gum::ITI< AttributeSelection, isScalar > Class Template Reference

Learn a graphical representation of a function as a decision tree. More...

#include <agrum/FMDP/planning/FunctionGraph/iti.h>

Inheritance diagram for gum::ITI< AttributeSelection, isScalar >:

Collaboration diagram for gum::ITI< AttributeSelection, isScalar >:

Public Member Functions
Size	size ()
Constructor & destructor.
	ITI (MultiDimFunctionGraph< double > *target, double attributeSelectionThreshold, Set< const DiscreteVariable * > attributeListe, const DiscreteVariable *learnedValue)
	ITI constructor for functions describing the behaviour of one variable according to a set of other variable such as conditionnal probabilities. More...
	ITI (MultiDimFunctionGraph< double > *target, double attributeSelectionThreshold, Set< const DiscreteVariable * > attributeListe)
	ITI constructeur for real functions. More...
	~ITI ()
	Default destructor. More...
Visit Methods
NodeId	root () const
bool	isTerminal (NodeId ni) const
const DiscreteVariable *	nodeVar (NodeId ni) const
NodeId	nodeSon (NodeId ni, Idx modality) const
Idx	nodeNbObservation (NodeId ni) const
virtual void	insertSetOfVars (MultiDimFunctionGraph< double > *ret) const

Protected Attributes
MultiDimFunctionGraph< double > *	target_
	The final diagram we're building. More...
Set< const DiscreteVariable *>	setOfVars_
const DiscreteVariable *	value_
Sequence< ValueType >	valueAssumed_
bool	needUpdate_
Model handling datastructures
NodeGraphPart	model_
	The source of nodeId. More...
NodeId	root_
	The root of the ordered tree. More...
HashTable< NodeId, const DiscreteVariable *>	nodeVarMap_
	Gives for any node its associated variable. More...
HashTable< NodeId, NodeId *>	nodeSonsMap_
	A table giving for any node a table mapping to its son idx is the modality of associated variable. More...
HashTable< const DiscreteVariable *, LinkedList< NodeId > *>	var2Node_
	Associates to any variable the list of all nodes associated to this variable. More...
HashTable< NodeId, NodeDatabase< AttributeSelection, isScalar > *>	nodeId2Database_
	This hashtable binds every node to an associated NodeDatabase which handles every observation that concerns that node. More...
HashTable< NodeId, Set< const Observation *> *>	leafDatabase_
	This hashtable binds to every leaf an associated set of all hte observations compatible with it. More...

Protected Member Functions
void	insertSetOfVars_ (MultiDimFunctionGraph< double > *ret)
	insertSetOfVars_ More...

Graph Structure update methods
virtual void	updateVar (const DiscreteVariable *)
	If a new modality appears to exists for given variable, call this method to turn every associated node to this variable into leaf. Graph has then indeed to be revised. More...
void	updateNode_ (NodeId nody, Set< const DiscreteVariable * > &bestVars)
	From the given sets of node, selects randomly one and installs it on given node. More...
virtual void	convertNode2Leaf_ (NodeId)
	Turns the given node into a leaf if not already so. More...
virtual void	transpose_ (NodeId, const DiscreteVariable *)
	Installs given variable to the given node, ensuring that the variable is not present in its subtree. More...
virtual NodeId	insertInternalNode_ (NodeDatabase< AttributeSelection, isScalar > *nDB, const DiscreteVariable *boundVar, NodeId *sonsMap)
	inserts a new internal node in internal graph More...
virtual NodeId	insertLeafNode_ (NodeDatabase< AttributeSelection, isScalar > *nDB, const DiscreteVariable *boundVar, Set< const Observation * > *obsSet)
	inserts a new leaf node in internal graohs More...

New Observation insertion methods
void	addObservation (const Observation *obs)
	Inserts a new observation. More...
void	updateNodeWithObservation_ (const Observation *newObs, NodeId currentNodeId)
	Will update internal graph's NodeDatabase of given node with the new observation. More...

Graph Structure update methods
void	updateGraph ()
	Updates the internal graph after a new observation has been added. More...
NodeId	insertNode_ (NodeDatabase< AttributeSelection, isScalar > *nDB, const DiscreteVariable *boundVar)
	inserts a new node in internal graph More...
void	chgNodeBoundVar_ (NodeId chgedNodeId, const DiscreteVariable *desiredVar)
	Changes the associated variable of a node. More...
void	removeNode_ (NodeId removedNodeId)
	Removes a node from the internal graph. More...

Function Graph Updating methods
void	updateFunctionGraph ()
	Updates target to currently learned graph structure. More...
NodeId	_insertNodeInFunctionGraph_ (NodeId src)
	Inserts an internal node in the target. More...
NodeId	_insertTerminalNode_ (NodeId src)
	Insert a terminal node in the target. More...
NodeId	_insertTerminalNode_ (NodeId src, Int2Type< true >)
	Insert a terminal node in the target. More...
NodeId	_insertTerminalNode_ (NodeId src, Int2Type< false >)
	Insert a terminal node in the target. More...

Detailed Description

template<TESTNAME AttributeSelection, bool isScalar = false>
class gum::ITI< AttributeSelection, isScalar >

Learn a graphical representation of a function as a decision tree.

This learning is done incrementaly. Hence first observation are add to the structure and then the structure is updated. Maintains two graph function : one internal for the learning and a target which is updated on demand.

Definition at line 61 of file iti.h.

Constructor & Destructor Documentation

ITI() [1/2]

template<TESTNAME AttributeSelection, bool isScalar>

gum::ITI< AttributeSelection, isScalar >::ITI	(	MultiDimFunctionGraph< double > *	target,
		double	attributeSelectionThreshold,
		Set< const DiscreteVariable * >	attributeListe,
		const DiscreteVariable *	learnedValue
	)

ITI constructor for functions describing the behaviour of one variable according to a set of other variable such as conditionnal probabilities.

Parameters

target	: the MultiDimFunctionGraph in which we load the structure
attributeSelectionThreshold	: threshold under which a node is not installed (pe-pruning)
attributeListe	: Set of vars on which we rely to explain the behaviour of learned variable
learnedValue	: the variable from which we try to learn the behaviour
target	: the MultiDimFunctionGraph in which we load the structure
attributeSelectionThreshold	: threshold under which a node is not installed (pe-pruning)
temporaryAPIfix	: Issue in API in regard to IMDDI
attributeListe	: Set of vars on which we rely to explain the behaviour of learned variable
learnedValue	: the variable from which we try to learn the behaviour

Definition at line 61 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                                        :
      IncrementalGraphLearner< AttributeSelection, isScalar >(target, attributeListe, learnedValue),
      _nbTotalObservation_(0), _attributeSelectionThreshold_(attributeSelectionThreshold) {
    GUM_CONSTRUCTOR(ITI);
    _staleTable_.insert(this->root_, false);
  }

Here is the call graph for this function:

ITI() [2/2]

template<TESTNAME AttributeSelection, bool isScalar>

gum::ITI< AttributeSelection, isScalar >::ITI	(	MultiDimFunctionGraph< double > *	target,
		double	attributeSelectionThreshold,
		Set< const DiscreteVariable * >	attributeListe
	)

ITI constructeur for real functions.

We try to predict the output of a function f given a set of variable

Parameters

target	: the MultiDimFunctionGraph in which we load the structure
attributeSelectionThreshold	: threshold under which a node is not installed (pe-pruning)
attributeListe	: Set of vars on which we rely to explain the behaviour of learned function

We try to predict the output of a function f given a set of variable

Parameters

target	: the MultiDimFunctionGraph in which we load the structure
attributeSelectionThreshold	: threshold under which a node is not installed (pe-pruning)
temporaryAPIfix	: Issue in API in regard to IMDDI
attributeListeSet	of vars on which we rely to explain the behaviour of learned function

Definition at line 84 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                                          :
      IncrementalGraphLearner< AttributeSelection, isScalar >(
         target,
         attributeListe,
         new LabelizedVariable("Reward", "", 2)),
      _nbTotalObservation_(0), _attributeSelectionThreshold_(attributeSelectionThreshold) {
    GUM_CONSTRUCTOR(ITI);
    _staleTable_.insert(this->root_, false);
  }

Here is the call graph for this function:

~ITI()

template<TESTNAME AttributeSelection, bool isScalar = false>

gum::ITI< AttributeSelection, isScalar >::~ITI ( )

inline

Default destructor.

Definition at line 104 of file iti.h.

           {
      GUM_DESTRUCTOR(ITI);
      ;
    }

Member Function Documentation

_insertNodeInFunctionGraph_()

template<TESTNAME AttributeSelection, bool isScalar>

NodeId gum::ITI< AttributeSelection, isScalar >::_insertNodeInFunctionGraph_ ( NodeId  currentNodeId )

private

Inserts an internal node in the target.

Parameters

src	the source node in internal graph

Returns

the mathcing node id in the target

Parameters

the	source node in internal graph

Returns

the mathcing node id in the target

Definition at line 265 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                                              {
    if (this->nodeVarMap_[currentNodeId] == this->value_) {
      NodeId nody = _insertTerminalNode_(currentNodeId);
      return nody;
    }

    if (!this->target_->variablesSequence().exists(this->nodeVarMap_[currentNodeId])) {
      this->target_->add(*(this->nodeVarMap_[currentNodeId]));
    }

    NodeId nody = this->target_->manager()->addInternalNode(this->nodeVarMap_[currentNodeId]);
    for (Idx moda = 0; moda < this->nodeVarMap_[currentNodeId]->domainSize(); ++moda) {
      NodeId son = this->_insertNodeInFunctionGraph_(this->nodeSonsMap_[currentNodeId][moda]);
      this->target_->manager()->setSon(nody, moda, son);
    }

    return nody;
  }

Here is the call graph for this function:

_insertTerminalNode_() [1/3]

template<TESTNAME AttributeSelection, bool isScalar = false>

NodeId gum::ITI< AttributeSelection, isScalar >::_insertTerminalNode_ ( NodeId  src )

inlineprivate

Insert a terminal node in the target.

This function is a dispatcher that will call the right function according to the value of the template isScalar

Parameters

src	the source node in the learned graph

Returns

the matching node in the target

Definition at line 209 of file iti.h.

                                            {
      return _insertTerminalNode_(src, Int2Type< isScalar >());
    }

_insertTerminalNode_() [2/3]

template<TESTNAME AttributeSelection, bool isScalar>

NodeId gum::ITI< AttributeSelection, isScalar >::_insertTerminalNode_ ( NodeId  currentNodeId, Int2Type< true >    )

private

Insert a terminal node in the target.

This function is called if we're learning a real value function. Inserts then a single value in target.

Parameters

src	the source node in the learned graph

Returns

the matching node in the target

This function is called if we're learning the behaviour of a variable. Inserts then this variable and the relevant value beneath into target.

Parameters

the	source node in the learned graph

Returns

the matching node in the target

Definition at line 325 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                                     {
    double value = 0.0;
    for (auto valIter = this->nodeId2Database_[currentNodeId]->cbeginValues();
         valIter != this->nodeId2Database_[currentNodeId]->cendValues();
         ++valIter) {
      value += (double)valIter.key() * valIter.val();
    }
    if (this->nodeId2Database_[currentNodeId]->nbObservation())
      value /= (double)this->nodeId2Database_[currentNodeId]->nbObservation();
    NodeId nody = this->target_->manager()->addTerminalNode(value);
    return nody;
  }

Here is the call graph for this function:

_insertTerminalNode_() [3/3]

template<TESTNAME AttributeSelection, bool isScalar>

NodeId gum::ITI< AttributeSelection, isScalar >::_insertTerminalNode_ ( NodeId  currentNodeId, Int2Type< false >    )

private

Insert a terminal node in the target.

This function is called if we're learning the behaviour of a variable. Inserts then this variable and the relevant value beneath into target.

Parameters

src	the source node in the learned graph

Returns

the matching node in the target

This function is called if we're learning a real value function. Inserts then a single value in target.

Parameters

the	source node in the learned graph

Returns

the matching node in the target

Definition at line 295 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                                      {
    if (!this->target_->variablesSequence().exists(this->value_))
      this->target_->add(*(this->value_));

    Size tot = this->nodeId2Database_[currentNodeId]->nbObservation();
    if (tot == Size(0)) return this->target_->manager()->addTerminalNode(0.0);

    NodeId* sonsMap
       = static_cast< NodeId* >(SOA_ALLOCATE(sizeof(NodeId) * this->value_->domainSize()));
    for (Idx modality = 0; modality < this->value_->domainSize(); ++modality) {
      double newVal = 0.0;
      newVal = (double)this->nodeId2Database_[currentNodeId]->effectif(modality) / (double)tot;
      sonsMap[modality] = this->target_->manager()->addTerminalNode(newVal);
    }
    NodeId nody = this->target_->manager()->addInternalNode(this->value_, sonsMap);
    return nody;
  }

Here is the call graph for this function:

addObservation()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::ITI< AttributeSelection, isScalar >::addObservation ( const Observation *  obs )

virtual

Inserts a new observation.

Parameters

obs	the new observation to learn
the	new observation to learn

Reimplemented from gum::IncrementalGraphLearner< AttributeSelection, isScalar >.

Definition at line 108 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                                 {
    _nbTotalObservation_++;
    IncrementalGraphLearner< AttributeSelection, isScalar >::addObservation(obs);
  }

Here is the call graph for this function:

chgNodeBoundVar_()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::ITI< AttributeSelection, isScalar >::chgNodeBoundVar_ ( NodeId  currentNodeId, const DiscreteVariable *  desiredVar  )

protectedvirtual

Changes the associated variable of a node.

Parameters

chgedNodeId	: the node to change
desiredVar	: its new associated variable

Reimplemented from gum::IncrementalGraphLearner< AttributeSelection, isScalar >.

Definition at line 220 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                                                 {
    if (this->nodeVarMap_[currentNodeId] != desiredVar) {
      _staleTable_[currentNodeId] = true;
      IncrementalGraphLearner< AttributeSelection, isScalar >::chgNodeBoundVar_(currentNodeId,
                                                                                desiredVar);
    }
  }

Here is the call graph for this function:

convertNode2Leaf_()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::IncrementalGraphLearner< AttributeSelection, isScalar >::convertNode2Leaf_ ( NodeId  currentNodeId )

protectedvirtualinherited

Turns the given node into a leaf if not already so.

Definition at line 200 of file incrementalGraphLearner_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                           {
    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_);
    }
  }

Here is the call graph for this function:

insertInternalNode_()

template<TESTNAME AttributeSelection, bool isScalar>

NodeId gum::IncrementalGraphLearner< AttributeSelection, isScalar >::insertInternalNode_ ( NodeDatabase< AttributeSelection, isScalar > *  nDB, const DiscreteVariable *  boundVar, NodeId *  sonsMap  )

protectedvirtualinherited

inserts a new internal node in internal graph

Parameters

nDB	: the associated database
boundVar	: the associated variable
sonsMap	: a table giving node's sons node

Returns

the newly created node's id

Definition at line 362 of file incrementalGraphLearner_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                            {
    NodeId newNodeId = this->insertNode_(nDB, boundVar);
    nodeSonsMap_.insert(newNodeId, sonsMap);
    return newNodeId;
  }

Here is the call graph for this function:

insertLeafNode_()

template<TESTNAME AttributeSelection, bool isScalar>

NodeId gum::IncrementalGraphLearner< AttributeSelection, isScalar >::insertLeafNode_ ( NodeDatabase< AttributeSelection, isScalar > *  nDB, const DiscreteVariable *  boundVar, Set< const Observation * > *  obsSet  )

protectedvirtualinherited

inserts a new leaf node in internal graohs

Parameters

nDB	: the associated database
boundVar	: the associated variable
obsSet	: the set of observation this leaf retains

Returns

the newly created node's id

Reimplemented in gum::IMDDI< AttributeSelection, isScalar >.

Definition at line 382 of file incrementalGraphLearner_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                           {
    NodeId newNodeId = this->insertNode_(nDB, boundVar);
    leafDatabase_.insert(newNodeId, obsSet);
    return newNodeId;
  }

Here is the call graph for this function:

insertNode_()

template<TESTNAME AttributeSelection, bool isScalar>

NodeId gum::ITI< AttributeSelection, isScalar >::insertNode_ ( NodeDatabase< AttributeSelection, isScalar > *  nDB, const DiscreteVariable *  boundVar  )

protectedvirtual

inserts a new node in internal graph

inserts a new node in internal graohs

Parameters

nDB	: the associated database
boundVar	: the associated variable

Returns

the newly created node's id

Reimplemented from gum::IncrementalGraphLearner< AttributeSelection, isScalar >.

Definition at line 203 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                             {
    NodeId n = IncrementalGraphLearner< AttributeSelection, isScalar >::insertNode_(nDB, boundVar);
    _staleTable_.insert(n, true);
    return n;
  }

Here is the call graph for this function:

insertSetOfVars()

template<TESTNAME AttributeSelection, bool isScalar = false>

virtual void gum::IncrementalGraphLearner< AttributeSelection, isScalar >::insertSetOfVars ( MultiDimFunctionGraph< double > *  ret ) const

inlinevirtualinherited

Implements gum::IVisitableGraphLearner.

Reimplemented in gum::IMDDI< AttributeSelection, isScalar >.

Definition at line 326 of file incrementalGraphLearner.h.

                                                                             {
      for (SetIteratorSafe< const DiscreteVariable* > varIter = setOfVars_.beginSafe();
           varIter != setOfVars_.endSafe();
           ++varIter)
        ret->add(**varIter);
    }

insertSetOfVars_()

template<TESTNAME AttributeSelection, bool isScalar = false>

void gum::ITI< AttributeSelection, isScalar >::insertSetOfVars_ ( MultiDimFunctionGraph< double > *  ret )

inlineprotected

insertSetOfVars_

Parameters

ret

Definition at line 245 of file iti.h.

                                                                {
      for (SetIteratorSafe< const DiscreteVariable* > varIter = this->setOfVars_.beginSafe();
           varIter != this->setOfVars_.endSafe();
           ++varIter)
        ret->add(**varIter);
    }

isTerminal()

template<TESTNAME AttributeSelection, bool isScalar = false>

bool gum::IncrementalGraphLearner< AttributeSelection, isScalar >::isTerminal ( NodeId  ni ) const

inlinevirtualinherited

Implements gum::IVisitableGraphLearner.

Definition at line 306 of file incrementalGraphLearner.h.

{ return !this->nodeSonsMap_.exists(ni); }

nodeNbObservation()

template<TESTNAME AttributeSelection, bool isScalar = false>

Idx gum::IncrementalGraphLearner< AttributeSelection, isScalar >::nodeNbObservation ( NodeId  ni ) const

inlinevirtualinherited

Implements gum::IVisitableGraphLearner.

Definition at line 321 of file incrementalGraphLearner.h.

{ return this->nodeId2Database_[ni]->nbObservation(); }

nodeSon()

template<TESTNAME AttributeSelection, bool isScalar = false>

NodeId gum::IncrementalGraphLearner< AttributeSelection, isScalar >::nodeSon ( NodeId  ni, Idx  modality  ) const

inlinevirtualinherited

Implements gum::IVisitableGraphLearner.

Definition at line 316 of file incrementalGraphLearner.h.

{ return this->nodeSonsMap_[ni][modality]; }

nodeVar()

template<TESTNAME AttributeSelection, bool isScalar = false>

const DiscreteVariable* gum::IncrementalGraphLearner< AttributeSelection, isScalar >::nodeVar ( NodeId  ni ) const

inlinevirtualinherited

Implements gum::IVisitableGraphLearner.

Definition at line 311 of file incrementalGraphLearner.h.

{ return this->nodeVarMap_[ni]; }

removeNode_()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::ITI< AttributeSelection, isScalar >::removeNode_ ( NodeId  currentNodeId )

protectedvirtual

Removes a node from the internal graph.

Parameters

removedNodeId

: the node to remove

Reimplemented from gum::IncrementalGraphLearner< AttributeSelection, isScalar >.

Definition at line 237 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                            {
    IncrementalGraphLearner< AttributeSelection, isScalar >::removeNode_(currentNodeId);
    _staleTable_.erase(currentNodeId);
  }

Here is the call graph for this function:

root()

template<TESTNAME AttributeSelection, bool isScalar = false>

NodeId gum::IncrementalGraphLearner< AttributeSelection, isScalar >::root ( ) const

inlinevirtualinherited

Implements gum::IVisitableGraphLearner.

Definition at line 301 of file incrementalGraphLearner.h.

{ return this->root_; }

size()

template<TESTNAME AttributeSelection, bool isScalar = false>

Size gum::IncrementalGraphLearner< AttributeSelection, isScalar >::size ( )

inlineinherited

Definition at line 290 of file incrementalGraphLearner.h.

{ return nodeVarMap_.size(); }

transpose_()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::IncrementalGraphLearner< AttributeSelection, isScalar >::transpose_ ( NodeId  currentNodeId, const DiscreteVariable *  desiredVar  )

protectedvirtualinherited

Installs given variable to the given node, ensuring that the variable is not present in its subtree.

Definition at line 227 of file incrementalGraphLearner_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                         {
    // **************************************************************************************
    // 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);
  }

Here is the call graph for this function:

updateFunctionGraph()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::ITI< AttributeSelection, isScalar >::updateFunctionGraph ( )

virtual

Updates target to currently learned graph structure.

Implements gum::IncrementalGraphLearner< AttributeSelection, isScalar >.

Definition at line 251 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                {
    this->target_->clear();
    this->target_->manager()->setRootNode(this->_insertNodeInFunctionGraph_(this->root_));
  }

Here is the call graph for this function:

updateGraph()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::ITI< AttributeSelection, isScalar >::updateGraph ( )

virtual

Updates the internal graph after a new observation has been added.

Implements gum::IncrementalGraphLearner< AttributeSelection, isScalar >.

Definition at line 139 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                        {
    std::vector< NodeId > filo;
    filo.push_back(this->root_);
    HashTable< NodeId, Set< const DiscreteVariable* >* > potentialVars;
    potentialVars.insert(this->root_, new Set< const DiscreteVariable* >(this->setOfVars_));


    while (!filo.empty()) {
      NodeId currentNodeId = filo.back();
      filo.pop_back();

      // First we look for the best var to install on the node
      double                         bestValue = _attributeSelectionThreshold_;
      Set< const DiscreteVariable* > bestVars;

      for (auto varIter = potentialVars[currentNodeId]->cbeginSafe();
           varIter != potentialVars[currentNodeId]->cendSafe();
           ++varIter)
        if (this->nodeId2Database_[currentNodeId]->isTestRelevant(*varIter)) {
          double varValue = this->nodeId2Database_[currentNodeId]->testValue(*varIter);
          if (varValue >= bestValue) {
            if (varValue > bestValue) {
              bestValue = varValue;
              bestVars.clear();
            }
            bestVars.insert(*varIter);
          }
        }

      // Then We installed Variable a test on that node
      this->updateNode_(currentNodeId, bestVars);

      // The we move on the children if needed
      if (this->nodeVarMap_[currentNodeId] != this->value_) {
        for (Idx moda = 0; moda < this->nodeVarMap_[currentNodeId]->domainSize(); moda++) {
          Set< const DiscreteVariable* >* itsPotentialVars
             = new Set< const DiscreteVariable* >(*potentialVars[currentNodeId]);
          itsPotentialVars->erase(this->nodeVarMap_[currentNodeId]);
          NodeId sonId = this->nodeSonsMap_[currentNodeId][moda];
          if (_staleTable_[sonId]) {
            filo.push_back(sonId);
            potentialVars.insert(sonId, itsPotentialVars);
          }
        }
      }
    }

    for (HashTableIteratorSafe< NodeId, Set< const DiscreteVariable* >* > nodeIter
         = potentialVars.beginSafe();
         nodeIter != potentialVars.endSafe();
         ++nodeIter)
      delete nodeIter.val();
  }

Here is the call graph for this function:

updateNode_()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::IncrementalGraphLearner< AttributeSelection, isScalar >::updateNode_ ( NodeId  updatedNode, Set< const DiscreteVariable * > &  varsOfInterest  )

protectedinherited

From the given sets of node, selects randomly one and installs it on given node.

Chechks of course if node's current variable is not in that set first.

Parameters

nody	: the node we update
bestVars	: the set of interessting vars to be installed here

Chechks of course if node's current variable is not in that set first.

Parameters

nody	: the node we update
bestVar	: the set of interessting vars to be installed here

Definition at line 169 of file incrementalGraphLearner_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                     {
    // 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);
  }

Here is the call graph for this function:

updateNodeWithObservation_()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::ITI< AttributeSelection, isScalar >::updateNodeWithObservation_ ( const Observation *  newObs, NodeId  currentNodeId  )

protectedvirtual

Will update internal graph's NodeDatabase of given node with the new observation.

Parameters

newObs
currentNodeId

Reimplemented from gum::IncrementalGraphLearner< AttributeSelection, isScalar >.

Definition at line 122 of file iti_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                                                                             {
    IncrementalGraphLearner< AttributeSelection, isScalar >::updateNodeWithObservation_(
       newObs,
       currentNodeId);
    _staleTable_[currentNodeId] = true;
  }

Here is the call graph for this function:

updateVar()

template<TESTNAME AttributeSelection, bool isScalar>

void gum::IncrementalGraphLearner< AttributeSelection, isScalar >::updateVar ( const DiscreteVariable *  var )

virtualinherited

If a new modality appears to exists for given variable, call this method to turn every associated node to this variable into leaf. Graph has then indeed to be revised.

Definition at line 147 of file incrementalGraphLearner_tpl.h.

References gum::Set< Key, Alloc >::emplace().

                                  {
    Link< NodeId >* nodIter = var2Node_[var]->list();
    Link< NodeId >* nni     = nullptr;
    while (nodIter) {
      nni = nodIter->nextLink();
      convertNode2Leaf_(nodIter->element());
      nodIter = nni;
    }
  }

Here is the call graph for this function:

Member Data Documentation

_attributeSelectionThreshold_

template<TESTNAME AttributeSelection, bool isScalar = false>

double gum::ITI< AttributeSelection, isScalar >::_attributeSelectionThreshold_

private

The threshold above which we consider variables to be dependant.

Definition at line 262 of file iti.h.

_nbTotalObservation_

template<TESTNAME AttributeSelection, bool isScalar = false>

Idx gum::ITI< AttributeSelection, isScalar >::_nbTotalObservation_

private

The total number of observation added to this tree.

Definition at line 259 of file iti.h.

_staleTable_

template<TESTNAME AttributeSelection, bool isScalar = false>

HashTable< NodeId, bool > gum::ITI< AttributeSelection, isScalar >::_staleTable_

private

Hashtable indicating if given node has been modified (upon receiving new exemple or through a transpose) The aim is not if we have revise the installed variable on that node.

Definition at line 256 of file iti.h.

leafDatabase_

template<TESTNAME AttributeSelection, bool isScalar = false>

HashTable< NodeId, Set< const Observation* >* > gum::IncrementalGraphLearner< AttributeSelection, isScalar >::leafDatabase_

protectedinherited

This hashtable binds to every leaf an associated set of all hte observations compatible with it.

Definition at line 379 of file incrementalGraphLearner.h.

model_

template<TESTNAME AttributeSelection, bool isScalar = false>

NodeGraphPart gum::IncrementalGraphLearner< AttributeSelection, isScalar >::model_

protectedinherited

The source of nodeId.

Definition at line 345 of file incrementalGraphLearner.h.

needUpdate_

template<TESTNAME AttributeSelection, bool isScalar = false>

bool gum::IncrementalGraphLearner< AttributeSelection, isScalar >::needUpdate_

protectedinherited

Definition at line 392 of file incrementalGraphLearner.h.

nodeId2Database_

template<TESTNAME AttributeSelection, bool isScalar = false>

HashTable< NodeId, NodeDatabase< AttributeSelection, isScalar >* > gum::IncrementalGraphLearner< AttributeSelection, isScalar >::nodeId2Database_

protectedinherited

This hashtable binds every node to an associated NodeDatabase which handles every observation that concerns that node.

Definition at line 373 of file incrementalGraphLearner.h.

nodeSonsMap_

template<TESTNAME AttributeSelection, bool isScalar = false>

HashTable< NodeId, NodeId* > gum::IncrementalGraphLearner< AttributeSelection, isScalar >::nodeSonsMap_

protectedinherited

A table giving for any node a table mapping to its son idx is the modality of associated variable.

Definition at line 361 of file incrementalGraphLearner.h.

nodeVarMap_

template<TESTNAME AttributeSelection, bool isScalar = false>

HashTable< NodeId, const DiscreteVariable* > gum::IncrementalGraphLearner< AttributeSelection, isScalar >::nodeVarMap_

protectedinherited

Gives for any node its associated variable.

Definition at line 355 of file incrementalGraphLearner.h.

root_

template<TESTNAME AttributeSelection, bool isScalar = false>

NodeId gum::IncrementalGraphLearner< AttributeSelection, isScalar >::root_

protectedinherited

The root of the ordered tree.

Definition at line 350 of file incrementalGraphLearner.h.

setOfVars_

template<TESTNAME AttributeSelection, bool isScalar = false>

Set< const DiscreteVariable* > gum::IncrementalGraphLearner< AttributeSelection, isScalar >::setOfVars_

protectedinherited

Definition at line 387 of file incrementalGraphLearner.h.

target_

template<TESTNAME AttributeSelection, bool isScalar = false>

MultiDimFunctionGraph< double >* gum::IncrementalGraphLearner< AttributeSelection, isScalar >::target_

protectedinherited

The final diagram we're building.

Definition at line 385 of file incrementalGraphLearner.h.

value_

template<TESTNAME AttributeSelection, bool isScalar = false>

const DiscreteVariable* gum::IncrementalGraphLearner< AttributeSelection, isScalar >::value_

protectedinherited

Definition at line 389 of file incrementalGraphLearner.h.

valueAssumed_

template<TESTNAME AttributeSelection, bool isScalar = false>

Sequence< ValueType > gum::IncrementalGraphLearner< AttributeSelection, isScalar >::valueAssumed_

protectedinherited

Definition at line 390 of file incrementalGraphLearner.h.

var2Node_

template<TESTNAME AttributeSelection, bool isScalar = false>

HashTable< const DiscreteVariable*, LinkedList< NodeId >* > gum::IncrementalGraphLearner< AttributeSelection, isScalar >::var2Node_

protectedinherited

Associates to any variable the list of all nodes associated to this variable.

Definition at line 367 of file incrementalGraphLearner.h.

---

The documentation for this class was generated from the following files:

• agrum/FMDP/learning/datastructure/iti.h
• agrum/FMDP/learning/datastructure/iti_tpl.h