credalNet.h

Go to the documentation of this file.

/**
 *
 *   Copyright (c) 2005-2021 by Pierre-Henri WUILLEMIN(@LIP6) & Christophe GONZALES(@AMU)
 *   info_at_agrum_dot_org
 *
 *  This library is free software: you can redistribute it and/or modify
 *  it under the terms of the GNU Lesser General Public License as published by
 *  the Free Software Foundation, either version 3 of the License, or
 *  (at your option) any later version.
 *
 *  This library is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU Lesser General Public License for more details.
 *
 *  You should have received a copy of the GNU Lesser General Public License
 *  along with this library.  If not, see <http://www.gnu.org/licenses/>.
 *
 */


#ifndef __CREDAL_NET__H__
#define __CREDAL_NET__H__

/**
 * @file
 * @brief Class representing Credal Networks
 * @author Matthieu HOURBRACQ and Pierre-Henri WUILLEMIN(@LIP6)
 */

#include <agrum/agrum.h>

#include <iostream>
#include <vector>

#ifdef HAVE_UNISTD_H
#  include <unistd.h>
#else
#  include <agrum/tools/core/mvsc/unistd.h>
#endif

//#include <sys/wait.h>
#include <algorithm>
#include <cstdlib>
#include <fcntl.h>
#include <fstream>
#include <set>
#include <sstream>
#include <string>
#include <sys/stat.h>

#include <utility>   /// c++11 stuff, like declval ( decltype from prototype without a default constructor )

#include <agrum/tools/core/math/math_utils.h>
#include <agrum/BN/BayesNet.h>
#include <agrum/BN/io/BIF/BIFReader.h>
#include <agrum/BN/io/BIF/BIFWriter.h>
#include <agrum/tools/core/exceptions.h>

#include <agrum/tools/core/math/pow.h>   // custom pow functions with integers, faster implementation
#include <agrum/tools/core/math/rational.h>   // custom decimal to rational

#include <agrum/CN/polytope/LrsWrapper.h>

#include <agrum/tools/core/OMPThreads.h>

// 64 bits for windows (long is 32 bits)
#ifdef _MSC_VER
typedef __int64          int64_t;
typedef unsigned __int64 uint64_t;
#else
#  include <stdint.h>
#endif

namespace gum {
  namespace credal {

    /**
     * @class CredalNet
     * @headerfile credalNet.h <agrum/CN/credalNet.h>
     * @brief Class template representing a Credal Network.
     * @ingroup cn_group
     * @tparam GUM_SCALAR A floating type ( float, GUM_SCALAR, long GUM_SCALAR
     * ... ).
     * @author Matthieu HOURBRACQ and Pierre-Henri WUILLEMIN(@LIP6)
     */
    template < typename GUM_SCALAR >
    class CredalNet {
      public:
      /** @brief NodeType to speed-up computations in some algorithms */
      enum class NodeType : char
      {
        Precise,
        Credal,
        Vacuous,
        Indic
      };

      /// @name Constructors / Destructors
      /// @{

      /**
       * Constructor used to create a CredalNet step by step, i.e. node by node,
       * arc
       * by arc, manually filling potentials.
       */
      CredalNet();

      /**
       * Constructor for interval defined credal network which takes 2 BayesNet
       * file path. One can also provide a single BayesNet in order to perturb it's
       * probability distributions into credal sets according to another BayesNet
       * containing the number of cases, for each node, of each parent
       * instantiation met during learning, i.e. \f$ p(X = 0 \mid pa(X) = j) =
       * N_{pa(X) = j} \f$.
       *
       * @param src_min_num The path to a BayesNet which contains lower
       *probabilities.
       * @param src_max_den The ( optional ) path to a BayesNet which contains
       *upper probabilities.
       */
      CredalNet(const std::string& src_min_num, const std::string& src_max_den = "");

      /**
       * Constructor for interval defined credal network which takes 2 BayesNet.
       * One can also provide a single BayesNet in order to perturb it's
       * probability distributions into credal sets according to another BayesNet
       * containing the number of cases, for each node, of each parent
       * instantiation met during learning, i.e. \f$ p(X = 0 \mid pa(X) = j) =
       * N_{pa(X) = j} \f$.
       *
       * @param src_min_num The BayesNet which contains lower probabilities.
       * @param src_max_den The ( optional ) BayesNet which contains upper
       *probabilities.
       */
      CredalNet(const BayesNet< GUM_SCALAR >& src_min_num,
                const BayesNet< GUM_SCALAR >& src_max_den = BayesNet< GUM_SCALAR >());

      /**
       * Destructor.
       */
      ~CredalNet();

      /// @}

      /// @name Credal network creation
      /// @{

      /**
       * @brief Adds a discrete node into the network
       * @param name The name of the discrete variable to be added
       * @param card The cardinality of the variable
       * @return The \c NodeId of the variable in the network
       */
      NodeId addVariable(const std::string& name, const Size& card);

      /**
       * @brief Adds an arc between two nodes.
       * @param tail The \c NodeId of the tail node
       * @param head The \c NodeId of the head node
       */
      void addArc(const NodeId& tail, const NodeId& head);

      /**
       * @brief %Set the vertices of the credal sets ( all of the conditionals )
       *of a
       *given node
       * @param id The \c NodeId of the node
       * @param cpt The vertices of every credal set ( for each instantiation of
       *the
       *parents )
       *
       * @warning : Does not change the \c BayesNet (s) associated to this
       *credal net
       *!
       *
       * First dimension is instantiation position ( from 0 to K - 1 ).
       * Second is the credal set vertice index
       * Third is the vertex
       */
      void setCPTs(const NodeId&                                                  id,
                   const std::vector< std::vector< std::vector< GUM_SCALAR > > >& cpt);

      /**
       * @brief %Set the vertices of one credal set of a given node ( any
       *instantiation index )
       * @param id The \c NodeId of the node
       * @param entry The index of the instantiation ( from 0 to K - 1 )
       *excluding
       *the given node ( only the parents are used to compute the index of the
       *credal
       *set )
       * @param cpt The vertices of every credal set ( for each instantiation of
       *the
       *parents )
       *
       * Use this with either \c LRSWrapper or \c LpInterface to get the
       *vertices of
       *a credal set represented by linear constraints.
       * @warning : Does not change the \c BayesNet (s) associated to this
       *credal net
       *!
       */
      void
         setCPT(const NodeId& id, Size& entry, const std::vector< std::vector< GUM_SCALAR > >& cpt);

      /**
       * @brief %Set the vertices of one credal set of a given node ( any
       *instantiation )
       * @param id The \c NodeId of the node
       * @param ins The \c Instantiation ( only the parents matter to find the
       *credal
       *set index )
       * @param cpt The vertices of every credal set ( for each instantiation of
       *the
       *parents )
       *
       * Use this with either \c LRSWrapper or \c LpInterface to get the
       *vertices of a credal set represented by linear constraints.
       * @warning : Does not change the \c BayesNet (s) associated to this
       *credal net
       *!
       *
       * @note we forget the master ref of \c ins to check variable order in the
       *instantiation ( to get index ), therefor we pass it by value
       */
      void setCPT(const NodeId&                                   id,
                  Instantiation                                   ins,
                  const std::vector< std::vector< GUM_SCALAR > >& cpt);

      /**
       * @brief %Set the interval constraints of the credal sets of a given node
       *(all instantiations )
       * @param id The \c NodeId of the node
       * @param lower The lower value for each probability in correct order
       * @param upper The upper value for each probability in correct order
       *
       * You need to call intervalToCredal when done filling all constraints.
       *
       * @warning : DOES change the \c BayesNet (s) associated to this credal
       *net !
       * @note we forget the master ref of \c ins to check variable order in the
       *instantiation ( to get index ), therefore we pass it by value
       */
      void fillConstraints(const NodeId&                    id,
                           const std::vector< GUM_SCALAR >& lower,
                           const std::vector< GUM_SCALAR >& upper);

      /**
       * @brief %Set the interval constraints of a credal set of a given node (
       *from an instantiation index )
       * @param id The \c NodeId of the node
       * @param entry The index of the instantiation excluding the given node (
       * only the parents are used to compute the index of the credal set )
       * @param lower The lower value for each probability in correct order
       * @param upper The upper value for each probability in correct order
       *
       * You need to call intervalToCredal when done filling all constraints.
       *
       * @warning : DOES change the \c BayesNet (s) associated to this credal
       *net !
       * @note we forget the master ref of \c ins to check variable order in the
       *instantiation ( to get index ), therefore we pass it by value
       */
      void fillConstraint(const NodeId&                    id,
                          const Idx&                       entry,
                          const std::vector< GUM_SCALAR >& lower,
                          const std::vector< GUM_SCALAR >& upper);

      /**
       * @brief %Set the interval constraints of a credal sets of a given node (
       *from
       *an instantiation )
       * @param id The \c NodeId of the node
       * @param ins The \c Instantiation
       * @param lower The lower value for each probability in correct order
       * @param upper The upper value for each probability in correct order
       *
       * You need to call intervalToCredal when done filling all constraints.
       *
       * @warning : DOES change the \c BayesNet (s) associated to this credal
       *net !
       * @note we forget the master ref of \c ins to check variable order in the
       *instantiation ( to get index ), therefor we pass it by value
       */
      void fillConstraint(const NodeId&                    id,
                          Instantiation                    ins,
                          const std::vector< GUM_SCALAR >& lower,
                          const std::vector< GUM_SCALAR >& upper);

      /**
       * @brief %Get an \c Instantiation from a node id, usefull to fill the
       * constraints of the network
       * @param id The \c NodeId we want an instantiation from
       * @return The instantiation
       */
      Instantiation instantiation(const NodeId& id);

      /**
       * @brief %Get the cardinality of a node
       * @param id The \c NodeId of the node
       * @return The cardinality of the node
       */
      Size domainSize(const NodeId& id);

      /// @}

      /// @name Public manipulation methods
      /// @{

      /**
       * Perturbates the BayesNet provided as input for this CredalNet by
       *generating intervals instead of point probabilities and then computes each
       *vertex of each credal set.
       *
       * The perturbations are done according to the number of cases met for
       *each node and each of it's parent instantiation, i.e. \f$ \epsilon =
       *\beta^{ln(N_{pa(X) = j} + 1)} \f$ is the imprecision introduced which
       *leads to \f$ \underline{p}(X = i \mid pa(X) = j) = (1 - \epsilon) p(X = i
       *\mid pa(X) = j) \f$ and \f$ \overline{p}(X = i \mid pa(X) = j)
       *=\underline{p}(X =i \mid pa(X) = j) + \epsilon \f$. Use this method when
       *using a single BayesNet storing counts of events with \c oneNet set to \c
       *TRUE or when using two BayesNet, one with lower probabilities and one with
       *upper probabilities, with \c oneNet set to \c FALSE.
       *
       * @param beta The beta used to perturbate the network. \f$ 0 \leq \beta
       *\leq 1 \f$.
       * @param oneNet Boolean used as a flag. %Set to \c TRUE if one BayesNet
       * if provided with counts, to \c FALSE if two BayesNet are provided; one
       *with probabilities (the lower net) and one with denominators over the first
       *modalities (the upper net).
       * @param keepZeroes Boolean used as a flag as whether or not -
       *respectively \c TRUE or \c FALSE - we keep zeroes as zeroes. Default is \c
       *FALSE, i.e. zeroes are not kept.
       */
      void bnToCredal(const GUM_SCALAR beta, const bool oneNet, const bool keepZeroes = false);

      /**
       * @deprecated Use intervalToCredal ( lrsWrapper with no input / output
       *files needed ).
       *
       * Computes the vertices of each credal set according to their interval
       *definition (uses lrs).
       *
       * Use this method when using a single BayesNet storing counts of events.
       */
      void intervalToCredalWithFiles();

      /**
       * Computes the vertices of each credal set according to their interval
       *definition (uses lrs).
       *
       * Use this method when using two BayesNet, one with lower probabilities
       *and one with upper probabilities.
       */
      void intervalToCredal();

      /**
       * Normalize counts of a BayesNet storing counts of each events such that
       *no probability is 0.
       *
       * Use this method when using a single BayesNet storing counts of events.
       * Lagrange normalization. This call is irreversible and modify counts
       *stored by \c  _src_bn_.
       *
       * Doest not performs computations of the parameters but keeps normalized
       *counts of events only. Call \c idmLearning to compute the probabilities
       *(with any parameter value).
       */
      void lagrangeNormalization();

      /**
       * Learns parameters from a BayesNet storing counts of events.
       *
       * Use this method when using a single BayesNet storing counts of events.
       * IDM model if \c s > 0, standard point probability if \c s = 0 (default
       *value if none precised).
       * @param s The IDM parameter.
       * @param keepZeroes Boolean used as a flag as whether or not -
       *respectively \c TRUE or \c FALSE - we keep zeroes as zeroes. Default is \c
       *FALSE, i.e. zeroes are not kept.
       */
      void idmLearning(const Idx s = 0, const bool keepZeroes = false);

      /**
       * Approximate binarization. Each bit has a lower and upper probability
       * which is the lowest - resp. highest - over all vertices of the credal set.
       * Enlarge the original credal sets and may induce huge imprecision.
       *
       * @warning Enlarge the original credal sets and therefore induce huge
       *imprecision by propagation. Not recommended, use MCSampling or something
       *else
       *instead.
       */
      void approximatedBinarization();

      /// @}

      // other utility member methods
      // PH void saveCNet ( const std::string &cn_path ) const;
      // PH void loadCNet ( const std::string &src_cn_path );

      /**
       * If this CredalNet was built over a perturbed BayesNet, one can save the
       *intervals as two BayesNet.
       *
       * @param min_path The path to save the BayesNet which contains the lower
       *probabilities of each node X, i.e. \f$ \underline{p}(X = i \mid pa(X) =
       *j)
       *\f$.
       * @param max_path The path to save the BayesNet which contains the upper
       *probabilities of each node X, i.e. \f$ \overline{p}(X = i \mid pa(X) =
       *j)
       *\f$.
       */
      void saveBNsMinMax(const std::string& min_path, const std::string& max_path) const;

      // PH void vacants ( int &result ) const;
      // PH void notVacants ( int &result ) const;
      // PH void averageVertices ( GUM_SCALAR &result ) const;

      /**
       * @return Returns the string representation of this CredalNet, i.e. it's
       * CPTs
       * (which also represent arcs).
       */
      std::string toString() const;
      // PH void toPNG ( const std::string &png_path ) const;

      /**
       * Used with binary networks to speed-up L2U inference.
       *
       * Store the lower and upper probabilities of each node X over the "true"
       *modality, i.e. respectively \f$ \underline{p}(X = 1 \mid pa(X) = j) \f$
       *and
       *\f$ \overline{p}(X = 1 \mid pa(X) = j) \f$.
       */
      void computeBinaryCPTMinMax();   // REDO THIS IN PRIVATE !!!

      /// @name Getters and setters
      /// @{
      /**
       * @return Returns a constant reference to the original BayesNet (used as
       * a
       * DAG, it's CPTs does not matter).
       */
      const BayesNet< GUM_SCALAR >& src_bn() const;

      /**
       * @return Returs a constant reference to the actual BayesNet (used as a
       * DAG,
       * it's CPTs does not matter).
       */
      const BayesNet< GUM_SCALAR >& current_bn() const;

      /**
       * @return Returns a constant reference to the ( up-to-date ) CredalNet
       * CPTs.
       */
      const NodeProperty< std::vector< std::vector< std::vector< GUM_SCALAR > > > >&
         credalNet_currentCpt() const;

      /**
       * @return Returns a constant reference to the ( up-to-date ) CredalNet
       * CPTs.
       */
      const NodeProperty< std::vector< std::vector< std::vector< GUM_SCALAR > > > >&
         credalNet_srcCpt() const;

      /**
       * @param id The constant reference to the choosen NodeId
       * @return Returns the type of the choosen node in the ( up-to-date )
       * CredalNet
       *  _current_bn_ if any,  _src_bn_ otherwise.
       */
      NodeType currentNodeType(const NodeId& id) const;

      /**
       * @param id The constant reference to the choosen NodeId
       * @return Returns the type of the choosen node in the ( up-to-date )
       * CredalNet
       * in  _src_bn_.
       */
      NodeType nodeType(const NodeId& id) const;

      /**
       * @return Returns a constant reference to the lowest perturbation of the
       * BayesNet provided as input for this CredalNet.
       */
      const GUM_SCALAR& epsilonMin() const;

      /**
       * @return Returns a constant reference to the highest perturbation of the
       * BayesNet provided as input for this CredalNet.
       */
      const GUM_SCALAR& epsilonMax() const;

      /**
       * @return Returns a constant reference to the average perturbation of the
       * BayesNet provided as input for this CredalNet.
       */
      const GUM_SCALAR& epsilonMean() const;

      /**
       * @return Returns \c TRUE if this CredalNet is separately and interval
       * specified, \c FALSE otherwise.
       */
      const bool isSeparatelySpecified() const;

      /**
       * @return Returns \c TRUE if this CredalNet has called
       * computeBinaryCPTMinMax() to speed-up inference with binary networks and
       * L2U. This needs to be reworked as it is too easy to forget to call it and
       * it can't be called within the inference engine (constness).
       */
      const bool hasComputedBinaryCPTMinMax() const;

      /**
       * Used with binary networks to speed-up L2U inference.
       *
       * @return Returns a constant reference to the lower probabilities of each
       *node
       *X over the "true" modality, i.e. \f$ \underline{p}(X = 1 \mid pa(X) = j)
       *\f$.
       */
      const std::vector< std::vector< GUM_SCALAR > >& get_binaryCPT_min() const;

      /**
       * Used with binary networks to speed-up L2U inference.
       *
       * @return Returns a constant reference to the upper probabilities of each
       *node
       *X over the "true" modality, i.e. \f$ \overline{p}(X = 1 \mid pa(X) = j)
       *\f$.
       */
      const std::vector< std::vector< GUM_SCALAR > >& get_binaryCPT_max() const;

      // PH const std::vector< std::vector< NodeId > > & var_bits() const;

      /// @}

      protected:
      private:
      /** @brief 1e6 by default, used by  _fracC_ as precision. */
      GUM_SCALAR _precisionC_;   // = 1e6;
      /** @brief 5 by default, used by  _fracC_ as number of decimals. */
      GUM_SCALAR _deltaC_;   // = 5;

      /** @brief The lowest perturbation of the BayesNet provided as input for
       * this
       * CredalNet. */
      GUM_SCALAR _epsilonMin_;
      /** @brief The highest perturbation of the BayesNet provided as input for
       * this
       * CredalNet. */
      GUM_SCALAR _epsilonMax_;
      /** @brief The average perturbation of the BayesNet provided as input for
       * this
       * CredalNet. */
      GUM_SCALAR _epsilonMoy_;

      /** @brief Value under which a decimal number is considered to be zero
       * when
       * computing redundant vertices. */
      GUM_SCALAR _epsRedund_;   //= 1e-6;

      /** @brief Value under which a decimal number is considered to be zero
       * when
       * using  _farey_. */
      GUM_SCALAR _epsF_;   // = 1e-6;
      /** @brief Highest possible denominator allowed when using  _farey_. A
       * value too
       * high may lead to lrs being unable to find vertices. */
      GUM_SCALAR _denMax_;   // = 1e6; // beware LRS

      /** @brief Precision used by  _frac_. */
      GUM_SCALAR _precision_;   // = 1e6; // beware LRS

      /** @brief \c TRUE if this CredalNet is separately and interval specified,
       * \c
       * FALSE otherwise. */
      bool _separatelySpecified_;

      /** @brief Original BayesNet (used as a DAG). Is never modified. */
      BayesNet< GUM_SCALAR > _src_bn_;

      /** @brief BayesNet used to store lower probabilities. */
      BayesNet< GUM_SCALAR > _src_bn_min_;
      /** @brief BayesNet used to store upper probabilities. */
      BayesNet< GUM_SCALAR > _src_bn_max_;

      /** @brief Up-to-date BayesNet (used as a DAG). */
      BayesNet< GUM_SCALAR >* _current_bn_;   // = nullptr;

      /** @brief This CredalNet original CPTs. */
      NodeProperty< std::vector< std::vector< std::vector< GUM_SCALAR > > > > _credalNet_src_cpt_;

      /** @brief This CredalNet up-to-date CPTs. */
      NodeProperty< std::vector< std::vector< std::vector< GUM_SCALAR > > > >*
         _credalNet_current_cpt_;   // =  nullptr;

      /** @deprecated @brief Corresponding bits of each variable. */
      NodeProperty< std::vector< NodeId > > _var_bits_;

      /** @brief The NodeType of each node from the ORIGINAL network. */
      NodeProperty< NodeType > _original_nodeType_;
      /** @brief The NodeType of each node from the up-to-date network. */
      NodeProperty< NodeType >* _current_nodeType_;   // = nullptr;

      /** @brief Used by L2U, to know if lower and upper probabilities over the
       * second modality has been stored in order to speed-up the algorithm. */
      bool _hasComputedBinaryCPTMinMax_;
      /**
       * @brief Used with binary networks to speed-up L2U inference. Store the
       * lower
       * probabilities of each node X over the "true" modality, i.e. \f$
       * \underline{p}(X = 1 \mid pa(X) = j) \f$.
       */
      typename std::vector< std::vector< GUM_SCALAR > > _binCptMin_;

      /**
       * @brief Used with binary networks to speed-up L2U inference. Store the
       * upper
       * probabilities of each node X over the "true" modality, i.e. \f$
       * \overline{p}(X = 1 \mid pa(X) = j) \f$.
       */
      typename std::vector< std::vector< GUM_SCALAR > > _binCptMax_;

      /** @brief %Set the NodeType of each node */
      void _sort_varType_();

      /**
       * @deprecated
       * @param var_cpt The reference to a node CPT which may need a Decision
       * Node.
       * @return Returns the cardinality of the Decision Node.
       */
      int _find_dNode_card_(
         const std::vector< std::vector< std::vector< GUM_SCALAR > > >& var_cpt) const;

      /**
       * Computes the vertices of each credal set according to their interval
       *definition (does not use lrs). Only works with credal sets defined such
       *that
       *when one modality reach it's upper probability, all others are at their
       *lowest.
       *
       * Called by bnToCredal and idmLearning.
       */
      void _intervalToCredal_();

      /** @brief Initialize private constant variables after the Constructor has
       * been
       * called */
      void _initParams_();

      /** @brief Initialize private BayesNet variables after the Constructor has
       * been
       * called */
      void _initCNNets_(const std::string& src_min_num, const std::string& src_max_den);

      /** @brief Initialize private BayesNet variables after the Constructor has
       * been
       * called */
      void _initCNNets_(const BayesNet< GUM_SCALAR >& src_min_num,
                        const BayesNet< GUM_SCALAR >& src_max_den);

      /**
       * @deprecated
       * @warning May be useless since the BayesNet copy constructor seems to
       *now
       *work well (parent order is preserved).
       *
       * Copy the up-to-date BayesNet associated with this CredalNet. Since all
       *we
       *care about is the DAG, only arcs are copied. Because the order with
       *which
       *arcs are created is important, the function iterates over the CPTs
       *variables
       *to be sure parent order stays the same from a net to it's copy.
       *
       * @param bn_dest The reference to the new copy
       */
      void _bnCopy_(BayesNet< GUM_SCALAR >& bn_dest);

      // void  _H2Vcdd_ ( const std::vector< std::vector< GUM_SCALAR > > & h_rep,
      // std::vector< std::vector< GUM_SCALAR > > & v_rep ) const;
      /**
       * @deprecated one should use the LrsWrapper class
       * Computes the V-representation of a credal set, i.e. it's vertices, from
       *it's
       *H-representation, i.e. the hyper-plan inequalities. Uses lrs.
       *
       * @param h_rep A constant reference to the H-representation of a credal
       *set.
       * @param v_rep A reference to the V-representation of the same credal
       *set.
       */
      void _H2Vlrs_(const std::vector< std::vector< GUM_SCALAR > >& h_rep,
                    std::vector< std::vector< GUM_SCALAR > >&       v_rep) const;

    };   // CredalNet


#ifndef GUM_NO_EXTERN_TEMPLATE_CLASS
    extern template class CredalNet< double >;
#endif

  }   // namespace credal
}   // namespace gum

#include <agrum/CN/credalNet_tpl.h>

#endif   //  __CREDAL_NET__H__