/*! \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 Abstract common class for mergers Merging is defined as a combination of information from \c source pdfs into a single pdf called the \c merger. \f[ f_{merged}(x) \leftarrow {f_1(x),f_2(x),f_3(x)} \f] In the basic form, the random variables of all source pdfs must be identical, e.g. \f$ x \f$ from the example above. Extension of the merger to a more demanding scenarios with fragmental sources is available in offspring bdm::merger_mix. This abstract class defines only the interface to using merger: operation merge(), and operation merger() which returns the result. */ class MergerBase: public root{ public: //! weights of the sources -- no restrictions vec w; //! source pdfs -- conditioning allowed Array > sources; //! merge virtual void merge() NOT_IMPLEMENTED_VOID; //! //! check if all epdfs are on the same support virtual void set_sources(const Array > &A) { int n=A.length(); bdm_assert(n>0,"merger has no sources to merge"); // check if weights are ok if (w.length()!=n) { w = ones(n)/n; } // check compatibility of sources -- no types are needed int dim0 = A(0)->dimension(); for (int i=0; i(A(i).get()); if (si){ sources(i) = shared_ptr(const_cast(si)); } bdm_assert(sources(i)->dimension()==dim0, "Merger: Incompatible dimensions of sources"); if (sources(i)->isnamed()){ //check if rv match } } } //! return the result of merging virtual const epdf& merger()=0; }; //! \brief Merger into normal density, max entropy approximator for 2 moments (mean+variance) template class ENormMerger: public MergerBase{ protected: //!internal epdf enorm iepdf; public: ENormMerger():method(GEOMETRIC){}; MERGER_METHOD method; void merge(){ int n = sources.length(); int dim = sources(0)->dimension(); sq_T Var(zeros(dim)); vec mea=zeros(dim); // go through all sources epdf * si; for (int i=0; i(sources(i).get()); // transform pdf into epdf if (!si) {bdm_error("Can't merger this type of pdf: " + sources(i)->to_string());} sq_T Ci = si->covariance(); sq_T wCi = Ci; wCi*=w(i); vec mi = si->mean(); switch (method) { case ARITHMETIC:{ Var += wCi; Var.opupdt(mi,w(i)); // + mean * mean' mea += w(i)*mi; break; } case GEOMETRIC:{ Var += wCi; sq_T iCi; Ci.inv(iCi); mea += iCi.to_mat()*w(i)*mi; break; } default: bdm_error("Method not implemneted"); } } // post pocessing switch (method) { case ARITHMETIC: iepdf._R()=Var; iepdf._R().opupdt(mea,-1.0); // Var - mean * mean' iepdf._mu()=mea; break; case GEOMETRIC: iepdf._R()=Var; iepdf._mu() = Var.to_mat()*mea; // mean=sqrt(Var)*mea break; default: bdm_error("Method not implemneted"); } } //! access function const epdf& merger(){return iepdf;} }; /*! @brief Base class (interface) for 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$. For merging of more general cases, use offsprings merger_mix and merger_grid. \TODO use sources for extensions */ class MergerDiscrete : public MergerBase { protected: //! partial sources Array > part_sources; //! Data link for each pdf in pdfs Array dls; //! Array of rvs that are not modelled by pdfs at all, \f$ z_i \f$ Array rvzs; //! Data Links for extension \f$ f(z_i|x_i,y_i) \f$ Array 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 MergerDiscrete () : Npoints ( 0 ), Nsources ( 0 ), DBG ( false ), dbg_file ( 0 ) { } //!Constructor from sources MergerDiscrete ( const Array > &S ); //! Function setting the main internal structures void set_sources ( const Array > &PartSources ); //! 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 = GEOMETRIC, double beta0 = 1.2 ) { 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 ~MergerDiscrete() { 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 void merge (); //! Merge log-likelihood values in points using method specified by parameter METHOD vec merge_points ( mat &lW ); //!@} //! \name Access to attributes //! @{ //! Access function eEmp& _Smp() { return eSmp; } eEmp & merger() {return eSmp;} /*! Create object from the following structure \code class = 'MergerDiscrete'; method = 'arithmetic' or 'geometric' or 'lognormal'; % name of the model used for merging --- compulsory only for lognormal merging model --- beta = x; % scalar prior on the log-normal merging model --- optional fields --- dbg_file = '...'; % name of the file to store debug informations --- inherited fields --- bdm::MergerBase::from_setting \endcode */ void from_setting ( const Setting& set ); void to_setting (Setting &set) const ; void validate() ; //!@} }; UIREGISTER ( MergerDiscrete ); SHAREDPTR ( MergerDiscrete ); //! \brief Merger using importance sampling with mixture proposal density class merger_mix : public MergerDiscrete { 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 > &S ) : Ncoms ( 0 ), effss_coef ( 0 ), stop_niter ( 0 ) { set_sources ( S ); } //! Set sources and prepare all internal structures void set_sources ( const Array > &S ) { MergerDiscrete::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 ( merger()._rv() ); return tmp; } /*! Create object from the following structure \code class = 'merger_mix'; --- optional fields --- ncoms = []; % number of components in a mixture effss_coef = []; % coefficient of resampling from interval [0,1] stop_niter = []; % number of interations --- inherited fields --- bdm::merger_base::from_setting \endcode If the optional fields are not given, they will be filled as follows: \code ncoms = 10; effss_coef = 0.9; stop_niter = 10; \endcode */ void from_setting ( const Setting& set ); void to_setting (Setting &set) const; void validate(); // const emix &merger(){return *Mpred;} //! @} }; UIREGISTER ( merger_mix ); SHAREDPTR ( merger_mix ); } #endif // MER_H