public virtual DMatrixRMaj getU(DMatrixRMaj U, bool transpose, bool compact)
{
U = BidiagonalDecompositionRow_DDRM.handleU(U, false, compact, m, n, min);
if (compact)
{
// U = Q*U1
DMatrixRMaj Q1 = decompQRP.getQ(null, true);
DMatrixRMaj U1 = decompBi.getU(null, false, true);
CommonOps_DDRM.mult(Q1, U1, U);
}
else
{
// U = [Q1*U1 Q2]
DMatrixRMaj Q = decompQRP.getQ(U, false);
DMatrixRMaj U1 = decompBi.getU(null, false, true);
DMatrixRMaj Q1 = CommonOps_DDRM.extract(Q, 0, Q.numRows, 0, min);
DMatrixRMaj tmp = new DMatrixRMaj(Q1.numRows, U1.numCols);
CommonOps_DDRM.mult(Q1, U1, tmp);
CommonOps_DDRM.insert(tmp, Q, 0, 0);
}
if (transpose)
{
CommonOps_DDRM.transpose(U);
}
return(U);
}
/// <summary>
/// Copy matrix B into this matrix at location (insertRow, insertCol).
/// </summary>
public override void insertIntoThis(int insertRow, int insertCol, SimpleMatrixD B)
{
var bm = B.getMatrix();
CommonOps_DDRM.insert(bm, mat, insertRow, insertCol);
}
/** Revises the CMA-ES distribution to reflect the current fitness results in the provided subpopulation. */
public void UpdateDistribution(IEvolutionState state, Subpopulation subpop)
{
// % Sort by fitness and compute weighted mean into xmean
// [arfitness, arindex] = sort(arfitness); % minimization
// xmean = arx(:,arindex(1:mu))*weights; % recombination % Eq.39
// counteval += lambda;
// only need partial sort?
((List <Individual>)subpop.Individuals).Sort();
SimpleMatrixD artmp = new SimpleMatrixD(GenomeSize, mu);
SimpleMatrixD xold = xmean;
xmean = new SimpleMatrixD(GenomeSize, 1);
for (int i = 0; i < mu; i++)
{
DoubleVectorIndividual dvind = (DoubleVectorIndividual)subpop.Individuals[i];
// won't modify the genome
SimpleMatrixD arz = new SimpleMatrixD(GenomeSize, 1, true, dvind.genome);
arz = (arz.minus(xold).divide(sigma));
for (int j = 0; j < GenomeSize; j++)
{
xmean.set(j, 0, xmean.get(j, 0) + weights[i] * dvind.genome[j]);
artmp.set(j, i, arz.get(j, 0));
}
}
// % Cumulation: Update evolution paths
SimpleMatrixD y = xmean.minus(xold).divide(sigma);
SimpleMatrixD bz = invsqrtC.mult(y);
SimpleMatrixD bz_scaled = bz.scale(Math.Sqrt(cs * (2.0 - cs) * mueff));
ps = ps.scale(1.0 - cs).plus(bz_scaled);
double h_sigma_value =
((ps.dot(ps) / (1.0 - Math.Pow(1.0 - cs, 2.0 * (state.Generation + 1)))) / GenomeSize);
int hsig = (h_sigma_value < (2.0 + (4.0 / (GenomeSize + 1)))) ? 1 : 0;
SimpleMatrixD y_scaled = y.scale(hsig * Math.Sqrt(cc * (2.0 - cc) * mueff));
pc = pc.scale(1.0 - cc).plus(y_scaled);
// % Adapt covariance matrix C
c = c.scale(1.0 - c1 - cmu);
c = c.plus(pc.mult(pc.transpose()).plus(c.scale((1.0 - hsig) * cc * (2.0 - cc))).scale(c1));
c = c.plus((artmp.mult(SimpleMatrixD.diag(weights).mult(artmp.transpose()))).scale(cmu));
// % Adapt step-size sigma
sigma = sigma * Math.Exp((cs / damps) * (ps.normF() / chiN - 1.0));
// % Update B and D from C
if ((state.Generation - lastEigenDecompositionGeneration) > 1.0 / ((c1 + cmu) * GenomeSize * 10.0))
{
lastEigenDecompositionGeneration = state.Generation;
// make sure the matrix is symmetric (it should be already)
// not sure if this is necessary
for (int i = 0; i < GenomeSize; i++)
{
for (int j = 0; j < i; j++)
{
c.set(j, i, c.get(i, j));
}
}
// this copy gets modified by the decomposition
DMatrixRMaj copy = c.copy().getMatrix();
EigenDecomposition <DMatrixRMaj> eig = DecompositionFactory_DDRM.eig(GenomeSize, true, true);
if (eig.decompose(copy))
{
SimpleMatrixD dinv = new SimpleMatrixD(GenomeSize, GenomeSize);
for (int i = 0; i < GenomeSize; i++)
{
double eigrt = Math.Sqrt(eig.getEigenValue(i).real);
d.set(i, i, eigrt);
dinv.set(i, i, 1 / eigrt);
CommonOps_DDRM.insert(eig.getEigenVector(i), b.getMatrix(), 0, i);
}
invsqrtC = b.mult(dinv.mult(b.transpose()));
CommonOps_DDRM.mult(b.getMatrix(), d.getMatrix(), bd);
}
else
{
state.Output.Fatal("CMA-ES eigendecomposition failed. ");
}
}
CommonOps_DDRM.scale(sigma, bd, sbd);
// % Break, if fitness is good enough or condition exceeds 1e14, better termination methods are advisable
// if arfitness(1) <= stopfitness || max(D) > 1e7 * min(D)
// break;
// end
if (useAltTermination && CommonOps_DDRM.elementMax(d.diag().getMatrix()) >
1e7 * CommonOps_DDRM.elementMin(d.diag().getMatrix()))
{
state.Evaluator.SetRunCompleted("CMAESSpecies: Stopped because matrix condition exceeded limit.");
}
}
public override void Setup(IEvolutionState state, IParameter paramBase)
{
base.Setup(state, paramBase);
IMersenneTwister random = state.Random[0];
IParameter def = DefaultBase;
IParameter subpopBase = paramBase.Pop();
IParameter subpopDefaultBase = ECDefaults.ParamBase.Push(Subpopulation.P_SUBPOPULATION);
if (!state.Parameters.ParameterExists(paramBase.Push(P_SIGMA), def.Push(P_SIGMA)))
{
state.Output.Message("CMA-ES sigma was not provided, defaulting to 1.0");
sigma = 1.0;
}
else
{
sigma = state.Parameters.GetDouble(paramBase.Push(P_SIGMA), def.Push(P_SIGMA), 0.0);
if (sigma <= 0)
{
state.Output.Fatal("If CMA-ES sigma is provided, it must be > 0.0", paramBase.Push(P_SIGMA),
def.Push(P_SIGMA));
}
}
double[] cvals = new double[GenomeSize];
string covarianceInitialization =
state.Parameters.GetStringWithDefault(paramBase.Push(P_COVARIANCE), def.Push(P_COVARIANCE), V_IDENTITY);
string covs = "Initial Covariance: <";
for (int i = 0; i < GenomeSize; i++)
{
if (i > 0)
{
covs += ", ";
}
if (covarianceInitialization.Equals(V_SCALED))
{
cvals[i] = (MaxGenes[i] - MinGenes[i]);
}
else if (covarianceInitialization.Equals(V_IDENTITY))
{
cvals[i] = 1.0;
}
else
{
state.Output.Fatal("Invalid covariance initialization type " + covarianceInitialization,
paramBase.Push(P_COVARIANCE), def.Push(P_COVARIANCE));
}
// cvals is standard deviations, so we change them to variances now
cvals[i] *= cvals[i];
covs += cvals[i];
}
state.Output.Message(covs + ">");
// set myself up and define my initial distribution here
int n = GenomeSize;
b = SimpleMatrixD.identity(n);
c = new SimpleMatrixD(CommonOps_DDRM.diag(cvals));
d = SimpleMatrixD.identity(n);
bd = CommonOps_DDRM.identity(n, n);
sbd = CommonOps_DDRM.identity(n, n);
invsqrtC = SimpleMatrixD.identity(n);
// Here we do one FIRST round of eigendecomposition, because newIndividual needs
// a valid version of sbd. If c is initially the identity matrix (and sigma = 1),
// then sbd is too, and we're done. But if c is scaled in any way, we need to compute
// the proper value of sbd. Along the way we'll wind up computing b, d, bd, and invsqrtC
EigenDecomposition <DMatrixRMaj> eig = DecompositionFactory_DDRM.eig(GenomeSize, true, true);
if (eig.decompose(c.copy().getMatrix()))
{
SimpleMatrixD dinv = new SimpleMatrixD(GenomeSize, GenomeSize);
for (int i = 0; i < GenomeSize; i++)
{
double eigrt = Math.Sqrt(eig.getEigenValue(i).real);
d.set(i, i, eigrt);
dinv.set(i, i, 1 / eigrt);
CommonOps_DDRM.insert(eig.getEigenVector(i), b.getMatrix(), 0, i);
}
invsqrtC = b.mult(dinv.mult(b.transpose()));
CommonOps_DDRM.mult(b.getMatrix(), d.getMatrix(), bd);
}
else
{
state.Output.Fatal("CMA-ES eigendecomposition failed. ");
}
CommonOps_DDRM.scale(sigma, bd, sbd);
// End FIRST round of eigendecomposition
// Initialize dynamic (internal) strategy parameters and constants
pc = new SimpleMatrixD(n, 1);
ps = new SimpleMatrixD(n, 1); // evolution paths for C and sigma
chiN = Math.Sqrt(n) *
(1.0 - 1.0 / (4.0 * n) + 1.0 / (21.0 * n * n)); // expectation of ||N(0,I)|| == norm(randn(N,1))
xmean = new SimpleMatrixD(GenomeSize, 1);
bool meanSpecified = false;
string val = state.Parameters.GetString(paramBase.Push(P_MEAN), def.Push(P_MEAN));
if (val != null)
{
meanSpecified = true;
if (val.Equals(V_CENTER))
{
for (int i = 0; i < GenomeSize; i++)
{
xmean.set(i, 0, (MaxGenes[i] + MinGenes[i]) / 2.0);
}
}
else if (val.Equals(V_ZERO))
{
for (int i = 0; i < GenomeSize; i++)
{
xmean.set(i, 0, 0); // it is this anyway
}
}
else if (val.Equals(V_RANDOM))
{
for (int i = 0; i < GenomeSize; i++)
{
xmean.set(i, 0,
state.Random[0].NextDouble(true, true) * (MaxGenes[i] - MinGenes[i]) + MinGenes[i]);
}
}
else
{
state.Output.Fatal("Unknown mean value specified: " + val, paramBase.Push(P_MEAN), def.Push(P_MEAN));
}
}
else
{
state.Output.Fatal("No default mean value specified. Loading full mean from parameters.",
paramBase.Push(P_MEAN), def.Push(P_MEAN));
}
bool nonDefaultMeanSpecified = false;
for (int i = 0; i < GenomeSize; i++)
{
double m_i = 0;
try
{
m_i = state.Parameters.GetDouble(paramBase.Push(P_MEAN).Push("" + i), def.Push(P_MEAN).Push("" + i));
xmean.set(i, 0, m_i);
nonDefaultMeanSpecified = true;
}
catch (FormatException e)
{
if (!meanSpecified)
{
state.Output.Error(
"No default mean value was specified, but CMA-ES mean index " + i +
" is missing or not a number.", paramBase.Push(P_MEAN).Push("" + i),
def.Push(P_MEAN).Push("" + i));
}
}
}
state.Output.ExitIfErrors();
if (nonDefaultMeanSpecified && meanSpecified)
{
state.Output.Warning("A default mean value was specified, but certain mean values were overridden.");
}
string mes = "Initial Mean: <";
for (int i = 0; i < GenomeSize - 1; i++)
{
mes = mes + xmean.get(i, 0) + ", ";
}
mes = mes + xmean.get(GenomeSize - 1, 0) + ">";
state.Output.Message(mes);
if (!state.Parameters.ParameterExists(paramBase.Push(P_LAMBDA), def.Push(P_LAMBDA)))
{
lambda = 4 + (int)Math.Floor(3 * Math.Log(n));
}
else
{
lambda = state.Parameters.GetInt(paramBase.Push(P_LAMBDA), def.Push(P_LAMBDA), 1);
if (lambda <= 0)
{
state.Output.Fatal("If the CMA-ES lambda parameter is provided, it must be a valid integer > 0",
paramBase.Push(P_LAMBDA), def.Push(P_LAMBDA));
}
}
if (!state.Parameters.ParameterExists(paramBase.Push(P_MU), def.Push(P_MU)))
{
mu = (int)(Math.Floor(lambda / 2.0));
}
else
{
mu = state.Parameters.GetInt(paramBase.Push(P_MU), def.Push(P_MU), 1);
if (mu <= 0)
{
state.Output.Fatal("If the CMA-ES mu parameter is provided, it must be a valid integer > 0",
paramBase.Push(P_MU), def.Push(P_MU));
}
}
if (mu > lambda) // uh oh
{
state.Output.Fatal("CMA-ES mu must be <= lambda. Presently mu=" + mu + " and lambda=" + lambda);
}
weights = new double[mu];
bool weightsSpecified = false;
for (int i = 0; i < mu; i++)
{
if (state.Parameters.ParameterExists(paramBase.Push(P_WEIGHTS).Push("" + i), def.Push(P_WEIGHTS).Push("" + i)))
{
state.Output.Message("CMA-ES weight index " + i +
" specified. Loading all weights from parameters.");
weightsSpecified = true;
break;
}
}
if (weightsSpecified)
{
for (int i = 0; i < mu; i++)
{
double m_i = 0;
try
{
weights[i] = state.Parameters.GetDouble(paramBase.Push(P_WEIGHTS).Push("" + i),
def.Push(P_WEIGHTS).Push("" + i));
}
catch (FormatException e)
{
state.Output.Error("CMA-ES weight index " + i + " missing or not a number.",
paramBase.Push(P_WEIGHTS).Push("" + i), def.Push(P_WEIGHTS).Push("" + i));
}
}
state.Output.ExitIfErrors();
}
else
{
for (int i = 0; i < mu; i++)
{
weights[i] = Math.Log((lambda + 1.0) / (2.0 * (i + 1)));
}
}
// normalize
double sum = 0.0;
for (int i = 0; i < mu; i++)
{
sum += weights[i];
}
for (int i = 0; i < mu; i++)
{
weights[i] /= sum;
}
// compute mueff
double sumSqr = 0.0;
for (int i = 0; i < mu; i++)
{
sumSqr += weights[i] * weights[i];
}
mueff = 1.0 / sumSqr;
mes = "Weights: <";
for (int i = 0; i < weights.Length - 1; i++)
{
mes = mes + weights[i] + ", ";
}
mes = mes + (weights.Length - 1) + ">";
state.Output.Message(mes);
useAltTermination = state.Parameters.GetBoolean(paramBase.Push(P_ALTERNATIVE_TERMINATION),
def.Push(P_ALTERNATIVE_TERMINATION), false);
useAltGenerator = state.Parameters.GetBoolean(paramBase.Push(P_ALTERNATIVE_GENERATOR),
def.Push(P_ALTERNATIVE_GENERATOR), false);
altGeneratorTries = state.Parameters.GetIntWithDefault(paramBase.Push(P_ALTERNATIVE_GENERATOR_TRIES),
def.Push(P_ALTERNATIVE_GENERATOR_TRIES), DEFAULT_ALT_GENERATOR_TRIES);
if (altGeneratorTries < 1)
{
state.Output.Fatal(
"If specified (the default is " + DEFAULT_ALT_GENERATOR_TRIES +
"), alt-generation-tries must be >= 1",
paramBase.Push(P_ALTERNATIVE_GENERATOR_TRIES), def.Push(P_ALTERNATIVE_GENERATOR_TRIES));
}
if (!state.Parameters.ParameterExists(paramBase.Push(P_CC), def.Push(P_CC)))
{
cc = (4.0 + mueff / n) / (n + 4.0 + 2.0 * mueff / n); // time constant for cumulation for C
}
else
{
cc = state.Parameters.GetDoubleWithMax(paramBase.Push(P_CC), def.Push(P_CC), 0.0, 1.0);
if (cc < 0.0)
{
state.Output.Fatal(
"If the CMA-ES cc parameter is provided, it must be a valid number in the range [0,1]",
paramBase.Push(P_CC), def.Push(P_CC));
}
}
if (!state.Parameters.ParameterExists(paramBase.Push(P_CS), def.Push(P_CS)))
{
cs = (mueff + 2.0) / (n + mueff + 5.0); // t-const for cumulation for sigma control
}
else
{
cs = state.Parameters.GetDoubleWithMax(paramBase.Push(P_CS), def.Push(P_CS), 0.0, 1.0);
if (cs < 0.0)
{
state.Output.Fatal(
"If the CMA-ES cs parameter is provided, it must be a valid number in the range [0,1]",
paramBase.Push(P_CS), def.Push(P_CS));
}
}
if (!state.Parameters.ParameterExists(paramBase.Push(P_C1), def.Push(P_C1)))
{
c1 = 2.0 / ((n + 1.3) * (n + 1.3) + mueff); // learning rate for rank-one update of C
}
else
{
c1 = state.Parameters.GetDouble(paramBase.Push(P_C1), def.Push(P_C1), 0.0);
if (c1 < 0)
{
state.Output.Fatal("If the CMA-ES c1 parameter is provided, it must be a valid number >= 0.0",
paramBase.Push(P_C1), def.Push(P_C1));
}
}
if (!state.Parameters.ParameterExists(paramBase.Push(P_CMU), def.Push(P_CMU)))
{
cmu = Math.Min(1.0 - c1, 2.0 * (mueff - 2.0 + 1.0 / mueff) / ((n + 2.0) * (n + 2.0) + mueff));
}
else
{
cmu = state.Parameters.GetDouble(paramBase.Push(P_CMU), def.Push(P_CMU), 0.0);
if (cmu < 0)
{
state.Output.Fatal("If the CMA-ES cmu parameter is provided, it must be a valid number >= 0.0",
paramBase.Push(P_CMU), def.Push(P_CMU));
}
}
if (c1 > (1 - cmu)) // uh oh
{
state.Output.Fatal("CMA-ES c1 must be <= 1 - cmu. You are using c1=" + c1 + " and cmu=" + cmu);
}
if (cmu > (1 - c1)) // uh oh
{
state.Output.Fatal("CMA-ES cmu must be <= 1 - c1. You are using cmu=" + cmu + " and c1=" + c1);
}
if (!state.Parameters.ParameterExists(paramBase.Push(P_DAMPS), def.Push(P_DAMPS)))
{
damps = 1.0 + 2.0 * Math.Max(0.0, Math.Sqrt((mueff - 1.0) / (n + 1.0)) - 1.0) + cs; // damping for sigma
}
else
{
damps = state.Parameters.GetDouble(paramBase.Push(P_DAMPS), def.Push(P_DAMPS), 0.0);
if (damps <= 0)
{
state.Output.Fatal("If the CMA-ES damps parameter is provided, it must be a valid number > 0.0",
paramBase.Push(P_DAMPS), def.Push(P_DAMPS));
}
}
double damps_min = 0.5;
double damps_max = 2.0;
if (damps > damps_max || damps < damps_min)
{
state.Output.Warning("CMA-ES damps ought to be close to 1. You are using damps = " + damps);
}
state.Output.Message("lambda: " + lambda);
state.Output.Message("mu: " + mu);
state.Output.Message("mueff: " + mueff);
state.Output.Message("cmu: " + cmu);
state.Output.Message("c1: " + c1);
state.Output.Message("cc: " + cc);
state.Output.Message("cs: " + cs);
state.Output.Message("damps: " + damps);
}
public void construct()
{
// make sure the last item is and end row
if (!items[items.Count - 1].endRow)
{
endRow();
}
// have to initialize some variable types first to get the actual size
for (int i = 0; i < items.Count; i++)
{
items[i].initialize();
}
setToRequiredSize(output.matrix);
int matrixRow = 0;
List <Item> row = new List <Item>();
for (int i = 0; i < items.Count; i++)
{
Item item = items[i];
if (item.endRow)
{
int expectedRows = 0;
int numCols = 0;
for (int j = 0; j < row.Count; j++)
{
Item v = row[j];
int numRows = v.getRows();
if (j == 0)
{
expectedRows = numRows;
}
else if (v.getRows() != expectedRows)
{
throw new InvalidOperationException("Row miss-matched. " + numRows + " " + v.getRows());
}
if (v.matrix)
{
CommonOps_DDRM.insert(v.getMatrix(), output.matrix, matrixRow, numCols);
}
else if (v.variable.getType() == VariableType.SCALAR)
{
output.matrix.set(matrixRow, numCols, v.getValue());
}
else if (v.variable.getType() == VariableType.INTEGER_SEQUENCE)
{
IntegerSequence sequence = ((VariableIntegerSequence)v.variable).sequence;
int col = numCols;
while (sequence.hasNext())
{
output.matrix.set(matrixRow, col++, sequence.next());
}
}
else
{
throw new ParseError("Can't insert a variable of type " + v.variable.getType() +
" inside a matrix!");
}
numCols += v.getColumns();
}
matrixRow += expectedRows;
row.Clear();
}
else
{
row.Add(item);
}
}
}
/**
* Computes the QR decomposition of the provided matrix.
*
* @param A Matrix which is to be decomposed. Not modified.
*/
public void decompose(DMatrixRMaj A)
{
this.QR = (DMatrixRMaj)A.copy();
int N = Math.Min(A.numCols, A.numRows);
gammas = new double[A.numCols];
DMatrixRMaj A_small = new DMatrixRMaj(A.numRows, A.numCols);
DMatrixRMaj A_mod = new DMatrixRMaj(A.numRows, A.numCols);
DMatrixRMaj v = new DMatrixRMaj(A.numRows, 1);
DMatrixRMaj Q_k = new DMatrixRMaj(A.numRows, A.numRows);
for (int i = 0; i < N; i++)
{
// reshape temporary variables
A_small.reshape(QR.numRows - i, QR.numCols - i, false);
A_mod.reshape(A_small.numRows, A_small.numCols, false);
v.reshape(A_small.numRows, 1, false);
Q_k.reshape(v.getNumElements(), v.getNumElements(), false);
// use extract matrix to get the column that is to be zeroed
CommonOps_DDRM.extract(QR, i, QR.numRows, i, i + 1, v, 0, 0);
double max = CommonOps_DDRM.elementMaxAbs(v);
if (max > 0 && v.getNumElements() > 1)
{
// normalize to reduce overflow issues
CommonOps_DDRM.divide(v, max);
// compute the magnitude of the vector
double tau = NormOps_DDRM.normF(v);
if (v.get(0) < 0)
{
tau *= -1.0;
}
double u_0 = v.get(0) + tau;
double gamma = u_0 / tau;
CommonOps_DDRM.divide(v, u_0);
v.set(0, 1.0);
// extract the submatrix of A which is being operated on
CommonOps_DDRM.extract(QR, i, QR.numRows, i, QR.numCols, A_small, 0, 0);
// A = (I - γ*u*u<sup>T</sup>)A
CommonOps_DDRM.setIdentity(Q_k);
CommonOps_DDRM.multAddTransB(-gamma, v, v, Q_k);
CommonOps_DDRM.mult(Q_k, A_small, A_mod);
// save the results
CommonOps_DDRM.insert(A_mod, QR, i, i);
CommonOps_DDRM.insert(v, QR, i, i);
QR.unsafe_set(i, i, -tau * max);
// save gamma for recomputing Q later on
gammas[i] = gamma;
}
}
}