@Override
protected void computeBeta(final HiddenMarkovModel hmm, final MLDataSet oseq) {
this.beta = new double[oseq.size()][hmm.getStateCount()];
for (int i = 0; i < hmm.getStateCount(); i++) {
this.beta[oseq.size() - 1][i] = 1. / this.ctFactors[oseq.size() - 1];
}
for (int t = oseq.size() - 2; t >= 0; t--) {
for (int i = 0; i < hmm.getStateCount(); i++) {
computeBetaStep(hmm, oseq.get(t + 1), t, i);
this.beta[t][i] /= this.ctFactors[t];
}
}
}
static HiddenMarkovModel buildDiscHMM() {
HiddenMarkovModel hmm = new HiddenMarkovModel(2, 2);
hmm.setPi(0, 0.95);
hmm.setPi(1, 0.05);
hmm.setStateDistribution(0, new DiscreteDistribution(new double[][] {{0.95, 0.05}}));
hmm.setStateDistribution(1, new DiscreteDistribution(new double[][] {{0.20, 0.80}}));
hmm.setTransitionProbability(0, 1, 0.05);
hmm.setTransitionProbability(0, 0, 0.95);
hmm.setTransitionProbability(1, 0, 0.10);
hmm.setTransitionProbability(1, 1, 0.90);
return hmm;
}
private void computeStep(
final HiddenMarkovModel hmm, final MLDataPair o, final int t, final int j) {
double minDelta = Double.MAX_VALUE;
int min_psy = 0;
for (int i = 0; i < hmm.getStateCount(); i++) {
final double thisDelta = this.delta[t - 1][i] - Math.log(hmm.getTransitionProbability(i, j));
if (minDelta > thisDelta) {
minDelta = thisDelta;
min_psy = i;
}
}
this.delta[t][j] = minDelta - Math.log(hmm.getStateDistribution(j).probability(o));
this.psy[t][j] = min_psy;
}
@Override
protected void computeAlpha(final HiddenMarkovModel hmm, final MLDataSet oseq) {
this.alpha = new double[oseq.size()][hmm.getStateCount()];
for (int i = 0; i < hmm.getStateCount(); i++) {
computeAlphaInit(hmm, oseq.get(0), i);
}
scale(this.ctFactors, this.alpha, 0);
final Iterator<MLDataPair> seqIterator = oseq.iterator();
if (seqIterator.hasNext()) {
seqIterator.next();
}
for (int t = 1; t < oseq.size(); t++) {
final MLDataPair observation = seqIterator.next();
for (int i = 0; i < hmm.getStateCount(); i++) {
computeAlphaStep(hmm, observation, t, i);
}
scale(this.ctFactors, this.alpha, t);
}
}
public ViterbiCalculator(final MLDataSet oseq, final HiddenMarkovModel hmm) {
if (oseq.size() < 1) {
throw new IllegalArgumentException("Must not have empty sequence");
}
this.delta = new double[oseq.size()][hmm.getStateCount()];
this.psy = new int[oseq.size()][hmm.getStateCount()];
this.stateSequence = new int[oseq.size()];
for (int i = 0; i < hmm.getStateCount(); i++) {
this.delta[0][i] =
-Math.log(hmm.getPi(i)) - Math.log(hmm.getStateDistribution(i).probability(oseq.get(0)));
this.psy[0][i] = 0;
}
final Iterator<MLDataPair> oseqIterator = oseq.iterator();
if (oseqIterator.hasNext()) {
oseqIterator.next();
}
int t = 1;
while (oseqIterator.hasNext()) {
final MLDataPair observation = oseqIterator.next();
for (int i = 0; i < hmm.getStateCount(); i++) {
computeStep(hmm, observation, t, i);
}
t++;
}
this.lnProbability = Double.MAX_VALUE;
for (int i = 0; i < hmm.getStateCount(); i++) {
final double thisProbability = this.delta[oseq.size() - 1][i];
if (this.lnProbability > thisProbability) {
this.lnProbability = thisProbability;
this.stateSequence[oseq.size() - 1] = i;
}
}
this.lnProbability = -this.lnProbability;
for (int t2 = oseq.size() - 2; t2 >= 0; t2--) {
this.stateSequence[t2] = this.psy[t2 + 1][this.stateSequence[t2 + 1]];
}
}
