2;i++){
for(int i=0;i<repTemp.getActualSize();i++){
// printVector(repTemp.getSolution(i).decisionVector, decisionNumber);
// printVector(repTemp.getSolution(i).objectiveVector, objectiveNumber);
// printf("sol:%d: %e\n", i, repTemp.getSolution(i).crowdingDistance);
for(int j=0;j<N;j++)
nonDominated[nonDom][j]=normalize(repTemp.getSolution(i).decisionVector[j], inferiorPositionLimit[j], superiorPositionLimit[j]);
nonDom++; //mu
}
int mu = nonDom;
double weights[mu];
//**************************** end of treatment of input data***************//
/***************setting default parameters according to****************/
//https://www.lri.fr/~hansen/hansenedacomparing.pdf
// in my strategy lambda = mu = solSize
double weightSquare=0, weightSum=0, value=0;
double smaller=MAXDOUBLE, larger=-MAXDOUBLE;
if(mu > 2){
if(!strcmp(weightDistribution, "metric")){ //uses the value of the ranking metric as weight
for(int i=mu-1;i>=0;i--){// find the smaller > 0
if(decomposition){
value=repTemp.getSolution(i).weightedDistance;
if(value < MAXDOUBLE){
larger=value;
break;
}
}
else{
value=repTemp.getSolution(i).crowdingDistance;
if(value > 0){ //a little higher than 0
smaller=value;
break;
}
}
}
for(int i=0;i<mu;i++){ //find the larger < MAXDOUBLE
if(decomposition){
value=repTemp.getSolution(i).weightedDistance;
if(value > 0){ //a little higher than 0
smaller=value;
break;
}
}
else{
value=repTemp.getSolution(i).crowdingDistance;
if(value < MAXDOUBLE){
larger=value;
break;
}
}
}
if(larger == -MAXDOUBLE || smaller == MAXDOUBLE){
fprintf(stderr, "\nERROR ON ASSIGNING WEIGHTS FOR CMA-ES!!\n");
fprintf(stderr, "\nsm: %f lg: %f mu: %d\n", smaller, larger, mu);
exit(1);
}
}
}else{
larger=1;
smaller=0;
}
for(int i=0;i<mu;i++){
if(!strcmp(weightDistribution, "equal"))
weights[i]=1.0; //equal weight distribution
if(!strcmp(weightDistribution, "log"))
weights[i]=log(mu+1)-log(i+1.0); //according to code
if(!strcmp(weightDistribution, "linear"))
weights[i]=mu-i; //according to code
if(!strcmp(weightDistribution, "metric")){ //uses the value of the ranking metric as weight
if(decomposition){
value=repTemp.getSolution(i).weightedDistance;
weights[i]=1-(normalize(value, (smaller-(smaller/100.0)), larger));
}else{
value=repTemp.getSolution(i).crowdingDistance;
weights[i]=(normalize(value, (smaller-(smaller/100.0)), larger));
}
// weights[i]=exp(normalize(repTemp.getSolution(i).crowdingDistance, smaller, larger));
// if(repTemp.getSolution(i).crowdingDistance<=0)
// weights[i]=exp(0);
// if(repTemp.getSolution(i).crowdingDistance>=MAXDOUBLE)
// weights[i]=exp(1);
// weights[i]=(normalize(value, (smaller-(smaller/100.0)), larger));
if(weights[i]<= 0)
weights[i]=normalize(smaller, (smaller-(smaller/100.0)), larger);//weights[i-1]/2.0;
if(weights[i]>1)
weights[i]=1;
// weights[i]=log(1+weights[i]);
}
weightSum+=weights[i]; //sum of the weights for posterior normalization
}
for(int i=0;i<mu;i++){ //normalization of the weights
weightSquare+=weights[i]*weights[i];
weights[i]/=weightSum;
// printf("\n %f - %f",repTemp.getSolution(i).weightedDistance, weights[i]);
}
// muEff=1.0/weightSquare;//paper
muEff=weightSum*weightSum/weightSquare;//code
cSigma= (muEff+2.0)/(N+muEff+3.0); //code
// cSigma= (muEff+2.0)/(N+muEff+5.0); //paper
dSigma= (1.0 + 2.0*std::max(0.0, sqrt((muEff-1.0)/(N+1.0)) - 1.0)) * std::max(0.3, 1.0 -(double)N / (1e-6+maxIterations)) + cSigma;//code
// dSigma= 1.0+ 2.0*( std::max(0.0, sqrt( (muEff-1.0)/(N+1.0) )-1.0 )) +cSigma; //paper
cc= 4.0/(N+4.0); //code
// cc= (4.0+(muEff/N))/ (N+4.0+(2.0*muEff/N)); //paper
muCov=muEff;
// cCov=((1.0/muCov)*(2.0/( (N+sqrt(2.0))*(N+sqrt(2.0)) )))+((1.0- (1.0/muCov))*std::min(1.0, ((2.0*muEff)-1.0)/( ((N+2.0)*(N+2.0))+muEff ) )); //paper (probably more precise)
double t1 = 2.0 / ((N+1.4142)*(N+1.4142));
double t2 = (2.0*muEff-1.) / ((N+2.0)*(N+2.0)+muEff);
t2 = (t2 > 1) ? 1 : t2;
t2 = (1.0/muCov) * t1 + (1.0-1.0/muCov) * t2;
cCov = t2; //code
chiN = sqrt((double) N) * (1.0 - 1.0/(4.0*N) + 1./(21.0*N*N));
/***************setting default parameters ****************/
if(sw.init){
for (int i = 0; i < N; ++i){ //avoid memory garbage
for (int j = 0; j < N; j++)
sw.B[i][j]=0.0; //need to keep
sw.B[i][i]=1.0; //need to keep
initialStds[i] = 0.3; //does not change //code
trace += initialStds[i]*initialStds[i]; //does not change
}
sw.sigma=sqrt(trace/N); //need to keep
for (int i = 0; i < N; ++i){ //avoid memory garbage
for (int j = 0; j < N; j++)
sw.C[i][j]=0.0; //need to keep
sw.pC[i] = sw.pSigma[i] = 0.0; //need to keep
// sw.mean[i]=0.5; //need to keep
// sw.mean[i]=normalize(sw.centroid[i], inferiorPositionLimit[i], superiorPositionLimit[i]);
if(sw.gen>0)
sw.mean[i]=(rand()/(double)RAND_MAX);//already initialized within the swarm, when restart does not need to restart, does it?
}
for (int i = 0; i < N; ++i) {
sw.C[i][i] = sw.D[i] = initialStds[i] * sqrt(N / trace); //need to keep
sw.C[i][i] *= sw.C[i][i]; //need to keep
}
sw.gen=0;
sw.init=false;
if(logCMAES)
printf("\n INIT \n");
// cmaes_init(&evo, decisionNumber, NULL, NULL, 0, mu, "cmaes_initials.par");
// printf("\nlambda: %d\n",evo.sp.lambda);
}else
sw.gen++;
//exit(1);
//******************************************//
//http://arxiv.org/pdf/1110.4181v1.pdf https://hal.inria.fr/hal-01146738/document //injecting external solutions into moead
sw.solsKept=0;
double clipped[mu][N], in[N], t[N];
for (int i = 0; i < N; i++)
oldMean[i]=sw.mean[i];
for(int s=0;s<mu;s++){ //for all solutions
bool injected=true;
for(int r=0;r<sw.getSize();r++){//check if the current solution is in the population from the previous generation, otherwise it was injected
for(int d=0;d<N;d++)
t[d]=normalize(sw.particles[r].solution.decisionVector[d], inferiorPositionLimit[d], superiorPositionLimit[d]);
if(isEqual(nonDominated[s],t,N)){//if this solution is in the population, it is not injected
injected=false;
sw.solsKept++;
}
}
for (int i = 0; i < N; i++)// eq(2) turning x_i into y_i
in[i]=(nonDominated[s][i]-oldMean[i])/sw.sigma;
if(injected){//first step eq(3) : y_i <-- clip(cy,||C^{-1/2}y_i||)
double sum, tmp[N], sum2=0;
for (int i = 0; i < N; i++) {
sum=0;
for (int j = 0; j < N; ++j)
sum += sw.B[j][i]*in[i];//B^T*y_i
tmp[i] = sum / sw.D[i]; //B^T*y_i*D^-1
}
for (int i = 0; i < N; i++) {
sum=0;
for (int j = 0; j < N; ++j)
sum += sw.B[i][j] * tmp[j];//B^T*y_i*B*D^-1
sum2+=sum*sum;
}
sum2=sqrt(sum2);
double cy=sqrt(N)+((2*N)/N+2);
double clip=cy/sum2; //clip(c,x) = min(1,c/x)
clip=std::min(1.0,clip);
// double * in = nonDominated[s];//x_i
// double sum2=0;//C^{-1/2} * y_i
// for (int i = 0; i < N; ++i) {
// double sum = 0.0;
// for (int j = 0; j < N; ++j)
// sum += sw.B[i][j] * sw.D[i] * ((in[i]-oldMean[i])/sw.sigma ); //y_i
// sum2+=sum*sum;
// }
// sum2=sqrt(sum2);
// double cy=sqrt(N)+((2*N)/N+2);
// double clip=cy/sum2; //clip(c,x) = min(1,c/x)
// clip=std::min(1.0,clip);
// printf(" -- %f %f %f\n----------------------\n",sum2, cy, clip);
for (int i = 0; i < N; ++i)
clipped[s][i]=in[i]*clip;
}else
for (int i = 0; i < N; ++i)
clipped[s][i]=in[i];
}//end of first step, clipping y_i
//second step: \Delta m
//As far as I understood, the differentiation in eq (4) is optional, to be used only if we want a strong impact with an injection, it is not used in the MOEA/D-CMA-ES paper, and is not going to be used here.
double deltaM[N];
for (int i = 0; i < N; ++i){ //update of the delta mean using the second part of eq (4)
deltaM[i]=0;
for(int j=0;j<mu;++j)
deltaM[i]+=weights[j]*clipped[j][i];
}//end of delta m update
//mean update eq(5)
double cm=1;
for (int i = 0; i < N; ++i)
sw.mean[i]=oldMean[i]+cm*sw.sigma*deltaM[i];
//end of mean update
// // //delta mean clip eq(6), but only if strong injection of mean shift, we do not use this
// // if(injected){
// // double * in = deltaM;//deltaM
// // double sum, tmp[N], sum2=0;
// // for (int i = 0; i < N; i++) {
// // sum=0;
// // for (int j = 0; j < N; ++j)
// // sum += sw.B[j][i]*in[i];//B^T*deltaM
// // tmp[i] = sum / sw.D[i]; //B^T*deltaM*D^-1
// // }
// // for (int i = 0; i < N; i++) {
// // sum=0;
// // for (int j = 0; j < N; ++j)
// // sum += sw.B[i][j] * tmp[j];//B^T*deltaM*B*D^-1
// // sum2+=sum*sum;
// // }
// // sum2=sqrt(sum2);
// // double cym=sqrt(2*N)+((2*N)/N+2);
// // double clip=cym/(sqrt(muEff)*sum2); //clip(c,x) = min(1,c/x)
// // clip=std::min(1.0,clip);
// // for (int i = 0; i < N; ++i)
// // deltaM[i]*=clip;//if not injected, deltaM keeps the same
// // }
// // //end of the delta mean clip
//Equations 7 and 9 were updated by changing the BDz calculation, now it is sqrt(muEff)*deltaM
//Equation 8 is a little different, but there is no apparent reason for it, so we did not change it yet
//Equation 10 is changed in the cov matrix
//Equation 11 is changed in the sigma update
//------------------------------------------------------------//
// for (int i = 0; i < N; ++i) {//update of the mean and BDz //original CMA-ES update
// oldMean[i]=sw.mean[i];
// sw.mean[i]=0;
// for(int j=0;j<mu;++j)
// sw.mean[i]+=weights[j]*nonDominated[j][i]; // in the paper sw.mean = m ;; y_w = (mean-oldMean)/sigma
// BDz[i] = sqrt(muEff)*(sw.mean[i] - oldMean[i])/sw.sigma; // BDz * sqrt(muEff) = y_w * sqrt(muEff)
// }
for (int i = 0; i < N; ++i)//BDz with deltaM, as in the paper of injecting solutions into CMA-ES
BDz[i]=sqrt(muEff)*deltaM[i];
// if(decomposition && sw.neighborhood[0].index==swarmNumber/2){//if the first neighbor of it (itself) is zero
// double sum=0;
// for(int i=0;i<N;i++)
// sum+=abs(sw.mean[i]-oldMean[i]);
// printf("%d -- %f\n",repTemp.getActualSize(), sw.sigma);
// for(int i=0;i<mu;i++){
// // printf("%f -- ", repTemp.getSolution(i).weightedDistance);
// for(int d=0;d<N;d++){
// printf("%.4f ", nonDominated[i][d]);
// // printf("%.4f ", repTemp.getSolution(i).decisionVector[d]);
// }
// printf(" (%.4f)\n", getScalarValue(repTemp.getSolution(i).objectiveVector, sw.repository.weight));
// }
// printf("-----------------------------------------\ncovMat:\n");
// for(int i=0;i<N;i++){
// for(int j=0;j<N;j++)
// printf("%.4f ",sw.C[i][j]);
// printf("\n");
// }
// printf("---------------------------------------\n");
// printf("mean: ");
// for(int d=0;d<N;d++){
// // printf("%.4f ", sw.mean[i]);
// printf("%.4f ", sw.mean[d]);
// }
// printf("\nbest: ");
// for(int d=0;d<N;d++){
// printf("%.4f ", normalize(sw.repository.getSolution(0).decisionVector[d], inferiorPositionLimit[d], superiorPositionLimit[d]) );
// }
// printf("\n---------------------------------------------------\n");
// }
double sum, tmp[N], psxps=0.0;
//calculate z := D^(-1) * B^(-1) * BDz into tmp // orignal comment
for (int i = 0; i < N; i++) {
sum=0;
for (int j = 0; j < N; ++j)
sum += sw.B[j][i] * BDz[j];
tmp[i] = sum / sw.D[i]; //tmp = z = D^(-1) * B' * BDz
}
//cumulation for sigma (ps) using B*z //original comment
for (int i = 0; i < N; i++) {
sum=0;
for (int j = 0; j < N; ++j)
sum += sw.B[i][j] * tmp[j]; //sum = sqrt(muEff)*C^(1/2)*y_w = sqrt(muEff)*BD^−1B'*y_w
sw.pSigma[i] = (1.0 - cSigma) * sw.pSigma[i] + sqrt(cSigma * (2.0 - cSigma)) * sum;
/* calculate norm(ps)^2 */
psxps += sw.pSigma[i] * sw.pSigma[i];
}
/* cumulation for covariance matrix (pc) using B*D*z~N(0,C) */
hsig = sqrt(psxps) / sqrt(1.0 - pow(1.0-cSigma, 2*(sw.gen+1))) / chiN < 1.4 + 2.0/(N+1);
for (int i = 0; i < N; ++i)
sw.pC[i] = (1.0 - cc) * sw.pC[i] + hsig * sqrt(cc * (2.0 - cc)) * BDz[i];
//************ Adapt_C2 *********// Update covariance matrix
double ccov1 = std::min(cCov * (1.0/muCov) * 1.0, 1.0); //code
// double ccov1 = 2.0/(((N+1.3)*(N+1.3))+muEff); //paper
double ccovmu = std::min(cCov * (1-1.0/muCov)* 1.0, 1.0-ccov1); //code
// double ccovmu = std::min(1.0-ccov1, 2 * ( ( (muEff-2.0)+(1.0/muEff) )/( (N+2.0)*(N+2.0) + (2.0*muEff)/2.0 ) )); //paper
double sigmasquare = sw.sigma * sw.sigma;
/* update covariance matrix */
for (int i = 0; i < N; ++i)
for (int j = 0; j <= i; ++j) {
sw.C[i][j] = (1.0 - ccov1 - ccovmu) * sw.C[i][j] + ccov1* (sw.pC[i] * sw.pC[j] + (1.0-hsig)*cc*(2.0-cc) * sw.C[i][j]);
for (int k = 0; k < mu; ++k) { /* additional rank mu update */
// sw.C[i][j] += ccovmu * weights[k]* (nonDominated[k][i] - oldMean[i]) * (nonDominated[k][j] - oldMean[j])/ sigmasquare;//original equation
sw.C[i][j] += ccovmu * weights[k] * clipped[k][i] * clipped[k][j];//changed on eq (10) of the injecting solutions into cmaes paper
}
}
//**************************************//
// dSigma= (1.0 + 2.0*std::max(0.0, sqrt((muEff-1.0)/(N+1.0)) - 1.0)) * std::max(0.3, 1.0 -(double)N / (1e-6+std::min(evo.sp.stopMaxIter, evo.sp.stopMaxFunEvals/evo.sp.lambda))) + cSigma;//code (used to compare our algorithm to the implemented)
/* update of sigma */
// sw.sigma *= exp(((sqrt(psxps)/chiN)-1.0)*cSigma/dSigma);//original equation
double deltaSigmaMax=1;
sw.sigma *= exp(std::min(deltaSigmaMax, ((sqrt(psxps)/chiN)-1.0)*cSigma/dSigma ) );//Changed on eq (11) of the injecting solutions into cmaes paper
if(logCMAES){
printf("\n\nmueff %f, cSigma %f, dSigma %f, ccumcov %f, sigma %f", muEff, cSigma, dSigma, cc, sw.sigma);
printf("\nhsig %d, psxps %f, chiN %f, gen %d \n", hsig, psxps, chiN, sw.gen);
// printf("\n\nweights - mu: %d : ", mu);
// printVector(weights, mu);
printf("\navg_old : ");
printVector(oldMean, N);
printf("\navg : ");
printVector(sw.mean, N);
// printf("\n\nD mat:");
// printVector(sw.D, N);
printf("\n\nBDz: ");
printVector(BDz, N);
printf("\ntmp : ");
printVector(tmp, N);
printf("\nps : ");
printVector(sw.pSigma, N);
printf("\npc : ");
printVector(sw.pC, N);
}
// t->rgBDz[i] = sqrt(t->sp.mueff)*(t->rgxmean[i] - t->rgxold[i])/t->sigma;
// printf("\n\nrgD (new): ");
// printVector(D, N);
// printf("\n\nB (new): \n");
// for (int i = 0; i < N; ++i){
// printVector(B[i], N);
// printf("\n");
// }
// printf("\n\nCov mat: \n");
// for (int i = 0; i < N; ++i){
// printVector(sw.C[i], N);
// printf("\n");
// }
//****************comparing the implemented CMA-ES with our version**********************//
// evo.sp.mu=mu;
// evo.sp.lambda=mu;
//
// for(int i=0;i<mu;i++)
// for(int j=0;j<N;j++)
// evo.rgrgx[i][j]=nonDominated[i][j];
// cmaes_SamplePopulation(&evo); /* do not change content of pop */
// for(int i=0;i<mu;i++)
// for(int j=0;j<N;j++)
// evo.rgrgx[i][j]=nonDominated[i][j];
//
// for(int i=0;i<N;i++)
// evo.rgxmean[i]=oldMean[i];
//
// printf("\nccov1 %f", ccov1);
// printf("\nccovmu %f", ccovmu);
// printf("\nsigmasquare %.20f", sigmasquare);
// printf("\nhsig %d", hsig);
// printf("\nccumcov %f", cc);
// printf("\nccov %.20f", cCov);
// printf("\nmucov %.20f", muCov);
// printf("\nmueff: %.20f", muEff);
// printf("\nBDz: ");
// printVector(BDz, N);
// printf("\npSigma: ");
// printVector(sw.pSigma, N);
// printf("\npC: ");
// printVector(sw.pC, N);
// printf("\nD: ");
// for(int i=0;i<N;i++)
// printf("%.15f ",sw.D[i]);
// printf("\npsxps: %f", psxps);
// printf("\nchin: %f", chiN);
// printf("\ncSigma: %f", cSigma);
// printf("\ndSigma: %f", dSigma);
// printf("\nsigma: %f", sw.sigma);
// printf("\nmean: ");
// printVector(sw.mean, N);
// printf("\nold: ");
// printVector(oldMean, N);
// printf("\nweights: ");
// printVector(weights, mu);
//
// cmaes_UpdateDistribution(&evo, weights);
//
// if(evo.sp.ccov != cCov){
// printf("\ncCov different\n");
// exit(1);
// }
// if(evo.sp.mucov != muCov){
// printf("\nmuCov different\n");
// exit(1);
// }
// if(evo.sp.ccumcov != cc){
// printf("\ncc different\n");
// exit(1);
// }
// if(evo.sp.mueff != muEff){
// printf("\nmuEff different\n");
// exit(1);
// }
// if(evo.sp.cs != cSigma){
// printf("\ncSigma different\n");
// exit(1);
// }
// if(evo.sp.damps != dSigma){
// printf("\ndSigma different\n");
// exit(1);
// }
// for(int i=0;i<N;i++){
// if(evo.rgBDz[i] != BDz[i]){
// printf("\nBDz on %d different\n", i);
// exit(1);
// }
// if(evo.rgD[i] != sw.D[i]){
// printf("\nD on %d different\n", i);
// exit(1);
// }
// if(evo.rgdTmp[i] != tmp[i]){
// printf("\ntmp on %d different\n", i);
// exit(1);
// }
// if(evo.rgps[i] != sw.pSigma[i]){
// printf("\npSigma on %d different (%.20f)\n", i, abs(evo.rgps[i] - sw.pSigma[i]));
// exit(1);
// }
// if(evo.rgpc[i] != sw.pC[i]){
// printf("\npC on %d different\n", i);
// exit(1);
// }
// }
//
// printf("\nevo mean: ");
// printVector(evo.rgxmean, N);
// printf("\nevo old: ");
// printVector(evo.rgxold, N);
// printf("\nevo weights: ");
// printVector(evo.sp.weights, mu);
//
// printf("\n\norig-cmaes\n");
// for(int i=0; i<cmaes_Get(&evo, "dim");i++){
// for(int j=0;j<cmaes_Get(&evo, "dim");j++)
// printf("%f ",evo.C[i][j]);
// printf("\n");
// }
//
// printf("\nmodif-cmaes\n");
// for(int i=0; i<N;i++){
// for(int j=0;j<N;j++)
// printf("%f ",sw.C[i][j]);
// printf("\n");
// }
//
// for(int i=0; i<N;i++){
// for(int j=0;j<N;j++){
// if(((int)(evo.C[i][j]*1000000))/1000000.0 != ((int)(sw.C[i][j]*1000000))/1000000.0){
// printf("Different %f -- %f \t (%e) (%d,%d)\n", evo.C[i][j], sw.C[i][j], abs(evo.C[i][j] - sw.C[i][j]), i, j );
// maxDifference=std::max(maxDifference, abs(evo.C[i][j] - sw.C[i][j]));
// // exit(1);
// }
// }
// }
// printf("maxDifference: %.20f, sigmaDifference: %.20f\n", maxDifference, abs(evo.sigma-sw.sigma));
// if(maxDifference > 1 ){
// printf("%d\n", mu);
// exit(1);
// }
//**************************************************************************************//
}