@Override
public void nextIteration() {
double[] beta = getBestValuesEver();
for (int i = 0; i < beta.length; i++) {
double minBound = getMin(i);
double maxBound = getMax(i);
if (beta[i] == minBound) {
minBound -= 1;
setMin(i, minBound);
logging.log("decrease lower bound for beta(" + i + ") to + " + minBound);
}
if (beta[i] == maxBound) {
maxBound += 1;
setMax(i, maxBound);
logging.log("increase upper bound for beta(" + i + ") to + " + maxBound);
}
}
}
private double[] estimateVariance() {
double[] beta = getBestValuesEver();
Matrix hessian = new Matrix(beta.length, beta.length);
for (Example example : exampleSet) {
double[] values = new double[beta.length];
double eta = 0.0d;
int j = 0;
for (Attribute attribute : example.getAttributes()) {
double value = example.getValue(attribute);
values[j] = value;
eta += beta[j] * value;
j++;
}
if (addIntercept) {
values[beta.length - 1] = 1.0d;
eta += beta[beta.length - 1];
}
double pi = Math.exp(eta) / (1 + Math.exp(eta));
double weightValue = 1.0d;
if (weight != null) weightValue = example.getValue(weight);
for (int x = 0; x < beta.length; x++) {
for (int y = 0; y < beta.length; y++) {
// sum is second derivative of log likelihood function
double h = hessian.get(x, y) - values[x] * values[y] * weightValue * pi * (1 - pi);
hessian.set(x, y, h);
}
}
}
double[] variance = new double[beta.length];
Matrix varianceCovarianceMatrix = null;
try {
// asymptotic variance-covariance matrix is inverse of hessian matrix
varianceCovarianceMatrix = hessian.inverse();
} catch (Exception e) {
logging.logWarning("could not determine variance-covariance matrix, hessian is singular");
for (int j = 0; j < beta.length; j++) {
variance[j] = Double.NaN;
}
return variance;
}
for (int j = 0; j < beta.length; j++) {
// get diagonal elements
variance[j] = Math.abs(varianceCovarianceMatrix.get(j, j));
}
return variance;
}
