/** calcul du gradient en chaque beta */
private double[] computeGradBeta(
ArrayList<SimpleCacheKernel<T>> kernels, List<TrainingSample<T>> l) {
double grad[] = new double[kernels.size()];
for (int i = 0; i < kernels.size(); i++) {
double matrix[][] = kernels.get(i).getKernelMatrix(l);
double a[] = svm.getAlphas();
for (int x = 0; x < matrix.length; x++) {
int l1 = l.get(x).label;
for (int y = 0; y < matrix.length; y++) {
int l2 = l.get(y).label;
grad[i] += 0.5 * l1 * l2 * a[x] * a[y] * matrix[x][y];
}
}
}
debug.print(3, "gradDir : " + Arrays.toString(grad));
return grad;
}
private double performMKLStep(
double suma,
double[] grad,
ArrayList<SimpleCacheKernel<T>> kernels,
ArrayList<Double> weights,
List<TrainingSample<T>> l) {
debug.print(2, ".");
// compute objective function
double oldObjective = +suma;
for (int i = 0; i < grad.length; i++) {
oldObjective -= weights.get(i) * grad[i];
}
debug.println(3, "oldObjective : " + oldObjective + " sumAlpha : " + suma);
// compute optimal step
double newBeta[] = new double[grad.length];
for (int i = 0; i < grad.length; i++) {
if (grad[i] >= 0 && weights.get(i) >= 0) {
newBeta[i] = grad[i] * weights.get(i) * weights.get(i) / p_norm;
newBeta[i] = Math.pow(newBeta[i], 1 / ((double) 1 + p_norm));
} else newBeta[i] = 0;
}
// normalize
double norm = 0;
for (int i = 0; i < newBeta.length; i++) norm += Math.pow(newBeta[i], p_norm);
norm = Math.pow(norm, -1 / (double) p_norm);
if (norm < 0) {
debug.println(1, "Error normalization, norm < 0");
return -1;
}
for (int i = 0; i < newBeta.length; i++) newBeta[i] *= norm;
// regularize and renormalize
double R = 0;
for (int i = 0; i < kernels.size(); i++) R += Math.pow(weights.get(i) - newBeta[i], 2);
R = Math.sqrt(R / (double) p_norm) * eps_regul;
if (R < 0) {
debug.println(1, "Error regularization, R < 0");
return -1;
}
norm = 0;
for (int i = 0; i < kernels.size(); i++) {
newBeta[i] += R;
if (newBeta[i] < num_cleaning) newBeta[i] = 0;
norm += Math.pow(newBeta[i], p_norm);
}
norm = Math.pow(norm, -1 / (double) p_norm);
if (norm < 0) {
debug.println(1, "Error normalization, norm < 0");
return -1;
}
for (int i = 0; i < newBeta.length; i++) newBeta[i] *= norm;
// store new weights
for (int i = 0; i < weights.size(); i++) weights.set(i, newBeta[i]);
// compute objective function
double objective = +suma;
for (int i = 0; i < grad.length; i++) {
objective -= weights.get(i) * grad[i];
}
debug.println(3, "objective : " + objective + " sumAlpha : " + suma);
// return objective evolution
return objective / oldObjective;
}