root/bdm/estim/libPF.h @ 283

/*!
  \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;

        //! which resampling method will be used
        RESAMPLING_METHOD resmethod;

        //! \name Options
        //!@{

        //! Log all samples
        bool opt_L_smp;
        //! Log all samples
        bool opt_L_wei;
        //!@}

public:
        //! \name Constructors
        //!@{
        PF ( ) :est(), _w ( est._w() ),_samples ( est._samples() ), opt_L_smp ( false ), opt_L_wei ( false ) {LIDs.set_size ( 5 );};
        /*      PF ( mpdf *par0, mpdf *obs0, epdf *epdf0, int n0 ) :
                                est ( ),_w ( est._w() ),_samples ( est._samples() ),opt_L_smp(false), opt_L_wei(false)
                { set_parameters ( par0,obs0,n0 ); set_statistics ( ones ( n0 ),epdf0 ); };*/
        void set_parameters ( mpdf *par0, mpdf *obs0, int n0, RESAMPLING_METHOD rm=SYSTEMATIC )
        { par = par0; obs=obs0; n=n0; resmethod= rm;};
        void set_statistics ( const vec w0, epdf *epdf0 ) {est.set_statistics ( w0,epdf0 );};
        //!@}
        //! Set posterior density by sampling from epdf0
//      void set_est ( const epdf &epdf0 );
        void set_options ( const string &opt ) {
                BM::set_options(opt);
                opt_L_wei= ( opt.find ( "logweights" ) !=string::npos );
                opt_L_smp= ( opt.find ( "logsamples" ) !=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 {
        Array<BM_T*> BMs;

        //! 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() ) {
                };
                //! read statistics from MPF
                void read_statistics ( Array<BM_T*> &A ) {
                        dim = E.dimension() +A ( 0 )->posterior().dimension();
                        for ( int i=0; i<_w.length() ;i++ ) {Coms ( i ) = A ( i )->_e();}
                }
                //! needed in resampling
                void set_elements ( int &i, double wi, const epdf* ep )
                {_w ( i ) =wi; Coms ( i ) =ep;};

                void set_parameters ( int n ) {
                        E.set_parameters ( n, false );
                        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 ) );
                }
                void qbounds ( vec &lb, vec &ub, double perc=0.95 ) const {
                        //bounds on particles
                        vec lbp;
                        vec ubp;
                        E.qbounds ( lbp,ubp );

                        //bounds on Components
                        int dimC=Coms ( 0 )->dimension();
                        int j;
                        // temporary
                        vec lbc(dimC);
                        vec ubc(dimC);
                        // minima and maxima
                        vec Lbc(dimC);
                        vec Ubc(dimC);
                        Lbc = std::numeric_limits<double>::infinity();
                        Ubc = -std::numeric_limits<double>::infinity();

                        for ( int i=0;i<_w.length();i++ ) {
                                // check Coms
                                Coms ( i )->qbounds ( lbc,ubc );
                                for ( j=0;j<dimC; j++ ) {
                                        if ( lbc ( j ) <Lbc ( j ) ) {Lbc ( j ) =lbc ( j );}
                                        if ( ubc ( j ) >Ubc ( j ) ) {Ubc ( j ) =ubc ( j );}
                                }
                        }
                        lb=concat(lbp,Lbc);
                        ub=concat(ubp,Ubc);
                }

                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 () : PF (), jest ( est ) {};
        void set_parameters ( mpdf *par0, mpdf *obs0, int n0, RESAMPLING_METHOD rm=SYSTEMATIC ) {
                PF::set_parameters ( par0, obs0, n0, rm );
                jest.set_parameters ( n0 );//duplication of rm
                BMs.set_length ( n0 );
        }
        void set_statistics ( epdf *epdf0, const BM_T* BMcond0 ) {

                PF::set_statistics ( ones ( n ) /n, epdf0 );
                // copy
                for ( int i=0;i<n;i++ ) { BMs ( i ) = new BM_T ( *BMcond0 ); BMs ( i )->condition ( _samples ( i ) );}

                jest.read_statistics ( BMs );
                //options
        };

        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 = ( opt.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 ( resmethod );
                // 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