root/library/bdm/stat/merger.h @ 1064

/*!
  \file
  \brief Mergers for combination of pdfs
  \author Vaclav Smidl.

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

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

#ifndef MERGER_H
#define MERGER_H


#include "../estim/mixtures.h"
#include "discrete.h"

namespace bdm {
using std::string;

//!Merging methods
enum MERGER_METHOD {ARITHMETIC = 1, GEOMETRIC = 2, LOGNORMAL = 3};

/*!
@brief Base class for general combination of pdfs on discrete support

Mixtures of Gaussian densities are used internally. Switching to other densities should be trivial.

The merged pdfs are expected to be of the form:
 \f[ f(x_i|y_i),  i=1..n \f]
where the resulting merger is a density on \f$ \cup [x_i,y_i] \f$ .
Note that all variables will be joined.

As a result of this feature, each source must be extended to common support
\f[ f(z_i|y_i,x_i) f(x_i|y_i) f(y_i)  i=1..n \f]
where \f$ z_i \f$ accumulate variables that were not in the original source.
These extensions are calculated on-the-fly.

However, these operations can not be performed in general. Hence, this class merges only sources on common support, \f$ y_i={}, z_i={}, \forall i \f$.
For merging of more general cases, use offsprings merger_mix and merger_grid.
*/

class merger_base : public epdf {
protected:
    //! Elements of composition
    Array<shared_ptr<pdf> > pdfs;

    //! Data link for each pdf in pdfs
    Array<datalink_m2e*> dls;

    //! Array of rvs that are not modelled by pdfs at all, \f$ z_i \f$
    Array<RV> rvzs;

    //! Data Links for extension \f$ f(z_i|x_i,y_i) \f$
    Array<datalink_m2e*> zdls;

    //! number of support points
    int Npoints;

    //! number of sources
    int Nsources;

    //! switch of the methoh used for merging
    MERGER_METHOD METHOD;
    //! Default for METHOD
    static const MERGER_METHOD DFLT_METHOD;

    //!Prior on the log-normal merging model
    double beta;
    //! default for beta
    static const double DFLT_beta;

    //! Projection to empirical density (could also be piece-wise linear)
    eEmp eSmp;

    //! debug or not debug
    bool DBG;

    //! debugging file
    it_file* dbg_file;
public:
    //! \name Constructors
    //! @{

    //! Default constructor
    merger_base () : Npoints ( 0 ), Nsources ( 0 ), DBG ( false ), dbg_file ( 0 ) {
    }

    //!Constructor from sources
    merger_base ( const Array<shared_ptr<pdf> > &S );

    //! Function setting the main internal structures
    void set_sources ( const Array<shared_ptr<pdf> > &Sources );

    //! Set support points from rectangular grid
    void set_support ( rectangular_support &Sup );

    //! Set support points from dicrete grid
    void set_support ( discrete_support &Sup ) {
        Npoints = Sup.points();
        eSmp.set_parameters ( Sup._Spoints() );
        eSmp.validate();
    }


    //! set debug file
    void set_debug_file ( const string fname ) {
        if ( DBG ) delete dbg_file;
        dbg_file = new it_file ( fname );
        DBG = ( dbg_file != 0 );
    }

    //! Set internal parameters used in approximation
    void set_method ( MERGER_METHOD MTH = DFLT_METHOD, double beta0 = DFLT_beta ) {
        METHOD = MTH;
        beta = beta0;
    }
    //! Set support points from a pdf by drawing N samples
    void set_support ( const epdf &overall, int N ) {
        eSmp.set_statistics ( overall, N );
        Npoints = N;
    }

    //! Destructor
    virtual ~merger_base() {
        for ( int i = 0; i < Nsources; i++ ) {
            delete dls ( i );
            delete zdls ( i );
        }
        if ( DBG ) delete dbg_file;
    };
    //!@}

    //! \name Mathematical operations
    //!@{

    //!Merge given sources in given points
    virtual void merge ();

    //! Merge log-likelihood values in points using method specified by parameter METHOD
    vec merge_points ( mat &lW );


    //! sample from merged density
//! weight w is a
    vec mean() const;

    mat covariance() const;

    vec variance() const;

    //! Compute log-probability of argument \c val
    virtual double evallog ( const vec &val ) const NOT_IMPLEMENTED(0);

    //! Returns a sample, \f$ x \f$ from density \f$ f_x()\f$
    virtual vec sample() const NOT_IMPLEMENTED(0);

    //!@}

    //! \name Access to attributes
    //! @{

    //! Access function
    eEmp& _Smp() {
        return eSmp;
    }

    //! load from setting
    void from_setting ( const Setting& set );

    void to_setting  (Setting  &set) const ;

    void validate() ;
    //!@}
};
UIREGISTER ( merger_base );
SHAREDPTR ( merger_base );

//! Merger using importance sampling with mixture proposal density
class merger_mix : public merger_base {
protected:
    //!Internal mixture of EF models
    MixEF Mix;
    //!Number of components in a mixture
    int Ncoms;
    //! coefficient of resampling [0,1]
    double effss_coef;
    //! stop after niter iterations
    int stop_niter;

    //! default value for Ncoms
    static const int DFLT_Ncoms;
    //! default value for efss_coef;
    static const double DFLT_effss_coef;

public:
    //!\name Constructors
    //!@{
    merger_mix () : Ncoms ( 0 ), effss_coef ( 0 ), stop_niter ( 0 ) { }

    merger_mix ( const Array<shared_ptr<pdf> > &S ) :
        Ncoms ( 0 ), effss_coef ( 0 ), stop_niter ( 0 ) {
        set_sources ( S );
    }

    //! Set sources and prepare all internal structures
    void set_sources ( const Array<shared_ptr<pdf> > &S ) {
        merger_base::set_sources ( S );
        //Nsources = S.length();
    }

    //! Set internal parameters used in approximation
    void set_parameters ( int Ncoms0 = DFLT_Ncoms, double effss_coef0 = DFLT_effss_coef ) {
        Ncoms = Ncoms0;
        effss_coef = effss_coef0;
    }
    //!@}

    //! \name Mathematical operations
    //!@{

    //!Merge values using mixture approximation
    void merge ();

    //! sample from the approximating mixture
    vec sample () const {
        return Mix.posterior().sample();
    }
    //! loglikelihood computed on mixture models
    double evallog ( const vec &yt ) const {
        vec dtf = ones ( yt.length() + 1 );
        dtf.set_subvector ( 0, yt );
        return Mix.logpred ( dtf, vec(0) );
    }
    //!@}

    //!\name Access functions
    //!@{
//! Access function
    MixEF& _Mix() {
        return Mix;
    }
    //! Access function
    emix* proposal() {
        emix* tmp = Mix.epredictor();
        tmp->set_rv ( rv );
        return tmp;
    }
    //! from_settings
    void from_setting ( const Setting& set );
    void        to_setting  (Setting  &set) const;
    void validate();

    //! @}

};
UIREGISTER ( merger_mix );
SHAREDPTR ( merger_mix );

}

#endif // MER_H