/** Train L2 tree boost. */
private void train2(Attribute[] attributes, double[][] x, int[] y) {
int n = x.length;
int N = (int) Math.round(n * f);
int[] y2 = new int[n];
for (int i = 0; i < n; i++) {
if (y[i] == 1) {
y2[i] = 1;
} else {
y2[i] = -1;
}
}
y = y2;
double[] h = new double[n]; // current F(x_i)
double[] response = new double[n]; // response variable for regression tree.
double mu = Math.mean(y);
b = 0.5 * Math.log((1 + mu) / (1 - mu));
for (int i = 0; i < n; i++) {
h[i] = b;
}
int[][] order = SmileUtils.sort(attributes, x);
RegressionTree.NodeOutput output = new L2NodeOutput(response);
trees = new RegressionTree[T];
int[] perm = new int[n];
int[] samples = new int[n];
for (int i = 0; i < n; i++) {
perm[i] = i;
}
for (int m = 0; m < T; m++) {
Arrays.fill(samples, 0);
Math.permutate(perm);
for (int i = 0; i < N; i++) {
samples[perm[i]] = 1;
}
for (int i = 0; i < n; i++) {
response[i] = 2.0 * y[i] / (1 + Math.exp(2 * y[i] * h[i]));
}
trees[m] = new RegressionTree(attributes, x, response, J, order, samples, output);
for (int i = 0; i < n; i++) {
h[i] += shrinkage * trees[m].predict(x[i]);
}
}
}
/**
* Standard EM algorithm which iteratively alternates Expectation and Maximization steps until
* convergence.
*
* @param components the initial configuration.
* @param x the input data.
* @param gamma the regularization parameter.
* @param maxIter the maximum number of iterations.
* @return log Likelihood
*/
double EM(List<Component> components, double[][] x, double gamma, int maxIter) {
if (x.length < components.size() / 2) throw new IllegalArgumentException("Too many components");
if (gamma < 0.0 || gamma > 0.2)
throw new IllegalArgumentException("Invalid regularization factor gamma.");
int n = x.length;
int m = components.size();
double[][] posteriori = new double[m][n];
// Log Likelihood
double L = 0.0;
for (double[] xi : x) {
double p = 0.0;
for (Component c : components) p += c.priori * c.distribution.p(xi);
if (p > 0) L += Math.log(p);
}
// EM loop until convergence
int iter = 0;
for (; iter < maxIter; iter++) {
// Expectation step
for (int i = 0; i < m; i++) {
Component c = components.get(i);
for (int j = 0; j < n; j++) {
posteriori[i][j] = c.priori * c.distribution.p(x[j]);
}
}
// Normalize posteriori probability.
for (int j = 0; j < n; j++) {
double p = 0.0;
for (int i = 0; i < m; i++) {
p += posteriori[i][j];
}
for (int i = 0; i < m; i++) {
posteriori[i][j] /= p;
}
// Adjust posterior probabilites based on Regularized EM algorithm.
if (gamma > 0) {
for (int i = 0; i < m; i++) {
posteriori[i][j] *= (1 + gamma * Math.log2(posteriori[i][j]));
if (Double.isNaN(posteriori[i][j]) || posteriori[i][j] < 0.0) {
posteriori[i][j] = 0.0;
}
}
}
}
// Maximization step
ArrayList<Component> newConfig = new ArrayList<Component>();
for (int i = 0; i < m; i++)
newConfig.add(
((MultivariateExponentialFamily) components.get(i).distribution).M(x, posteriori[i]));
double sumAlpha = 0.0;
for (int i = 0; i < m; i++) sumAlpha += newConfig.get(i).priori;
for (int i = 0; i < m; i++) newConfig.get(i).priori /= sumAlpha;
double newL = 0.0;
for (double[] xi : x) {
double p = 0.0;
for (Component c : newConfig) {
p += c.priori * c.distribution.p(xi);
}
if (p > 0) newL += Math.log(p);
}
if (newL > L) {
L = newL;
components.clear();
components.addAll(newConfig);
} else {
break;
}
}
return L;
}
