incrementalGraphLearner_tpl.h

Go to the documentation of this file.

// =======================================================
#include <queue>
// =======================================================
#include <agrum/core/math/math.h>
#include <agrum/core/multiPriorityQueue.h>
#include <agrum/core/types.h>
// =======================================================
#include <agrum/FMDP/learning/core/chiSquare.h>
#include <agrum/FMDP/learning/datastructure/incrementalGraphLearner.h>
// =======================================================
#include <agrum/variables/discreteVariable.h>
// =======================================================

namespace gum {

  // ============================================================================
  // ============================================================================

  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  IncrementalGraphLearner< AttributeSelection, isScalar >::IncrementalGraphLearner(
     MultiDimFunctionGraph< double >* target,
     Set< const DiscreteVariable* >   varList,
     const DiscreteVariable*          value) :
      _target(target),
      _setOfVars(varList), _value(value) {
    GUM_CONSTRUCTOR(IncrementalGraphLearner);

    for (auto varIter = _setOfVars.cbeginSafe(); varIter != _setOfVars.cendSafe();
         ++varIter)
      _var2Node.insert(*varIter, new LinkedList< NodeId >());
    _var2Node.insert(_value, new LinkedList< NodeId >());

    _model.addNode();
    this->_root = _insertLeafNode(
       new NodeDatabase< AttributeSelection, isScalar >(&_setOfVars, _value),
       _value,
       new Set< const Observation* >());
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  IncrementalGraphLearner< AttributeSelection,
                           isScalar >::~IncrementalGraphLearner() {
    for (auto nodeIter = _nodeId2Database.beginSafe();
         nodeIter != _nodeId2Database.endSafe();
         ++nodeIter)
      delete nodeIter.val();

    for (auto nodeIter = _nodeSonsMap.beginSafe();
         nodeIter != _nodeSonsMap.endSafe();
         ++nodeIter)
      SOA_DEALLOCATE(nodeIter.val(),
                     sizeof(NodeId) * _nodeVarMap[nodeIter.key()]->domainSize());

    for (auto varIter = _var2Node.beginSafe(); varIter != _var2Node.endSafe();
         ++varIter)
      delete varIter.val();

    for (auto nodeIter = _leafDatabase.beginSafe();
         nodeIter != _leafDatabase.endSafe();
         ++nodeIter)
      delete nodeIter.val();

    __clearValue();

    GUM_DESTRUCTOR(IncrementalGraphLearner);
  }


  // ============================================================================
  // ============================================================================

  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  void IncrementalGraphLearner< AttributeSelection, isScalar >::addObservation(
     const Observation* newObs) {
    __assumeValue(newObs);

    // The we go across the tree
    NodeId currentNodeId = _root;

    while (_nodeSonsMap.exists(currentNodeId)) {
      // On each encountered node, we update the database
      _updateNodeWithObservation(newObs, currentNodeId);

      // The we select the next to go throught
      currentNodeId =
         _nodeSonsMap[currentNodeId]
                     [__branchObs(newObs, _nodeVarMap[currentNodeId])];
    }

    // On final insertion into the leave we reach
    _updateNodeWithObservation(newObs, currentNodeId);
    _leafDatabase[currentNodeId]->insert(newObs);
  }


  // ============================================================================
  // ============================================================================

  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  void IncrementalGraphLearner< AttributeSelection, isScalar >::updateVar(
     const DiscreteVariable* var) {
    Link< NodeId >* nodIter = _var2Node[var]->list();
    Link< NodeId >* nni = nullptr;
    while (nodIter) {
      nni = nodIter->nextLink();
      _convertNode2Leaf(nodIter->element());
      nodIter = nni;
    }
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  void IncrementalGraphLearner< AttributeSelection, isScalar >::_updateNode(
     NodeId updatedNode, Set< const DiscreteVariable* >& varsOfInterest) {
    // If this node has no interesting variable, we turn it into a leaf
    if (varsOfInterest.empty()) {
      _convertNode2Leaf(updatedNode);
      return;
    }

    // If this node has already one of the best variable intalled as test, we
    // move on
    if (_nodeVarMap.exists(updatedNode)
        && varsOfInterest.exists(_nodeVarMap[updatedNode])) {
      return;
    }

    // In any other case we have to install variable as best test
    Idx randy = (Idx)(std::rand() / RAND_MAX) * varsOfInterest.size(), basc = 0;
    SetConstIteratorSafe< const DiscreteVariable* > varIter;
    for (varIter = varsOfInterest.cbeginSafe(), basc = 0;
         varIter != varsOfInterest.cendSafe() && basc < randy;
         ++varIter, basc++)
      ;

    _transpose(updatedNode, *varIter);
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  void IncrementalGraphLearner< AttributeSelection, isScalar >::_convertNode2Leaf(
     NodeId currentNodeId) {
    if (_nodeVarMap[currentNodeId] != _value) {
      _leafDatabase.insert(currentNodeId, new Set< const Observation* >());

      // Resolving potential sons issue
      for (Idx modality = 0; modality < _nodeVarMap[currentNodeId]->domainSize();
           ++modality) {
        NodeId sonId = _nodeSonsMap[currentNodeId][modality];
        _convertNode2Leaf(sonId);
        (*_leafDatabase[currentNodeId]) =
           (*_leafDatabase[currentNodeId]) + *(_leafDatabase[sonId]);
        _removeNode(sonId);
      }

      SOA_DEALLOCATE(_nodeSonsMap[currentNodeId],
                     sizeof(NodeId) * _nodeVarMap[currentNodeId]->domainSize());
      _nodeSonsMap.erase(currentNodeId);

      _chgNodeBoundVar(currentNodeId, _value);
    }
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  void IncrementalGraphLearner< AttributeSelection, isScalar >::_transpose(
     NodeId currentNodeId, const DiscreteVariable* desiredVar) {
    // **************************************************************************************
    // Si le noeud courant contient déjà la variable qu'on souhaite lui amener
    // Il n'y a rien à faire
    if (_nodeVarMap[currentNodeId] == desiredVar) { return; }

    // **************************************************************************************
    // Si le noeud courant est terminal,
    // Il faut artificiellement insérer un noeud liant à la variable
    if (_nodeVarMap[currentNodeId] == _value) {
      // We turned this leaf into an internal node.
      // This mean that we'll need to install children leaves for each value of
      // desiredVar

      // First We must prepare these new leaves NodeDatabases and Sets<const
      // Observation*>
      NodeDatabase< AttributeSelection, isScalar >** dbMap =
         static_cast< NodeDatabase< AttributeSelection, isScalar >** >(
            SOA_ALLOCATE(sizeof(NodeDatabase< AttributeSelection, isScalar >*)
                         * desiredVar->domainSize()));
      Set< const Observation* >** obsetMap =
         static_cast< Set< const Observation* >** >(SOA_ALLOCATE(
            sizeof(Set< const Observation* >*) * desiredVar->domainSize()));
      for (Idx modality = 0; modality < desiredVar->domainSize(); ++modality) {
        dbMap[modality] =
           new NodeDatabase< AttributeSelection, isScalar >(&_setOfVars, _value);
        obsetMap[modality] = new Set< const Observation* >();
      }
      for (SetIteratorSafe< const Observation* > obsIter =
              _leafDatabase[currentNodeId]->beginSafe();
           _leafDatabase[currentNodeId]->endSafe() != obsIter;
           ++obsIter) {
        dbMap[__branchObs(*obsIter, desiredVar)]->addObservation(*obsIter);
        obsetMap[__branchObs(*obsIter, desiredVar)]->insert(*obsIter);
      }

      // Then we can install each new leaves (and put in place the sonsMap)
      NodeId* sonsMap = static_cast< NodeId* >(
         SOA_ALLOCATE(sizeof(NodeId) * desiredVar->domainSize()));
      for (Idx modality = 0; modality < desiredVar->domainSize(); ++modality)
        sonsMap[modality] =
           _insertLeafNode(dbMap[modality], _value, obsetMap[modality]);

      // Some necessary clean up
      SOA_DEALLOCATE(dbMap,
                     sizeof(NodeDatabase< AttributeSelection, isScalar >*)
                        * desiredVar->domainSize());
      SOA_DEALLOCATE(
         obsetMap, sizeof(Set< const Observation* >*) * desiredVar->domainSize());

      // And finally we can turn the node into an internal node associated to
      // desiredVar
      _chgNodeBoundVar(currentNodeId, desiredVar);
      _nodeSonsMap.insert(currentNodeId, sonsMap);

      return;
    }

    // *************************************************************************************
    // Remains the general case where currentNodeId is an internal node.

    // First we ensure that children node use desiredVar as variable
    for (Idx modality = 0; modality < _nodeVarMap[currentNodeId]->domainSize();
         ++modality)
      _transpose(_nodeSonsMap[currentNodeId][modality], desiredVar);

    //        Sequence<NodeDatabase<AttributeSelection, isScalar>*>
    //        sonsNodeDatabase =
    //        _nodeId2Database[currentNodeId]->splitOnVar(desiredVar);
    NodeId* sonsMap = static_cast< NodeId* >(
       SOA_ALLOCATE(sizeof(NodeId) * desiredVar->domainSize()));

    // Then we create the new mapping
    for (Idx desiredVarModality = 0; desiredVarModality < desiredVar->domainSize();
         ++desiredVarModality) {
      NodeId* grandSonsMap = static_cast< NodeId* >(
         SOA_ALLOCATE(sizeof(NodeId) * _nodeVarMap[currentNodeId]->domainSize()));
      NodeDatabase< AttributeSelection, isScalar >* sonDB =
         new NodeDatabase< AttributeSelection, isScalar >(&_setOfVars, _value);
      for (Idx currentVarModality = 0;
           currentVarModality < _nodeVarMap[currentNodeId]->domainSize();
           ++currentVarModality) {
        grandSonsMap[currentVarModality] =
           _nodeSonsMap[_nodeSonsMap[currentNodeId][currentVarModality]]
                       [desiredVarModality];
        sonDB->operator+=((*_nodeId2Database[grandSonsMap[currentVarModality]]));
      }

      sonsMap[desiredVarModality] =
         _insertInternalNode(sonDB, _nodeVarMap[currentNodeId], grandSonsMap);
    }

    // Finally we clean the old remaining nodes
    for (Idx currentVarModality = 0;
         currentVarModality < _nodeVarMap[currentNodeId]->domainSize();
         ++currentVarModality) {
      _removeNode(_nodeSonsMap[currentNodeId][currentVarModality]);
    }

    // We suppress the old sons map and remap to the new one
    SOA_DEALLOCATE(_nodeSonsMap[currentNodeId],
                   sizeof(NodeId) * _nodeVarMap[currentNodeId]->domainSize());
    _nodeSonsMap[currentNodeId] = sonsMap;

    _chgNodeBoundVar(currentNodeId, desiredVar);
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  NodeId IncrementalGraphLearner< AttributeSelection, isScalar >::_insertNode(
     NodeDatabase< AttributeSelection, isScalar >* nDB,
     const DiscreteVariable*                       boundVar) {
    NodeId newNodeId = _model.addNode();
    _nodeVarMap.insert(newNodeId, boundVar);
    _nodeId2Database.insert(newNodeId, nDB);
    _var2Node[boundVar]->addLink(newNodeId);

    _needUpdate = true;

    return newNodeId;
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  NodeId
     IncrementalGraphLearner< AttributeSelection, isScalar >::_insertInternalNode(
        NodeDatabase< AttributeSelection, isScalar >* nDB,
        const DiscreteVariable*                       boundVar,
        NodeId*                                       sonsMap) {
    NodeId newNodeId = this->_insertNode(nDB, boundVar);
    _nodeSonsMap.insert(newNodeId, sonsMap);
    return newNodeId;
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  NodeId IncrementalGraphLearner< AttributeSelection, isScalar >::_insertLeafNode(
     NodeDatabase< AttributeSelection, isScalar >* nDB,
     const DiscreteVariable*                       boundVar,
     Set< const Observation* >*                    obsSet) {
    NodeId newNodeId = this->_insertNode(nDB, boundVar);
    _leafDatabase.insert(newNodeId, obsSet);
    return newNodeId;
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  void IncrementalGraphLearner< AttributeSelection, isScalar >::_chgNodeBoundVar(
     NodeId currentNodeId, const DiscreteVariable* desiredVar) {
    if (_nodeVarMap[currentNodeId] == desiredVar) return;

    _var2Node[_nodeVarMap[currentNodeId]]->searchAndRemoveLink(currentNodeId);
    _var2Node[desiredVar]->addLink(currentNodeId);
    _nodeVarMap[currentNodeId] = desiredVar;

    if (_nodeVarMap[currentNodeId] != _value
        && _leafDatabase.exists(currentNodeId)) {
      delete _leafDatabase[currentNodeId];
      _leafDatabase.erase(currentNodeId);
    }

    if (_nodeVarMap[currentNodeId] == _value
        && !_leafDatabase.exists(currentNodeId)) {
      _leafDatabase.insert(currentNodeId, new Set< const Observation* >());
    }

    _needUpdate = true;
  }


  // ############################################################################
  // ############################################################################
  template < TESTNAME AttributeSelection, bool isScalar >
  void IncrementalGraphLearner< AttributeSelection, isScalar >::_removeNode(
     NodeId currentNodeId) {
    // Retriat de l'id
    _model.eraseNode(currentNodeId);

    // Retrait du vecteur fils
    if (_nodeSonsMap.exists(currentNodeId)) {
      SOA_DEALLOCATE(_nodeSonsMap[currentNodeId],
                     sizeof(NodeId) * _nodeVarMap[currentNodeId]->domainSize());
      _nodeSonsMap.erase(currentNodeId);
    }

    if (_leafDatabase.exists(currentNodeId)) {
      delete _leafDatabase[currentNodeId];
      _leafDatabase.erase(currentNodeId);
    }

    // Retrait de la variable
    _var2Node[_nodeVarMap[currentNodeId]]->searchAndRemoveLink(currentNodeId);
    _nodeVarMap.erase(currentNodeId);

    // Retrait du NodeDatabase
    delete _nodeId2Database[currentNodeId];
    _nodeId2Database.erase(currentNodeId);

    _needUpdate = true;
  }
}   // namespace gum