BayesNet_tpl.h

Go to the documentation of this file.

/***************************************************************************
 *   Copyright (C) 2005 by Pierre-Henri WUILLEMIN et Christophe GONZALES   *
 *   {prenom.nom}_at_lip6.fr                                               *
 *                                                                         *
 *   This program is free software; you can redistribute it and/or modify  *
 *   it under the terms of the GNU General Public License as published by  *
 *   the Free Software Foundation; either version 2 of the License, or     *
 *   (at your option) any later version.                                   *
 *                                                                         *
 *   This program 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 General Public License for more details.                          *
 *                                                                         *
 *   You should have received a copy of the GNU General Public License     *
 *   along with this program; if not, write to the                         *
 *   Free Software Foundation, Inc.,                                       *
 *   59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.             *
 ***************************************************************************/
#include <limits>
#include <set>

#include <agrum/BN/BayesNet.h>
#include <agrum/BN/IBayesNet.h>
#include <agrum/BN/IBayesNet.h>

#include <agrum/variables/rangeVariable.h>
#include <agrum/variables/labelizedVariable.h>

#include <agrum/multidim/aggregators/amplitude.h>
#include <agrum/multidim/aggregators/and.h>
#include <agrum/multidim/aggregators/count.h>
#include <agrum/multidim/aggregators/exists.h>
#include <agrum/multidim/aggregators/forall.h>
#include <agrum/multidim/aggregators/max.h>
#include <agrum/multidim/aggregators/median.h>
#include <agrum/multidim/aggregators/min.h>
#include <agrum/multidim/aggregators/or.h>

#include <agrum/multidim/ICIModels/multiDimNoisyAND.h>
#include <agrum/multidim/ICIModels/multiDimNoisyORCompound.h>
#include <agrum/multidim/ICIModels/multiDimNoisyORNet.h>

#include <agrum/multidim/ICIModels/multiDimLogit.h>

#include <agrum/BN/generator/simpleCPTGenerator.h>
#include <agrum/core/utils_string.h>

namespace gum {
  template < typename GUM_SCALAR >
  NodeId build_node(gum::BayesNet< GUM_SCALAR >& bn,
                    std::string                  node,
                    gum::Size                    domainSize) {
    std::string                name = node;
    auto                       ds = domainSize;
    std::vector< std::string > labels;

    // node like "n[5]"
    auto posBrack = node.find('[');
    if (posBrack != std::string::npos) {
      if (*(node.rbegin()) != ']')
        name = node;   // a name with  '[' inside but no ']' at the end
      else {
        name = node.substr(0, posBrack);
        ds = static_cast< Size >(
           std::stoi(node.substr(posBrack + 1, node.size() - posBrack - 2)));
      }
    }

    // node like "n{one|two|three}"
    posBrack = node.find('{');
    if (posBrack != std::string::npos) {
      if (*(node.rbegin()) != '}')
        name = node;   // a name with  '{' inside but no '}' at the end
      else {
        name = node.substr(0, posBrack);
        labels = split(node.substr(posBrack + 1, node.size() - posBrack - 2), "|");
        if (labels.size() < 2) {
          GUM_ERROR(InvalidArgument, "Not enough labels in node " << node);
        }
        if (!hasUniqueElts(labels)) {
          GUM_ERROR(InvalidArgument, "Duplicate labels in node " << node);
        }
        ds = static_cast< Size >(labels.size());
      }
    }

    if (ds == 0) {
      GUM_ERROR(InvalidArgument, "No value for variable " << name << ".");
    } else if (ds == 1) {
      GUM_ERROR(InvalidArgument,
                "Only one value for variable " << name
                                               << " (2 at least are needed).");
    }

    // now we add the node in the BN
    NodeId idVar;
    try {
      idVar = bn.idFromName(name);
    } catch (gum::NotFound&) {
      if (labels.empty()) {
        idVar = bn.add(RangeVariable(name, name, 0, ds - 1));
      } else {
        auto l = LabelizedVariable(name, name, 0);
        for (const auto& label : labels) {
          l.addLabel(label);
        }
        idVar = bn.add(l);
      }
    }

    return idVar;
  }

  template < typename GUM_SCALAR >
  BayesNet< GUM_SCALAR >
     BayesNet< GUM_SCALAR >::fastPrototype(const std::string& dotlike,
                                           Size               domainSize) {
    gum::BayesNet< GUM_SCALAR > bn;


    for (const auto& chaine : split(dotlike, ";")) {
      NodeId lastId = 0;
      bool   notfirst = false;
      for (const auto& souschaine : split(chaine, "->")) {
        bool forward = true;
        for (const auto& node : split(souschaine, "<-")) {
          auto idVar = build_node(bn, node, domainSize);
          if (notfirst) {
            if (forward) {
              bn.addArc(lastId, idVar);
              forward = false;
            } else {
              bn.addArc(idVar, lastId);
            }
          } else {
            notfirst = true;
            forward = false;
          }
          lastId = idVar;
        }
      }
    }
    bn.generateCPTs();
    bn.setProperty("name", dotlike);
    return bn;
  }

  template < typename GUM_SCALAR >
  INLINE BayesNet< GUM_SCALAR >::BayesNet() : IBayesNet< GUM_SCALAR >() {
    GUM_CONSTRUCTOR(BayesNet);
  }

  template < typename GUM_SCALAR >
  INLINE BayesNet< GUM_SCALAR >::BayesNet(std::string name) :
      IBayesNet< GUM_SCALAR >(name) {
    GUM_CONSTRUCTOR(BayesNet);
  }

  template < typename GUM_SCALAR >
  BayesNet< GUM_SCALAR >::BayesNet(const BayesNet< GUM_SCALAR >& source) :
      IBayesNet< GUM_SCALAR >(source), __varMap(source.__varMap) {
    GUM_CONS_CPY(BayesNet);

    __copyPotentials(source);
  }

  template < typename GUM_SCALAR >
  BayesNet< GUM_SCALAR >& BayesNet< GUM_SCALAR >::
                          operator=(const BayesNet< GUM_SCALAR >& source) {
    if (this != &source) {
      IBayesNet< GUM_SCALAR >::operator=(source);
      __varMap = source.__varMap;

      __clearPotentials();
      __copyPotentials(source);
    }

    return *this;
  }

  template < typename GUM_SCALAR >
  BayesNet< GUM_SCALAR >::~BayesNet() {
    GUM_DESTRUCTOR(BayesNet);
    for (const auto p : __probaMap) {
      delete p.second;
    }
  }

  template < typename GUM_SCALAR >
  INLINE const DiscreteVariable&
               BayesNet< GUM_SCALAR >::variable(NodeId id) const {
    return __varMap.get(id);
  }

  template < typename GUM_SCALAR >
  INLINE void
     BayesNet< GUM_SCALAR >::changeVariableName(NodeId             id,
                                                const std::string& new_name) {
    __varMap.changeName(id, new_name);
  }

  template < typename GUM_SCALAR >
  INLINE void BayesNet< GUM_SCALAR >::changeVariableLabel(
     NodeId id, const std::string& old_label, const std::string& new_label) {
    if (variable(id).varType() != VarType::Labelized) {
      GUM_ERROR(NotFound, "Variable " << id << " is not a LabelizedVariable.");
    }
    LabelizedVariable* var = dynamic_cast< LabelizedVariable* >(
       const_cast< DiscreteVariable* >(&variable(id)));

    var->changeLabel(var->posLabel(old_label), new_label);
  }


  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::nodeId(const DiscreteVariable& var) const {
    return __varMap.get(var);
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::add(const DiscreteVariable& var) {
    auto   ptr = new MultiDimArray< GUM_SCALAR >();
    NodeId res = 0;

    try {
      res = add(var, ptr);

    } catch (Exception&) {
      delete ptr;
      throw;
    }

    return res;
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::add(const std::string& name,
                                            unsigned int       nbrmod) {
    if (nbrmod < 2) {
      GUM_ERROR(OperationNotAllowed,
                "Variable " << name << "needs more than " << nbrmod
                            << " modalities");
    }

    LabelizedVariable v(name, name, nbrmod);
    return add(v);
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::add(
     const DiscreteVariable& var, MultiDimImplementation< GUM_SCALAR >* aContent) {
    NodeId proposedId = dag().nextNodeId();
    NodeId res = 0;

    res = add(var, aContent, proposedId);

    return res;
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::add(const DiscreteVariable& var,
                                            NodeId                  id) {
    auto   ptr = new MultiDimArray< GUM_SCALAR >();
    NodeId res = 0;

    try {
      res = add(var, ptr, id);

    } catch (Exception&) {
      delete ptr;
      throw;
    }

    return res;
  }

  template < typename GUM_SCALAR >
  NodeId
     BayesNet< GUM_SCALAR >::add(const DiscreteVariable&               var,
                                 MultiDimImplementation< GUM_SCALAR >* aContent,
                                 NodeId                                id) {
    __varMap.insert(id, var);
    this->_dag.addNodeWithId(id);

    auto cpt = new Potential< GUM_SCALAR >(aContent);
    (*cpt) << variable(id);
    __probaMap.insert(id, cpt);
    return id;
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::idFromName(const std::string& name) const {
    return __varMap.idFromName(name);
  }

  template < typename GUM_SCALAR >
  INLINE const DiscreteVariable&
               BayesNet< GUM_SCALAR >::variableFromName(const std::string& name) const {
    return __varMap.variableFromName(name);
  }

  template < typename GUM_SCALAR >
  INLINE const Potential< GUM_SCALAR >&
               BayesNet< GUM_SCALAR >::cpt(NodeId varId) const {
    return *(__probaMap[varId]);
  }

  template < typename GUM_SCALAR >
  INLINE const VariableNodeMap& BayesNet< GUM_SCALAR >::variableNodeMap() const {
    return __varMap;
  }

  template < typename GUM_SCALAR >
  INLINE void BayesNet< GUM_SCALAR >::erase(const DiscreteVariable& var) {
    erase(__varMap.get(var));
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::erase(NodeId varId) {
    if (__varMap.exists(varId)) {
      // Reduce the variable child's CPT
      const NodeSet& children = this->children(varId);

      for (const auto c : children) {
        __probaMap[c]->erase(variable(varId));
      }

      delete __probaMap[varId];

      __probaMap.erase(varId);
      __varMap.erase(varId);
      this->_dag.eraseNode(varId);
    }
  }

  template < typename GUM_SCALAR >
  INLINE void BayesNet< GUM_SCALAR >::addArc(NodeId tail, NodeId head) {
    this->_dag.addArc(tail, head);
    // Add parent in the child's CPT
    (*(__probaMap[head])) << variable(tail);
  }

  template < typename GUM_SCALAR >
  INLINE void BayesNet< GUM_SCALAR >::eraseArc(const Arc& arc) {
    if (__varMap.exists(arc.tail()) && __varMap.exists(arc.head())) {
      NodeId head = arc.head(), tail = arc.tail();
      this->_dag.eraseArc(arc);
      // Remove parent froms child's CPT
      (*(__probaMap[head])) >> variable(tail);
    }
  }

  template < typename GUM_SCALAR >
  INLINE void BayesNet< GUM_SCALAR >::eraseArc(NodeId tail, NodeId head) {
    eraseArc(Arc(tail, head));
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::reverseArc(const Arc& arc) {
    // check that the arc exsists
    if (!__varMap.exists(arc.tail()) || !__varMap.exists(arc.head())
        || !dag().existsArc(arc)) {
      GUM_ERROR(InvalidArc, "a nonexisting arc cannot be reversed");
    }

    NodeId tail = arc.tail(), head = arc.head();

    // check that the reversal does not induce a cycle
    try {
      DAG d = dag();
      d.eraseArc(arc);
      d.addArc(head, tail);
    } catch (Exception&) {
      GUM_ERROR(InvalidArc, "this arc reversal would induce a directed cycle");
    }

    // with the same notations as Shachter (1986), "evaluating influence
    // diagrams",
    // p.878, we shall first compute the product of probabilities:
    // pi_j^old (x_j | x_c^old(j) ) * pi_i^old (x_i | x_c^old(i) )
    Potential< GUM_SCALAR > prod{cpt(tail) * cpt(head)};

    // modify the topology of the graph: add to tail all the parents of head
    // and add to head all the parents of tail
    beginTopologyTransformation();
    NodeSet new_parents;
    for (const auto node : this->parents(tail))
      new_parents.insert(node);
    for (const auto node : this->parents(head))
      new_parents.insert(node);
    // remove arc (head, tail)
    eraseArc(arc);

    // add the necessary arcs to the tail
    for (const auto p : new_parents) {
      if ((p != tail) && !dag().existsArc(p, tail)) { addArc(p, tail); }
    }

    addArc(head, tail);
    // add the necessary arcs to the head
    new_parents.erase(tail);

    for (const auto p : new_parents) {
      if ((p != head) && !dag().existsArc(p, head)) { addArc(p, head); }
    }

    endTopologyTransformation();

    // update the conditional distributions of head and tail
    Set< const DiscreteVariable* > del_vars;
    del_vars << &(variable(tail));
    Potential< GUM_SCALAR > new_cpt_head = prod.margSumOut(del_vars);
    auto& cpt_head = const_cast< Potential< GUM_SCALAR >& >(cpt(head));
    cpt_head = std::move(new_cpt_head);

    Potential< GUM_SCALAR > new_cpt_tail{prod / cpt_head};
    auto& cpt_tail = const_cast< Potential< GUM_SCALAR >& >(cpt(tail));
    cpt_tail = std::move(new_cpt_tail);
  }

  template < typename GUM_SCALAR >
  INLINE void BayesNet< GUM_SCALAR >::reverseArc(NodeId tail, NodeId head) {
    reverseArc(Arc(tail, head));
  }


  //==============================================
  // Aggregators
  //=============================================
  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addAMPLITUDE(const DiscreteVariable& var) {
    return add(var, new aggregator::Amplitude< GUM_SCALAR >());
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addAND(const DiscreteVariable& var) {
    if (var.domainSize() > 2) GUM_ERROR(SizeError, "an AND has to be boolean");

    return add(var, new aggregator::And< GUM_SCALAR >());
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addCOUNT(const DiscreteVariable& var,
                                                 Idx                     value) {
    return add(var, new aggregator::Count< GUM_SCALAR >(value));
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addEXISTS(const DiscreteVariable& var,
                                                  Idx                     value) {
    if (var.domainSize() > 2) GUM_ERROR(SizeError, "an EXISTS has to be boolean");

    return add(var, new aggregator::Exists< GUM_SCALAR >(value));
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addFORALL(const DiscreteVariable& var,
                                                  Idx                     value) {
    if (var.domainSize() > 2) GUM_ERROR(SizeError, "an EXISTS has to be boolean");

    return add(var, new aggregator::Forall< GUM_SCALAR >(value));
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addMAX(const DiscreteVariable& var) {
    return add(var, new aggregator::Max< GUM_SCALAR >());
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addMEDIAN(const DiscreteVariable& var) {
    return add(var, new aggregator::Median< GUM_SCALAR >());
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addMIN(const DiscreteVariable& var) {
    return add(var, new aggregator::Min< GUM_SCALAR >());
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addOR(const DiscreteVariable& var) {
    if (var.domainSize() > 2) GUM_ERROR(SizeError, "an OR has to be boolean");

    return add(var, new aggregator::Or< GUM_SCALAR >());
  }


  //================================
  // ICIModels
  //================================
  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addNoisyOR(const DiscreteVariable& var,
                                                   GUM_SCALAR external_weight) {
    return addNoisyORCompound(var, external_weight);
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addNoisyORCompound(
     const DiscreteVariable& var, GUM_SCALAR external_weight) {
    return add(var, new MultiDimNoisyORCompound< GUM_SCALAR >(external_weight));
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addNoisyORNet(const DiscreteVariable& var,
                                                      GUM_SCALAR external_weight) {
    return add(var, new MultiDimNoisyORNet< GUM_SCALAR >(external_weight));
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addNoisyAND(const DiscreteVariable& var,
                                                    GUM_SCALAR external_weight) {
    return add(var, new MultiDimNoisyAND< GUM_SCALAR >(external_weight));
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addLogit(const DiscreteVariable& var,
                                                 GUM_SCALAR external_weight) {
    return add(var, new MultiDimLogit< GUM_SCALAR >(external_weight));
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addNoisyOR(const DiscreteVariable& var,
                                                   GUM_SCALAR external_weight,
                                                   NodeId     id) {
    return addNoisyORCompound(var, external_weight, id);
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addNoisyAND(const DiscreteVariable& var,
                                                    GUM_SCALAR external_weight,
                                                    NodeId     id) {
    return add(var, new MultiDimNoisyAND< GUM_SCALAR >(external_weight), id);
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addLogit(const DiscreteVariable& var,
                                                 GUM_SCALAR external_weight,
                                                 NodeId     id) {
    return add(var, new MultiDimLogit< GUM_SCALAR >(external_weight), id);
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addNoisyORCompound(
     const DiscreteVariable& var, GUM_SCALAR external_weight, NodeId id) {
    return add(
       var, new MultiDimNoisyORCompound< GUM_SCALAR >(external_weight), id);
  }

  template < typename GUM_SCALAR >
  INLINE NodeId BayesNet< GUM_SCALAR >::addNoisyORNet(const DiscreteVariable& var,
                                                      GUM_SCALAR external_weight,
                                                      NodeId     id) {
    return add(var, new MultiDimNoisyORNet< GUM_SCALAR >(external_weight), id);
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::addWeightedArc(NodeId     tail,
                                              NodeId     head,
                                              GUM_SCALAR causalWeight) {
    auto* CImodel =
       dynamic_cast< const MultiDimICIModel< GUM_SCALAR >* >(cpt(head).content());

    if (CImodel != 0) {
      // or is OK
      addArc(tail, head);

      CImodel->causalWeight(variable(tail), causalWeight);
    } else {
      GUM_ERROR(InvalidArc,
                "Head variable (" << variable(head).name()
                                  << ") is not a CIModel variable !");
    }
  }

  template < typename GUM_SCALAR >
  INLINE std::ostream& operator<<(std::ostream&                 output,
                                  const BayesNet< GUM_SCALAR >& bn) {
    output << bn.toString();
    return output;
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::beginTopologyTransformation() {
    for (const auto node : nodes())
      __probaMap[node]->beginMultipleChanges();
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::endTopologyTransformation() {
    for (const auto node : nodes())
      __probaMap[node]->endMultipleChanges();
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::__clearPotentials() {
    // Removing previous potentials
    for (const auto& elt : __probaMap) {
      delete elt.second;
    }

    __probaMap.clear();
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::__copyPotentials(
     const BayesNet< GUM_SCALAR >& source) {
    // Copying potentials

    for (const auto src : source.__probaMap) {
      // First we build the node's CPT
      Potential< GUM_SCALAR >* copy_array = new Potential< GUM_SCALAR >();
      copy_array->beginMultipleChanges();
      for (gum::Idx i = 0; i < src.second->nbrDim(); i++) {
        (*copy_array) << variableFromName(src.second->variable(i).name());
      }
      copy_array->endMultipleChanges();
      copy_array->copyFrom(*(src.second));

      // We add the CPT to the CPT's hashmap
      __probaMap.insert(src.first, copy_array);
    }
  }

  template < typename GUM_SCALAR >
  INLINE void BayesNet< GUM_SCALAR >::generateCPTs() const {
    for (const auto node : nodes())
      generateCPT(node);
  }
  template < typename GUM_SCALAR >
  INLINE void BayesNet< GUM_SCALAR >::generateCPT(NodeId node) const {
    SimpleCPTGenerator< GUM_SCALAR > generator;

    generator.generateCPT(cpt(node).pos(variable(node)), cpt(node));
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::changePotential(NodeId                   id,
                                               Potential< GUM_SCALAR >* newPot) {
    if (cpt(id).nbrDim() != newPot->nbrDim()) {
      GUM_ERROR(OperationNotAllowed,
                "cannot exchange potentials with different "
                "dimensions for variable with id "
                   << id);
    }

    for (Idx i = 0; i < cpt(id).nbrDim(); i++) {
      if (&cpt(id).variable(i) != &(newPot->variable(i))) {
        GUM_ERROR(OperationNotAllowed,
                  "cannot exchange potentials because, for variable with id "
                     << id << ", dimension " << i << " differs. ");
      }
    }

    _unsafeChangePotential(id, newPot);
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::_unsafeChangePotential(
     NodeId id, Potential< GUM_SCALAR >* newPot) {
    delete __probaMap[id];
    __probaMap[id] = newPot;
  }

  template < typename GUM_SCALAR >
  void BayesNet< GUM_SCALAR >::changePotential(const std::string&       name,
                                               Potential< GUM_SCALAR >* newPot) {
    changePotential(idFromName(name), newPot);
  }

} /* namespace gum */