baumwelchscaledlearner.java

java实现的隐马尔科夫模型

/* jahmm package - v0.3.1 */

/*
 *  Copyright (c) 2004, Jean-Marc Francois.
 *
 *  This file is part of Jahmm.
 *  Jahmm is free software; you can redistribute it and/or modify
 *  it under the terms of the GNU General Public License as published by
 *  the Free Software Foundation; either version 2 of the License, or
 *  (at your option) any later version.
 *
 *  Jahmm is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with Jahmm; if not, write to the Free Software
 *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 */

package be.ac.ulg.montefiore.run.jahmm.learn;

import java.util.*;
import be.ac.ulg.montefiore.run.jahmm.*;


/**
 * An implementation of the Baum-Welch learning algorithm.  It uses a
 * scaling mecanism so as to avoid underflows.
 * <p>
 *  For more information on the scaling procedure, read <i>Rabiner</i> and 
 * <i>Juang</i>'s <i>Fundamentals of speech recognition</i> (Prentice Hall,
 * 1993).
 */
public class BaumWelchScaledLearner 
    extends BaumWelchLearner {

    /**
     * Initializes a Baum-Welch algorithm implementation.  This algorithm
     * finds a HMM that models a set of observation sequences.
     *
     * @param nbStates  The number of states the resulting HMM will be made of.
     * @param opdfFactory A class that builds the observation probability
     *                    distributions associated to the states of the HMM.
     * @param obsSeqs A vector of observation sequences.  Each observation
     *                sequences is a vector of
     *                {@link be.ac.ulg.montefiore.run.jahmm.Observation
     *                observations}.
     */
    public BaumWelchScaledLearner(int nbStates, OpdfFactory opdfFactory,
				  Vector obsSeqs) {
	super(nbStates, opdfFactory, obsSeqs);
    }
    
    ForwardBackwardCalculator generateForwardBackwardCalculator(Vector obsSeq,
								Hmm hmm) {
	return new ForwardBackwardScaledCalculator(obsSeq, hmm, 
						   ForwardBackwardCalculator.
						   COMPUTE_ALPHA |
						   ForwardBackwardCalculator.
						   COMPUTE_BETA);
    }
    
    /* Here, the xi (and, thus, gamma) values are not divided by the
       probability of the sequence because this probability might be
       too small and induce an underflow. xi[t][i][j] still can be
       interpreted as P[q_t = i and q_(t+1) = j | obsSeq, hmm] because
       we assume that the scaling factors are such that their product
       is equal to the inverse of the probability of the sequence. */
    double[][][] estimation_xi(Vector obsSeq, ForwardBackwardCalculator fbc, 
				Hmm hmm) {
	double xi[][][] = 
	    new double[obsSeq.size() - 1][hmm.nbStates()][hmm.nbStates()];
	
	for (int t = 0; t < obsSeq.size() - 1; t++)
	    for (int i = 0; i < hmm.nbStates(); i++)
		for (int j = 0; j < hmm.nbStates(); j++)
		    xi[t][i][j] = fbc.alphaElement(t, i) *
			hmm.getAij(i, j) * 
			hmm.getOpdf(j).probability((Observation)
						   obsSeq.elementAt(t + 1)) *
			fbc.betaElement(t + 1, j);
	
	return xi;
    }
}