root/library/bdm/estim/mixtures.h @ 943

/*!
  \file
  \brief Bayesian Filtering for mixtures of exponential family (EF) members
  \author Vaclav Smidl.

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

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

#ifndef MIXTURES_H
#define MIXTURES_H


#include "../math/functions.h"
#include "../stat/exp_family.h"
#include "../stat/emix.h"

namespace bdm {

//! enum switch for internal approximation used in MixEF
enum MixEF_METHOD { EM = 0, QB = 1};

/*!
* \brief Mixture of Exponential Family Densities

An approximate estimation method for models with latent discrete variable, such as
mixture models of the following kind:
\f[
f(y_t|\psi_t, \Theta) = \sum_{i=1}^{n} w_i f(y_t|\psi_t, \theta_i)
\f]
where  \f$\psi\f$ is a known function of past outputs, \f$w=[w_1,\ldots,w_n]\f$ are component weights, and component parameters \f$\theta_i\f$ are assumed to be mutually independent. \f$\Theta\f$ is an aggregation af all component parameters and weights, i.e. \f$\Theta = [\theta_1,\ldots,\theta_n,w]\f$.

The characteristic feature of this model is that if the exact values of the latent variable were known, estimation of the parameters can be handled by a single model. For example, for the case of mixture models, posterior density for each component parameters would be a BayesianModel from Exponential Family.

This class uses EM-style type algorithms for estimation of its parameters. Under this simplification, the posterior density is a product of exponential family members, hence under EM-style approximate estimation this class itself belongs to the exponential family.

TODO: Extend BM to use rvc.
*/
class MixEF: public BMEF {
protected:
        //!Number of components
        int n;
        //! Models for Components of \f$\theta_i\f$
        Array<BMEF*> Coms;
        //! Statistics for weights
        multiBM weights;
        //aux
        friend class eprod_mix;
        //!Posterior on component parameters
        class eprod_mix: public eprod_base {
                protected:
                const MixEF &mix; // pointer to parent.n
                public:
                eprod_mix(const MixEF &m):mix(m){}
                const epdf* factor(int i) const {return (i==(mix.n-1)) ? &mix.weights.posterior() : &mix.Coms(i)->posterior();}
                const int no_factors()const {return mix.n+1;}
        } est;
        ////!indices of component rvc in common rvc

        //! Flag for a method that is used in the inference
        MixEF_METHOD method;

public:
        //! Full constructor
        MixEF ( const Array<BMEF*> &Coms0, const vec &alpha0 ) :
                        BMEF ( ), n ( Coms0.length() ), Coms ( n ),
                        weights (), est(*this), method ( QB ) {
                for ( int i = 0; i < n; i++ ) {
                        Coms ( i ) = ( BMEF* ) Coms0 ( i )->_copy();
                }
                weights.set_parameters(alpha0);
                weights.validate();
        }

        //! Constructor of empty mixture
        MixEF () :
                        BMEF ( ), n ( 0 ), Coms ( 0 ),
                        weights (), est(*this), method ( QB ) {
        }
        //! Copy constructor
        MixEF ( const MixEF &M2 ) : BMEF ( ), n ( M2.n ), Coms ( n ),
                        weights ( M2.weights ), est(*this), method ( M2.method ) {
                for ( int i = 0; i < n; i++ ) {
                        Coms ( i ) = (BMEF*) M2.Coms ( i )->_copy();
                }
        }

        //! Initializing the mixture by a random pick of centroids from data
        //! \param Com0 Initial component - necessary to determine its type.
        //! \param Data Data on which the initialization will be done
        //! \param c Initial number of components, default=5
        void init ( BMEF* Com0, const mat &Data, const int c = 5 );
        //Destructor
        //! Recursive EM-like algorithm (QB-variant), see Karny et. al, 2006
        void bayes ( const vec &yt, const vec &cond );
        //! EM algorithm
        void bayes ( const mat &yt, const vec &cond );
        //! batch weighted Bayes rule
        void bayes_batch ( const mat &yt, const mat &cond, const vec &wData );
        double logpred ( const vec &yt ) const;
        //! return correctly typed posterior (covariant return)
        const eprod_mix& posterior() const {
                return est;
        }

        emix* epredictor(const vec &cond=vec()) const;
        //! Flatten the density as if it was not estimated from the data
        void flatten ( const BMEF* M2 );
        //! Access function
        BMEF* _Coms ( int i ) {
                return Coms ( i );
        }

        //!Set which method is to be used
        void set_method ( MixEF_METHOD M ) {
                method = M;
        }

        void to_setting ( Setting &set ) const {
                BMEF::to_setting( set );
                UI::save ( Coms, set, "Coms" );
                UI::save ( &weights, set, "weights" );
        }
};
UIREGISTER ( MixEF );

}
#endif // MIXTURES_H