gibbssampler.h

gibbs

#ifndef GIBBSSAMPLER_H
#define GIBBSSAMPLER_H
#include "InferenceMethod.h"
#include "BayesNet.h"
#include <vector>
#include <list>
#include "Distribution.h"

class GibbsSampler : public InferenceMethod
{
private:
    const BayesNet& model;
    int numChains;
    double convergenceRatio;
    int burnInIters;
    int minIters;
    bool fixedIters;
    vector<Distribution> marginals;
    Distribution jointDistrib;

public:
    GibbsSampler(const BayesNet& bn, int chains, double ratio, int burnIn, 
            int min, bool fixed) 
        : model(bn), numChains(chains), convergenceRatio(ratio),
          burnInIters(burnIn), minIters(min), fixedIters(fixed),
          marginals(model.getNumVars())
    { /* NOP */ }

    virtual ~GibbsSampler() { /* NOP */ }

    virtual void runMarginalInference(const VarSet& evidence);
    virtual void runJointInference(const list<int>& queryVars,
            const VarSet& evidence);
    //virtual Distribution getMarginal(int i) const { return marginals[i]; }
    virtual double singleConditionalLogProb( const list<int>& queryVars,
        const VarSet& evidence, const VarSet& answer) const;

    virtual double getMarginalProb(int i, double value) const {
        return marginals[i][(int)value];
    }

    virtual double getJointProb(const VarConfig& query) const {
        // TODO -- update this to handle continuous vars...
        return jointDistrib[query.getIndex()];
    }

    virtual int getRange(int i) const { return model.getRange(i); }

private:
    int configIndex(VarSet& allVars, list<int>& testIndices) const;
    void burnInChain(VarSet &chain, const VarSet& evidence, int burnInIters) 
        const;
    unsigned int sampleFromDist(const vector<double>& dist) const;
    double testConvergence(vector<vector<double> > summaries, 
            vector<vector<double> > sqSummaries, int n) const;
    double predictIters(vector<vector<double> > summaries, 
            vector<vector<double> > sqSummaries, int n) const;
};
#endif // ndef GIBBSSAMPLER_H