meanestimateforfixedvariancegaussian.java

一个基于PlaceLab的室内和室外的智能导航系统

/*
 * Created on Aug 3, 2004
 *
 */
package org.placelab.particlefilter;

import org.placelab.collections.UnsupportedOperationException;

/**
 * 
 *
 * This class represents a probability distrubution over potential means for a fixed-variance
 * Gaussian.  That is, we're trying to approximate some set of data with a Gaussian, and we know 
 * that Gaussian's variance but not its mean--the distribution represented by 'this' class lets
 * us keep a distrubtion over potential means.  Because the distribution over means happens to
 * be itself Gaussian, there are actually two Gaussians involved whenever you're using this class.
 * 
 * NOTE that even though we refer to the approximating distribution as having a fixed VARIANCE,
 * the constructors actually take in the fixed SIGMA (stdev) value.
 */
public class MeanEstimateForFixedVarianceGaussian extends Gaussian {
	
	private double fixedSigma; // the fixed sigma of the distribution whose mean is 
		                          // being approximated by this distribution
	
	/**
	 * @param mean
	 * @param sigma
	 * 
	 * This constructs a new distribution.  The mean and sigma given can be thought of as 
	 * priors.  Larger sigmas indicate a smaller/weaker prior, and smaller sigmas indicate
	 * a larger/stronger prior; in either case the prior is biased toward the given mean.
	 */
	public MeanEstimateForFixedVarianceGaussian(double mean, double sigma, double fixedSigma) {
		super(mean, sigma);
		this.fixedSigma = fixedSigma;
	}
	public MeanEstimateForFixedVarianceGaussian(Gaussian g, double fixedSigma) {
		this(g.getMean(), g.getSigma(), fixedSigma);
	}
	public MeanEstimateForFixedVarianceGaussian(MeanEstimateForFixedVarianceGaussian other) {
		this(other.getMean(), other.getSigma(), other.getFixedSigma());
	}
	
	/* updateWithSample
	 * ****************
	 * This class represents a distribution over possible mean values for a separate Gaussian
	 * distribution approximating a set of data.  Given a sample from that data, this method
	 * updates this distribution, to give a more accurate estimate of what the mean of the
	 * fixed-variance Gaussian approximating the actual data might look like.
	 * 
	 * Note: assumes the data is 1D.
	 */
	public void updateWithSample(double sample, double weight) {
		
		
		// debugging /*POSLOG*/
		if (sample < 0) {
			System.out.println("Trouble is a brewing!\n  You forgot to use positive log signal strength");
		}
		
		if ((weight < 0.0) || (weight > 1.0)) 
			throw new UnsupportedOperationException("Error: weight "+weight+" is out of range [0.0, 1.0]");
		if (weight == 0) return; // no update to be done if we aren't counting this sample!
		
		double om = getMean(); // om = old mean
		double os = getSigma(); // os = old sigma
		double fs = fixedSigma; // fs = fixed sigma
		double numerator = ((1.0/(os*os))*om) + ((weight/(fs*fs))*sample);
		double denominator = (1.0/(os*os)) + (weight/(fs*fs));
		setMean(numerator/denominator);
		double sigmaSqInv = (1.0/(os*os))+(weight/(fs*fs));
		double sigmaSq = (1.0/sigmaSqInv);
		setSigma(Math.sqrt(sigmaSq));
	}
	public void updateWithSample(double sample) {
		updateWithSample(sample, 1.0);
	}
	
	/* updateWithMultipleSamples
	 * *************************
	 * In actuality there is a better way to update with many samples, since the mean of all the
	 * samples is a sufficient statistic for the update, but I'm too lazy to write it out so
	 * we're going to do it the easy way.
	 */
	public void updateWithMultipleSamples(double samples[]) {
		for(int i=0; i<samples.length; i++) {
			updateWithSample(samples[i]);
		}
	}
	public void updateWithMultipleSamples(double samples[], double weights[]) {
		if (samples.length != weights.length) 
			throw new UnsupportedOperationException("Error: sample and weights arrays are not of the same lengths");
		for(int i=0; i<samples.length; i++) {
			updateWithSample(samples[i], weights[i]);
		}
	}
	
	/* getLikelihoodOfSample
	 * *********************
	 * THIS IS NOT THE SAME AS Gaussian.getHeightAt!!!!!!!!!!
	 * 
	 * Gaussian.getHeightAt will return the height of this Gaussian at the location of the given sample.
	 * This method integrates the distribution of the mean (this gaussian) with the fixed-variance 
	 * gaussian to determine the likelihood of the given sample.  That is, it determines the likelihood
	 * by "averaging" the likelihood of the sample given that the fixed-variance gaussian has mean x, where
	 * x ranges over all possible values represented by 'this' Gaussian.
	 */
	public double getLikelihoodOfSample(double sample) {
		
		// create the fixed-variance gaussian, and place it at the peak of the mean distribution
		Gaussian g = new Gaussian(this.getMean(), this.getFixedSigma());
		
		// convolve the gaussian with 'this' distribution to get the overall sample probability curve
		g.smoothBy(this);
		
		// determine the probability based upon the convolved, final curve
		return g.getHeightAt(sample);
	}
	
	/* Getters/setters
	 * ***************
	 * 
	 *
	 */
	public double getFixedSigma() {
		return fixedSigma;
	}
	public void setFixedSigma(double fs) {
		fixedSigma = fs;
	}

}