baumwelchscaled.java

java实现的隐马尔科夫模型

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

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


/* When the sequence generated by MarkovGenerator mg has a length > 1000,
   the unscaled Baum-Welch learner can't give any result because of
   underflows. */

public class BaumWelchScaled {

    /* A simple example of the HMM package */
    /* See http://www.run.montefiore.ulg.ac.be/~francois/software/hmm/ */
    static public void main(String[] argv) 
	throws java.io.IOException {

	Hmm hmm = buildHmm();
	
	/* Observation sequence generation */
	MarkovGenerator mg = new MarkovGenerator(hmm);
	Vector sequence = mg.observationSequence(100000);
	Vector sequences = new Vector();
	sequences.add(sequence);

	/* Baum-Welch learning */
	BaumWelchScaledLearner bwl =
	    new BaumWelchScaledLearner(2, new OpdfIntegerFactory(2), sequences);

	Hmm learntHmm = bwl.iterate(buildInitHmm());
	(new HmmIntegerDrawer()).write(learntHmm, "learntHmm.dot");
    }

    
    /* The HMM this example is based on */
    static Hmm buildHmm() {
       	Hmm hmm = new Hmm(2, new OpdfIntegerFactory(2));
	
	hmm.setPi(0, 0.95);
	hmm.setPi(1, 0.05);

	hmm.setOpdf(0, new OpdfInteger(new double[] {0.95, 0.05}));
	hmm.setOpdf(1, new OpdfInteger(new double[] {0.2, 0.8}));

	hmm.setAij(0, 1, 0.05);
	hmm.setAij(0, 0, 0.95);
	hmm.setAij(1, 0, 0.1);
	hmm.setAij(1, 1, 0.9);

	return hmm;
    }

    
    /* Initial guess for the Baum-Welch algorithm */
    static Hmm buildInitHmm() {
	Hmm hmm = new Hmm(2, new OpdfIntegerFactory(2));
	
	hmm.setPi(0, 0.90);
	hmm.setPi(1, 0.10);

	hmm.setOpdf(0, new OpdfInteger(new double[] {0.8, 0.2}));
	hmm.setOpdf(1, new OpdfInteger(new double[] {0.1, 0.9}));

	hmm.setAij(0, 1, 0.2);
	hmm.setAij(0, 0, 0.8);
	hmm.setAij(1, 0, 0.15);
	hmm.setAij(1, 1, 0.85);

	return hmm;
	}
}