private void testAndCorrectNumerics(
CMAEvolutionaryAlgorithm algorithm, int N, List<Individual> individuals) {
/*
* Flat fitness, Test if funtion values are identical
*/
/*
* we should sort the individuals by fitness value
*/
Collections.sort(individuals, algorithm.getComparator());
if (algorithm.getGenerations() > 0) {
if (individuals.get(0).getFitness()
== individuals
.get((int) Math.min(algorithm.getLambda() - 1, algorithm.getLambda() / 2.0 + 1) - 1)
.getFitness()) {
if (debug) {
System.err.println(
"WARNING - Flat fitness landscape, consider reformulation of fitness,"
+ "step-size increased");
}
algorithm.setSigma(
algorithm.getSigma() * Math.exp((0.2 + algorithm.getCs() / algorithm.getDamps())));
}
}
/*
* Align (renormalize) scale C and (consequently sigma)
*/
/*
* e.g. for infinite stationary state simulations (noise handling needs to be introduces for
* that)
*/
/*
* handling needs to be introduced for that)
*/
double fac = 1;
if (StatUtils.max(algorithm.getDiag()) < 1.0e-6) {
fac = 1.0 / StatUtils.max(algorithm.getDiag());
} else if (StatUtils.min(algorithm.getDiag()) > 1.0e+4) {
fac = 1.0 / StatUtils.min(algorithm.getDiag());
}
if (fac != 1.0) {
algorithm.setSigma(algorithm.getSigma() / fac);
for (int i = 0; i < N; i++) {
algorithm.getPc()[i] *= fac;
algorithm.getDiag()[i] *= fac;
for (int j = 0; j <= i; j++) {
algorithm.getC()[i][j] *= fac * fac;
}
}
}
}
private void eidgendecomposition(CMAEvolutionaryAlgorithm algorithm, int N, int flgforce) {
if (countCUpdatesSinceEigenupdate == 0 && flgforce < 2) {
return;
}
if (!algorithm.getFlgDiag()
&& flgforce <= 0
&& countCUpdatesSinceEigenupdate < 1.0 / algorithm.getCcov() / N / 5.0) {
return;
}
if (algorithm.getFlgDiag()) {
for (int i = 0; i < N; i++) {
algorithm.getDiag()[i] = Math.sqrt(algorithm.getC()[i][i]);
}
algorithm.setCountCupdatesSinceEigenupdate(0);
} else {
// set B <- C
for (int i = 0; i < N; i++) {
for (int j = 0; j <= i; j++) {
algorithm.getB()[i][j] = algorithm.getB()[j][i] = algorithm.getC()[i][j];
}
}
// eigendecomposition:
double[] offdiag = new double[N];
tred2(N, algorithm.getB(), algorithm.getDiag(), offdiag);
tql2(N, algorithm.getDiag(), offdiag, algorithm.getB());
checkEigenSystem(N, algorithm.getC(), algorithm.getDiag(), algorithm.getB());
for (int i = 0; i < N; i++) {
if (algorithm.getDiag()[i] < 0.0) {
if (algorithm.isDebug()) {
System.err.println("ERROR - An eigenvalue has become negative.");
}
}
algorithm.getDiag()[i] = Math.sqrt(algorithm.getDiag()[i]);
}
algorithm.setCountCupdatesSinceEigenupdate(0);
}
/*
* TODO: Check if necessary
*/
if (StatUtils.min(algorithm.getDiag()) == 0.0) {
algorithm.setAxisRatio(Double.POSITIVE_INFINITY);
} else {
algorithm.setAxisRatio(
StatUtils.max(algorithm.getDiag()) / StatUtils.min(algorithm.getDiag()));
}
} // eigendecomposition
@Override
public List<Individual> operate(EvolutionaryAlgorithm algorithm, List<Individual> individuals)
throws OperatorException {
List<Individual> new_population = new ArrayList<Individual>(individuals.size());
CMAEvolutionaryAlgorithm alg = (CMAEvolutionaryAlgorithm) algorithm;
debug = algorithm.isDebug();
double[] chromosome;
int N = individuals.get(0).getDimension();
countCUpdatesSinceEigenupdate = alg.getCountCupdatesSinceEigenupdate();
// latest possibility to generate B and diagD:
eidgendecomposition(alg, N, 0);
// ensure minimal and maximal standard deviation:
double minsqrtdiagC = Math.sqrt(StatUtils.min(alg.diag(alg.getC())));
double maxsqrtdiagC = Math.sqrt(StatUtils.max(alg.diag(alg.getC())));
/*
* TODO: test if it is necessary *
*/
for (int i = 0; i < individuals.size(); i++) {
new_population.add((Individual) individuals.get(i).clone());
}
if (lowerStandardDeviation != null && lowerStandardDeviation.length > 0) {
for (int i = 0; i < N; i++) {
double d = lowerStandardDeviation[Math.min(i, lowerStandardDeviation.length - 1)];
if (d > alg.getSigma() * minsqrtdiagC) {
alg.setSigma(d / minsqrtdiagC);
}
}
}
if (upperStandardDeviation != null && upperStandardDeviation.length > 0) {
for (int i = 0; i < N; i++) {
double d = upperStandardDeviation[Math.min(i, upperStandardDeviation.length - 1)];
if (d < alg.getSigma() * maxsqrtdiagC) {
alg.setSigma(d / maxsqrtdiagC);
}
}
}
testAndCorrectNumerics(alg, N, new_population);
double[] artmp = new double[N];
/*
* Sample the distribution
*/
for (int iNk = 0; iNk < alg.getLambda(); iNk++) {
chromosome = new_population.get(iNk).getChromosomeAt(0);
if (alg.getFlgDiag()) {
for (int i = 0; i < N; i++) {
chromosome[i] =
this.checkBounds(
alg,
alg.getxMean()[i] + alg.getSigma() * alg.getDiag()[i] * EAFRandom.nextGaussian());
}
} else {
for (int i = 0; i < N; i++) {
artmp[i] = alg.getDiag()[i] * EAFRandom.nextGaussian();
}
/*
* 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 += alg.getB()[i][j] * artmp[j];
}
chromosome[i] = this.checkBounds(alg, alg.getxMean()[i] + alg.getSigma() * sum);
}
}
new_population.get(iNk).setChromosomeAt(0, chromosome);
}
return new_population;
}