root/bdm/estim/libPF.h @ 281

/*!
  \file
  \brief Bayesian Filtering using stochastic sampling (Particle Filters)
  \author Vaclav Smidl.

  -----------------------------------
  BDM++ - C++ library for Bayesian Decision Making under Uncertainty

  Using IT++ for numerical operations
  -----------------------------------
*/

#ifndef PF_H
#define PF_H


#include "../stat/libEF.h"

namespace bdm {

/*!
* \brief Trivial particle filter with proposal density equal to parameter evolution model.

Posterior density is represented by a weighted empirical density (\c eEmp ).
*/

class PF : public BM {
protected:
        //!number of particles;
        int n;
        //!posterior density
        eEmp est;
        //! pointer into \c eEmp
        vec &_w;
        //! pointer into \c eEmp
        Array<vec> &_samples;
        //! Parameter evolution model
        mpdf *par;
        //! Observation model
        mpdf *obs;

        //! \name Options
        //!@{

        //! Log resampling times
        bool opt_L_res;
        //! Log all samples
        bool opt_L_smp;
        //!@}

public:
        //! \name Constructors
        //!@{
        PF ( ) :est(), _w ( est._w() ),_samples ( est._samples() ) {};
        PF ( mpdf *par0, mpdf *obs0, epdf *epdf0, int n0 ) :
                        est ( ),_w ( est._w() ),_samples ( est._samples() )
        { set_parameters ( par0,obs0,n0 ); set_statistics ( ones ( n0 ),epdf0 ); };
        void set_parameters ( mpdf *par0, mpdf *obs0, int n0 )
        { par = par0; obs=obs0; n=n0; est.set_n ( n );};
        void set_statistics ( const vec w0, epdf *epdf0 ) {est.set_parameters ( w0,epdf0 );};
        //!@}
        //! Set posterior density by sampling from epdf0
        void set_est ( const epdf &epdf0 );
        void set_options(const string &opt){
                opt_L_res= ( s.find ( "logres" ) !=string::npos );
                opt_L_smp= ( s.find ( "logsmp" ) !=string::npos );              
        }
        void bayes ( const vec &dt );
        //!access function
        vec* __w() {return &_w;}
};

/*!
\brief Marginalized Particle filter

Trivial version: proposal = parameter evolution, observation model is not used. (it is assumed to be part of BM).
*/

template<class BM_T>

class MPF : public PF {
        BM_T* Bms[10000];

        //! internal class for MPDF providing composition of eEmp with external components

class mpfepdf : public epdf  {
        protected:
                eEmp &E;
                vec &_w;
                Array<const epdf*> Coms;
        public:
                mpfepdf ( eEmp &E0 ) :
                                epdf ( ), E ( E0 ),  _w ( E._w() ),
                                Coms ( _w.length() ) {
                };
                void set_elements ( int &i, double wi, const epdf* ep )
                {_w ( i ) =wi; Coms ( i ) =ep;};

                void set_n ( int n ) {E.set_n ( n ); Coms.set_length ( n );}
                vec mean() const {
                        // ugly
                        vec pom=zeros ( Coms ( 0 )->dimension() );
                        for ( int i=0; i<_w.length(); i++ ) {pom += Coms ( i )->mean() * _w ( i );}
                        return concat ( E.mean(),pom );
                }
                vec variance() const {
                        // ugly
                        vec pom=zeros ( Coms ( 0 )->dimension() );
                        vec pom2=zeros ( Coms ( 0 )->dimension() );
                        for ( int i=0; i<_w.length(); i++ ) {
                                pom += Coms ( i )->mean() * _w ( i );
                                pom2 += ( Coms ( i )->variance() + pow ( Coms ( i )->mean(),2 ) ) * _w ( i );
                        }
                        return concat ( E.variance(),pom2-pow ( pom,2 ) );
                }

                vec sample() const {it_error ( "Not implemented" );return 0;}

                double evallog ( const vec &val ) const {it_error ( "not implemented" ); return 0.0;}
        };

        //! Density joining PF.est with conditional parts
        mpfepdf jest;

        //! Log means of BMs
        bool opt_L_mea;
        
public:
        //! Default constructor.
        MPF ( mpdf *par0, mpdf *obs0, int n, const BM_T &BMcond0 ) : PF (), jest ( est ) {
                PF::set_parameters ( par0,obs0,n );
                jest.set_n ( n );
                //
                //TODO test if rv and BMcond.rv are compatible.
//              rv.add ( rvlin );
                //

                if ( n>10000 ) {it_error ( "increase 10000 here!" );}

                for ( int i=0;i<n;i++ ) {
                        Bms[i] = new BM_T ( BMcond0 ); //copy constructor
                        const epdf& pom=Bms[i]->posterior();
                        jest.set_elements ( i,1.0/n,&pom );
                }
        };

        ~MPF() {
        }

        void bayes ( const vec &dt );
        const epdf& posterior() const {return jest;}
        const epdf* _e() const {return &jest;} //Fixme: is it useful?
        //! Set postrior of \c rvc to samples from epdf0. Statistics of Bms are not re-computed! Use only for initialization!
        void set_est ( const epdf& epdf0 ) {
                PF::set_est ( epdf0 );  // sample params in condition
                // copy conditions to BMs

                for ( int i=0;i<n;i++ ) {Bms[i]->condition ( _samples ( i ) );}
        }
        void set_options(const string &opt){
                PF:set_options(opt);
                opt_L_mea = ( s.find ( "logmeans" ) !=string::npos );
        }
        
        //!Access function
        BM* _BM ( int i ) {return Bms[i];}
};

template<class BM_T>
void MPF<BM_T>::bayes ( const vec &dt ) {
        int i;
        vec lls ( n );
        vec llsP ( n );
        ivec ind;
        double mlls=-std::numeric_limits<double>::infinity();

#pragma omp parallel for
        for ( i=0;i<n;i++ ) {
                //generate new samples from paramater evolution model;
                _samples ( i ) = par->samplecond ( _samples ( i ) );
                llsP ( i ) = par->_e()->evallog ( _samples ( i ) );
                Bms[i]->condition ( _samples ( i ) );
                Bms[i]->bayes ( dt );
                lls ( i ) = Bms[i]->_ll(); // lls above is also in proposal her must be lls(i) =, not +=!!
                if ( lls ( i ) >mlls ) mlls=lls ( i ); //find maximum likelihood (for numerical stability)
        }

        double sum_w=0.0;
        // compute weights
#pragma omp parallel for
        for ( i=0;i<n;i++ ) {
                _w ( i ) *= exp ( lls ( i ) - mlls ); // multiply w by likelihood
                sum_w+=_w ( i );
        }

        if ( sum_w  >0.0 ) {
                _w /=sum_w; //?
        }
        else {
                cout<<"sum(w)==0"<<endl;
        }


        double eff = 1.0/ ( _w*_w );
        if ( eff < ( 0.3*n ) ) {
                ind = est.resample();
                // Resample Bms!

#pragma omp parallel for
                for ( i=0;i<n;i++ ) {
                        if ( ind ( i ) !=i ) {//replace the current Bm by a new one
                                //fixme this would require new assignment operator
                                // *Bms[i] = *Bms[ind ( i ) ];

                                // poor-man's solution: replicate constructor here
                                // copied from MPF::MPF
                                delete Bms[i];
                                Bms[i] = new BM_T ( *Bms[ind ( i ) ] ); //copy constructor
                                const epdf& pom=Bms[i]->posterior();
                                jest.set_elements ( i,1.0/n,&pom );
                        }
                };
                cout << '.';
        }
}

}
#endif // KF_H