SVED_tpl.h

Go to the documentation of this file.

/***************************************************************************
 *   Copyright (C) 2005 by Christophe GONZALES and Pierre-Henri WUILLEMIN  *
 *   {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 <agrum/PRM/inference/SVED.h>

namespace gum {
  namespace prm {

    template < typename GUM_SCALAR >
    SVED< GUM_SCALAR >::~SVED() {
      GUM_DESTRUCTOR(SVED);

      for (const auto& elt : __elim_orders)
        delete elt.second;

      if (__class_elim_order != nullptr) delete __class_elim_order;
    }

    template < typename GUM_SCALAR >
    void
       SVED< GUM_SCALAR >::__eliminateNodes(const PRMInstance< GUM_SCALAR >* query,
                                            NodeId                           node,
                                            BucketSet&                       pool,
                                            BucketSet& trash) {
      Set< const PRMInstance< GUM_SCALAR >* > ignore;
      ignore.insert(query);
      // Extracting required attributes and slotchains
      Set< NodeId >& attr_set = __getAttrSet(query);
      Set< NodeId >& sc_set = __getSCSet(query);
      // Downward elimination
      List< const PRMInstance< GUM_SCALAR >* > elim_list;

      for (const auto attr : attr_set) {
        try {
          for (auto iter = query->getRefAttr(attr).begin();
               iter != query->getRefAttr(attr).end();
               ++iter)
            if ((!ignore.exists(iter->first)) && (__bb.exists(iter->first)))
              __eliminateNodesDownward(
                 query, iter->first, pool, trash, elim_list, ignore);
        } catch (NotFound&) {
          // Ok
        }
      }

      // Eliminating all nodes in query instance, except query
      InstanceBayesNet< GUM_SCALAR > bn(*query);
      DefaultTriangulation           t(&(bn.moralGraph()), &(bn.modalities()));
      std::vector< const DiscreteVariable* > elim_order;

      if (this->hasEvidence(query)) __insertEvidence(query, pool);

      for (const auto attr : attr_set)
        pool.insert(
           &(const_cast< Potential< GUM_SCALAR >& >(query->get(attr).cpf())));

      for (size_t idx = 0; idx < t.eliminationOrder().size(); ++idx) {
        if (t.eliminationOrder()[idx] != node) {
          auto        var_id = t.eliminationOrder()[idx];
          const auto& var = bn.variable(var_id);
          elim_order.push_back(&var);
        }
      }

      eliminateNodes(elim_order, pool, trash);
      // Eliminating instance in elim_list
      List< const PRMInstance< GUM_SCALAR >* > tmp_list;
      __reduceElimList(query, elim_list, tmp_list, ignore, pool, trash);

      while (!elim_list.empty()) {
        if (__checkElimOrder(query, elim_list.front())) {
          if ((!ignore.exists(elim_list.front()))
              && (__bb.exists(elim_list.front())))
            __eliminateNodesDownward(
               query, elim_list.front(), pool, trash, elim_list, ignore);
        } else if (__bb.exists(elim_list.front())) {
          tmp_list.insert(elim_list.front());
        }

        elim_list.popFront();
      }

      // Upward elimination
      for (const auto chain : sc_set)
        for (const auto parent : query->getInstances(chain))
          if ((!ignore.exists(parent)) && (__bb.exists(*parent)))
            __eliminateNodesUpward(parent, pool, trash, tmp_list, ignore);
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::__eliminateNodesDownward(
       const PRMInstance< GUM_SCALAR >*          from,
       const PRMInstance< GUM_SCALAR >*          i,
       BucketSet&                                pool,
       BucketSet&                                trash,
       List< const PRMInstance< GUM_SCALAR >* >& elim_list,
       Set< const PRMInstance< GUM_SCALAR >* >&  ignore) {
      ignore.insert(i);
      // Extracting required attributes and slotchains
      Set< NodeId >& attr_set = __getAttrSet(i);
      Set< NodeId >& sc_set = __getSCSet(i);
      // Calling elimination over child instance
      List< const PRMInstance< GUM_SCALAR >* > my_list;

      for (const auto attr : attr_set) {
        try {
          for (auto iter = i->getRefAttr(attr).begin();
               iter != i->getRefAttr(attr).end();
               ++iter)
            if ((!ignore.exists(iter->first)) && (__bb.exists(iter->first)))
              __eliminateNodesDownward(
                 i, iter->first, pool, trash, my_list, ignore);
        } catch (NotFound&) {
          // Ok
        }
      }

      // Eliminating all nodes in current instance
      if (this->hasEvidence(i)) {
        __eliminateNodesWithEvidence(i, pool, trash);
      } else {
        __insertLiftedNodes(i, pool, trash);

        for (const auto agg : i->type().aggregates())
          if (__bb.requisiteNodes(i).exists(agg->id()))
            pool.insert(__getAggPotential(i, agg));

        try {
          InstanceBayesNet< GUM_SCALAR >         bn(*i);
          std::vector< const DiscreteVariable* > elim_order;

          for (auto node : __getElimOrder(i->type())) {
            const auto& var = bn.variable(node);
            elim_order.push_back(&var);
          }

          eliminateNodes(elim_order, pool, trash);
        } catch (NotFound&) {
          // Raised if there is no inner nodes to eliminate
        }
      }

      // Calling elimination over child's parents
      while (!my_list.empty()) {
        if (__checkElimOrder(i, my_list.front())) {
          if ((!ignore.exists(my_list.front())) && (__bb.exists(my_list.front())))
            __eliminateNodesDownward(
               i, my_list.front(), pool, trash, my_list, ignore);
        } else if (__bb.exists(my_list.front())) {
          elim_list.insert(my_list.front());
        }

        my_list.popFront();
      }

      // Adding parents instance to elim_list
      for (const auto chain : sc_set)
        for (const auto parent : i->getInstances(chain))
          if ((!ignore.exists(parent)) && __bb.exists(parent) && (parent != from))
            elim_list.insert(parent);
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::__eliminateNodesUpward(
       const PRMInstance< GUM_SCALAR >*          i,
       BucketSet&                                pool,
       BucketSet&                                trash,
       List< const PRMInstance< GUM_SCALAR >* >& elim_list,
       Set< const PRMInstance< GUM_SCALAR >* >&  ignore) {
      ignore.insert(i);
      // Extracting required attributes and slotchains
      Set< NodeId >& attr_set = __getAttrSet(i);
      Set< NodeId >& sc_set = __getSCSet(i);

      // Downward elimination
      for (const auto attr : attr_set) {
        try {
          for (auto iter = i->getRefAttr(attr).begin();
               iter != i->getRefAttr(attr).end();
               ++iter)
            if ((!ignore.exists(iter->first)) && (__bb.exists(iter->first)))
              __eliminateNodesDownward(
                 i, iter->first, pool, trash, elim_list, ignore);
        } catch (NotFound&) {
          // Ok
        }
      }

      // Eliminating all nodes in i instance
      if (this->hasEvidence(i)) {
        __eliminateNodesWithEvidence(i, pool, trash);
      } else {
        __insertLiftedNodes(i, pool, trash);

        for (const auto agg : i->type().aggregates())
          if (__bb.requisiteNodes(i).exists(agg->id()))
            pool.insert(__getAggPotential(i, agg));

        try {
          InstanceBayesNet< GUM_SCALAR >         bn(*i);
          std::vector< const DiscreteVariable* > elim_order;

          for (auto node : __getElimOrder(i->type())) {
            const auto& var = bn.variable(node);
            elim_order.push_back(&var);
          }
          eliminateNodes(elim_order, pool, trash);
        } catch (NotFound&) {
          // Raised if there is no inner nodes to eliminate
        }
      }

      // Eliminating instance in elim_list
      List< const PRMInstance< GUM_SCALAR >* > tmp_list;

      while (!elim_list.empty()) {
        if (__checkElimOrder(i, elim_list.front())) {
          if ((!ignore.exists(elim_list.front()))
              && (__bb.exists(elim_list.front())))
            __eliminateNodesDownward(
               i, elim_list.front(), pool, trash, elim_list, ignore);
        } else if (__bb.exists(elim_list.front())) {
          ignore.insert(elim_list.front());
        }

        elim_list.popFront();
      }

      // Upward elimination
      for (const auto chain : sc_set)
        for (const auto parent : i->getInstances(chain))
          if ((!ignore.exists(parent)) && (__bb.exists(parent)))
            __eliminateNodesUpward(parent, pool, trash, tmp_list, ignore);
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::__eliminateNodesWithEvidence(
       const PRMInstance< GUM_SCALAR >* i, BucketSet& pool, BucketSet& trash) {
      // Adding required evidences
      for (const auto& elt : this->evidence(i))
        if (__bb.requisiteNodes(i).exists(elt.first))
          pool.insert(const_cast< Potential< GUM_SCALAR >* >(elt.second));

      // Adding potentials and eliminating the remaining nodes
      for (const auto& a : *i)
        if (__bb.requisiteNodes(i).exists(a.first))
          pool.insert(&(const_cast< Potential< GUM_SCALAR >& >(a.second->cpf())));

      InstanceBayesNet< GUM_SCALAR > bn(*i);
      DefaultTriangulation           t(&(bn.moralGraph()), &(bn.modalities()));
      const std::vector< NodeId >&   full_elim_order = t.eliminationOrder();

      for (auto var = full_elim_order.begin(); var != full_elim_order.end(); ++var)
        eliminateNode(&(i->get(*var).type().variable()), pool, trash);
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::__insertLiftedNodes(
       const PRMInstance< GUM_SCALAR >* i, BucketSet& pool, BucketSet& trash) {
      BucketSet* lifted_pool = nullptr;

      try {
        lifted_pool = __lifted_pools[&(__bb.requisiteNodes(i))];
      } catch (NotFound&) {
        __initLiftedNodes(i, trash);
        lifted_pool = __lifted_pools[&(__bb.requisiteNodes(i))];
      }

      for (const auto lifted_pot : *lifted_pool) {
        Potential< GUM_SCALAR >* pot = copyPotential(i->bijection(), *lifted_pot);
        pool.insert(pot);
        trash.insert(pot);
      }
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::__initLiftedNodes(const PRMInstance< GUM_SCALAR >* i,
                                               BucketSet& trash) {
      PRMClass< GUM_SCALAR >& c = const_cast< PRMClass< GUM_SCALAR >& >(i->type());
      BucketSet*              lifted_pool = new BucketSet();
      __lifted_pools.insert(&(__bb.requisiteNodes(i)), lifted_pool);

      for (const auto node : __bb.requisiteNodes(i))
        if (PRMClassElement< GUM_SCALAR >::isAttribute(c.get(node)))
          lifted_pool->insert(
             const_cast< Potential< GUM_SCALAR >* >(&(c.get(node).cpf())));

      NodeSet inners, outers, ignore;

      for (const auto& elt : *i) {
        if (__bb.requisiteNodes(*i).exists(elt.first)) {
          if (PRMClassElement< GUM_SCALAR >::isAttribute(c.get(elt.first))) {
            if (c.isOutputNode(c.get(elt.first)))
              outers.insert(elt.first);
            else if (!outers.exists(elt.first))
              inners.insert(elt.first);
          } else if (PRMClassElement< GUM_SCALAR >::isAggregate(
                        c.get(elt.first))) {
            outers.insert(elt.first);

            // We need to put in the output_elim_order aggregator's parents
            // which are
            // innner nodes
            for (const auto par : c.containerDag().parents(elt.first))
              if (PRMClassElement< GUM_SCALAR >::isAttribute(i->type().get(par))
                  && i->type().isInnerNode(i->type().get(par))
                  && __bb.requisiteNodes(i).exists(par)) {
                inners.erase(par);
                outers.insert(par);
              }
          }
        } else {
          ignore.insert(elt.first);
        }
      }

      // Now we proceed with the elimination of inner attributes
      ClassBayesNet< GUM_SCALAR > bn(c);
      List< NodeSet >             partial_ordering;

      if (inners.size()) partial_ordering.pushBack(inners);

      if (outers.size()) partial_ordering.pushBack(outers);

      if (ignore.size()) partial_ordering.pushBack(ignore);

      GUM_ASSERT(inners.size() || outers.size());
      PartialOrderedTriangulation t(
         &(bn.moralGraph()), &(bn.modalities()), &partial_ordering);

      for (size_t idx = 0; idx < inners.size(); ++idx)
        eliminateNode(&(c.get(t.eliminationOrder()[idx]).type().variable()),
                      *lifted_pool,
                      trash);

      // If there is not only inner and input Attributes
      if (outers.size()) {
        __elim_orders.insert(
           &c,
           new std::vector< NodeId >(t.eliminationOrder().begin() + inners.size(),
                                     t.eliminationOrder().end()));
      }
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::__initElimOrder() {
      ClassDependencyGraph< GUM_SCALAR >                        cdg(*(this->_prm));
      Sequence< const PRMClassElementContainer< GUM_SCALAR >* > class_elim_order;
      std::list< NodeId >                                       l;

      for (const auto node : cdg.dag().nodes()) {
        if (cdg.dag().parents(node).empty()) { l.push_back(node); }
      }

      Set< NodeId > visited_node;

      while (!l.empty()) {
        visited_node.insert(l.front());

        if (!class_elim_order.exists(cdg.get(l.front()).first)) {
          class_elim_order.insert(cdg.get(l.front()).first);
        }

        for (const auto child : cdg.dag().children(l.front())) {
          if (!visited_node.contains(child)) { l.push_back(child); }
        }

        l.pop_front();
      }

      __class_elim_order = new Sequence< std::string >();
      for (auto c : class_elim_order) {
        std::string name = c->name();
        auto        pos = name.find_first_of("<");
        if (pos != std::string::npos) { name = name.substr(0, pos); }
        try {
          __class_elim_order->insert(name);
        } catch (DuplicateElement&) {}
      }
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::_marginal(const Chain&             chain,
                                       Potential< GUM_SCALAR >& m) {
      const PRMInstance< GUM_SCALAR >*  i = chain.first;
      const PRMAttribute< GUM_SCALAR >* elt = chain.second;
      SVED< GUM_SCALAR >::BucketSet     pool, trash;
      __bb.compute(i, elt->id());
      __eliminateNodes(i, elt->id(), pool, trash);

      std::vector< const Potential< GUM_SCALAR >* > result;
      for (auto pot : pool) {
        if (pot->contains(*(m.variablesSequence().atPos(0))))
          result.push_back(pot);
      }

      while (result.size() > 1) {
        const auto& p1 = *(result.back());
        result.pop_back();
        const auto& p2 = *(result.back());
        result.pop_back();
        auto mult = new Potential< GUM_SCALAR >(p1 * p2);
        result.push_back(mult);
        trash.insert(mult);
      }

      m = *(result.back());
      m.normalize();

      GUM_ASSERT(m.nbrDim() == (Size)1);

      // cleaning up the mess
      for (const auto pot : trash)
        delete pot;

      for (const auto& elt : __lifted_pools)
        delete elt.second;

      __lifted_pools.clear();

      for (const auto& elt : __req_set) {
        delete elt.second.first;
        delete elt.second.second;
      }

      __req_set.clear();

      for (const auto elt : __elim_orders)
        delete elt.second;

      __elim_orders.clear();
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::_joint(const std::vector< Chain >& queries,
                                    Potential< GUM_SCALAR >&    j) {
      GUM_ERROR(FatalError, "Not implemented.");
    }

    template < typename GUM_SCALAR >
    void SVED< GUM_SCALAR >::__initReqSets(const PRMInstance< GUM_SCALAR >* i) {
      Set< NodeId >* attr_set = new Set< NodeId >();
      Set< NodeId >* sc_set = new Set< NodeId >();

      for (const auto node : __bb.requisiteNodes(i)) {
        switch (i->type().get(node).elt_type()) {
          case PRMClassElement< GUM_SCALAR >::prm_aggregate:
          case PRMClassElement< GUM_SCALAR >::prm_attribute: {
            attr_set->insert(node);
            break;
          }

          case PRMClassElement< GUM_SCALAR >::prm_slotchain: {
            sc_set->insert(node);
            break;
          }

          default: {
            GUM_ERROR(FatalError,
                      "There should not be elements other"
                      " than PRMAttribute<GUM_SCALAR> and SlotChain.");
          }
        }
      }

      __req_set.insert(
         &(__bb.requisiteNodes(i)),
         std::pair< Set< NodeId >*, Set< NodeId >* >(attr_set, sc_set));
    }

    template < typename GUM_SCALAR >
    INLINE SVED< GUM_SCALAR >::SVED(const PRM< GUM_SCALAR >&       prm,
                                    const PRMSystem< GUM_SCALAR >& model) :
        PRMInference< GUM_SCALAR >(prm, model),
        __class_elim_order(0), __bb(*this) {
      GUM_CONSTRUCTOR(SVED);
    }


    template < typename GUM_SCALAR >
    INLINE void
       SVED< GUM_SCALAR >::__insertEvidence(const PRMInstance< GUM_SCALAR >* i,
                                            BucketSet& pool) {
      for (const auto& elt : this->evidence(i))
        pool.insert(const_cast< Potential< GUM_SCALAR >* >(elt.second));
    }

    template < typename GUM_SCALAR >
    INLINE std::vector< NodeId >&
           SVED< GUM_SCALAR >::__getElimOrder(const PRMClass< GUM_SCALAR >& c) {
      return *(__elim_orders[&c]);
    }

    template < typename GUM_SCALAR >
    INLINE std::string SVED< GUM_SCALAR >::__trim(const std::string& s) {
      auto pos = s.find_first_of("<");
      if (pos != std::string::npos) { return s.substr(0, pos); }
      return s;
    }

    template < typename GUM_SCALAR >
    INLINE bool SVED< GUM_SCALAR >::__checkElimOrder(
       const PRMInstance< GUM_SCALAR >* first,
       const PRMInstance< GUM_SCALAR >* second) {
      if (__class_elim_order == 0) { __initElimOrder(); }

      auto first_name = __trim(first->type().name());
      auto second_name = __trim(second->type().name());
      return (__class_elim_order->pos(first_name)
              <= __class_elim_order->pos(second_name));
    }

    template < typename GUM_SCALAR >
    INLINE Potential< GUM_SCALAR >* SVED< GUM_SCALAR >::__getAggPotential(
       const PRMInstance< GUM_SCALAR >* i, const PRMAggregate< GUM_SCALAR >* agg) {
      return &(
         const_cast< Potential< GUM_SCALAR >& >(i->get(agg->safeName()).cpf()));
    }

    template < typename GUM_SCALAR >
    INLINE void SVED< GUM_SCALAR >::_evidenceAdded(
       const typename SVED< GUM_SCALAR >::Chain& chain) {
      // Do nothing
    }

    template < typename GUM_SCALAR >
    INLINE void SVED< GUM_SCALAR >::_evidenceRemoved(
       const typename SVED< GUM_SCALAR >::Chain& chain) {
      // Do nothing
    }

    template < typename GUM_SCALAR >
    INLINE Set< NodeId >&
           SVED< GUM_SCALAR >::__getAttrSet(const PRMInstance< GUM_SCALAR >* i) {
      try {
        return *(__req_set[&(__bb.requisiteNodes(i))].first);
      } catch (NotFound&) {
        __initReqSets(i);
        return *(__req_set[&(__bb.requisiteNodes(i))].first);
      }
    }

    template < typename GUM_SCALAR >
    INLINE Set< NodeId >&
           SVED< GUM_SCALAR >::__getSCSet(const PRMInstance< GUM_SCALAR >* i) {
      try {
        return *(__req_set[&(__bb.requisiteNodes(i))].second);
      } catch (NotFound&) {
        __initReqSets(i);
        return *(__req_set[&(__bb.requisiteNodes(i))].second);
      }
    }

    template < typename GUM_SCALAR >
    INLINE void SVED< GUM_SCALAR >::__reduceElimList(
       const PRMInstance< GUM_SCALAR >*          i,
       List< const PRMInstance< GUM_SCALAR >* >& elim_list,
       List< const PRMInstance< GUM_SCALAR >* >& reduced_list,
       Set< const PRMInstance< GUM_SCALAR >* >&  ignore,
       BucketSet&                                pool,
       BucketSet&                                trash) {
      while (!elim_list.empty()) {
        if (__checkElimOrder(i, elim_list.front())) {
          if ((!ignore.exists(elim_list.front()))
              && (__bb.exists(elim_list.front()))) {
            __eliminateNodesDownward(
               i, elim_list.front(), pool, trash, elim_list, ignore);
          }
        } else if (__bb.exists(elim_list.front())) {
          reduced_list.insert(elim_list.front());
        }

        elim_list.popFront();
      }
    }

    template < typename GUM_SCALAR >
    INLINE std::string SVED< GUM_SCALAR >::name() const {
      return "SVED";
    }

  } /* namespace prm */
} /* namespace gum */