Class used to perform Function Graph Operations in the FMDP Framework. More...

#include <agrum/tools/multidim/patterns/regress.h>

Collaboration diagram for gum::Regress< GUM_SCALAR, COMBINEOPERATOR, PROJECTOPERATOR, TerminalNodePolicy >:

Public Member Functions
Constructors / Destructors
	Regress (const MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy > vfunction, const MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy > probDist, const Set< const DiscreteVariable * > primedVars, const DiscreteVariable targetVar, const GUM_SCALAR neutral)
	Default constructor. More...

	~Regress ()
	Default destructor. More...

Main Method
MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy > *	compute ()
	Computes and builds the Function Graph that is the result of the operation. More...

Detailed Description

template<typename GUM_SCALAR, template< typename > class COMBINEOPERATOR, template< typename > class PROJECTOPERATOR, template< typename > class TerminalNodePolicy = ExactTerminalNodePolicy>
class gum::Regress< GUM_SCALAR, COMBINEOPERATOR, PROJECTOPERATOR, TerminalNodePolicy >

Class used to perform Function Graph Operations in the FMDP Framework.

Definition at line 54 of file regress.h.

Constructor & Destructor Documentation

◆ Regress()

template<typename GUM_SCALAR , template< typename > class COMBINEOPERATOR, template< typename > class PROJECTOPERATOR, template< typename > class TerminalNodePolicy>

INLINE gum::Regress< GUM_SCALAR, COMBINEOPERATOR, PROJECTOPERATOR, TerminalNodePolicy >::Regress	(	const MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy > *	vfunction,
		const MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy > *	probDist,
		const Set< const DiscreteVariable * > *	primedVars,
		const DiscreteVariable *	targetVar,
		const GUM_SCALAR	neutral
	)

Default constructor.

Definition at line 48 of file regress_tpl.h.

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

                                                                :
       DG1__(DG1),
       DG2__(DG2), neutral__(neutral), combine__(), project__(),
       DG1InstantiationNeeded__(DG1->realSize(), true, false),
       DG2InstantiationNeeded__(DG2->realSize(), true, false) {
     GUM_CONSTRUCTOR(Regress);
     rd__ = MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy >::
        getReducedAndOrderedInstance();
     nbVar__      = 0;
     default__    = nullptr;
     primedVars__ = primedVars;
     targetVar__  = targetVar;
   }

Here is the call graph for this function:

◆ ~Regress()

template<typename GUM_SCALAR , template< typename > class COMBINEOPERATOR, template< typename > class PROJECTOPERATOR, template< typename > class TerminalNodePolicy>

INLINE gum::Regress< GUM_SCALAR, COMBINEOPERATOR, PROJECTOPERATOR, TerminalNodePolicy >::~Regress ( )

Default destructor.

Definition at line 75 of file regress_tpl.h.

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

                    {
     GUM_DESTRUCTOR(Regress);
 
     for (auto instIter = DG1InstantiationNeeded__.beginSafe();
          instIter != DG1InstantiationNeeded__.endSafe();
          ++instIter)
       DEALLOCATE(instIter.val(), sizeof(short int) * nbVar__);
 
     for (auto instIter = DG2InstantiationNeeded__.beginSafe();
          instIter != DG2InstantiationNeeded__.endSafe();
          ++instIter)
       DEALLOCATE(instIter.val(), sizeof(short int) * nbVar__);
 
     if (nbVar__ != 0) DEALLOCATE(default__, sizeof(short int) * nbVar__);
   }

Here is the call graph for this function:

Member Function Documentation

◆ compute()

template<typename GUM_SCALAR , template< typename > class COMBINEOPERATOR, template< typename > class PROJECTOPERATOR, template< typename > class TerminalNodePolicy>

INLINE MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy > * gum::Regress< GUM_SCALAR, COMBINEOPERATOR, PROJECTOPERATOR, TerminalNodePolicy >::compute ( )

Computes and builds the Function Graph that is the result of the operation.

Definition at line 103 of file regress_tpl.h.

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

                   {
     establishVarOrder__();
     findRetrogradeVariables__(DG1__, DG1InstantiationNeeded__);
     findRetrogradeVariables__(DG2__, DG2InstantiationNeeded__);
 
     Idx* varInst = nullptr;
     if (nbVar__ != 0) {
       varInst = static_cast< Idx* >(ALLOCATE(sizeof(Idx) * nbVar__));
       for (Idx i = 0; i < nbVar__; i++)
         varInst[i] = (Idx)0;
     }
 
     O4DGContext conti(varInst, nbVar__);
     conti.setDG1Node(DG1__->root());
     conti.setDG2Node(DG2__->root());
 
     NodeId root = compute__(conti, (Idx)0 - 1);
     rd__->manager()->setRootNode(root);
 
     if (nbVar__ != 0) DEALLOCATE(varInst, sizeof(Idx) * nbVar__);
 
     rd__->erase(*targetVar__);
 
     return rd__;
   }

Here is the call graph for this function:

◆ compute__()

template<typename GUM_SCALAR , template< typename > class COMBINEOPERATOR, template< typename > class PROJECTOPERATOR, template< typename > class TerminalNodePolicy>

INLINE NodeId gum::Regress< GUM_SCALAR, COMBINEOPERATOR, PROJECTOPERATOR, TerminalNodePolicy >::compute__	(	O4DGContext &	currentSituation,
		Idx	lastInstVarPos
	)

private

The main recursion function.

Definition at line 327 of file regress_tpl.h.

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

                                                                      {
     NodeId newNode = 0;
 
     // If both current nodes are terminal,
     // we only have to compute the resulting value
     if (DG1__->isTerminalNode(currentSituation.DG1Node())
         && DG2__->isTerminalNode(currentSituation.DG2Node())) {
       // We have to compute new valueand we insert a new node in diagram with
       // this value, ...
       GUM_SCALAR newVal  = neutral__;
       GUM_SCALAR tempVal = combine__(DG1__->nodeValue(currentSituation.DG1Node()),
                                      DG2__->nodeValue(currentSituation.DG2Node()));
       for (Idx targetModa = 0; targetModa < targetVar__->domainSize();
            ++targetModa)
         newVal = project__(newVal, tempVal);
       return rd__->manager()->addTerminalNode(newVal);
     }
 
     // If not,
     // we'll have to do some exploration
 
     // First we ensure that we hadn't already visit this pair of node under hte
     // same circumstances
     short int* dg1NeededVar
        = DG1InstantiationNeeded__.exists(currentSituation.DG1Node())
           ? DG1InstantiationNeeded__[currentSituation.DG1Node()]
           : default__;
     Idx        dg1CurrentVarPos = DG1__->isTerminalNode(currentSituation.DG1Node())
                                    ? nbVar__
                                    : rd__->variablesSequence().pos(
                                DG1__->node(currentSituation.DG1Node())->nodeVar());
     short int* dg2NeededVar
        = DG2InstantiationNeeded__.exists(currentSituation.DG2Node())
           ? DG2InstantiationNeeded__[currentSituation.DG2Node()]
           : default__;
     Idx dg2CurrentVarPos = DG2__->isTerminalNode(currentSituation.DG2Node())
                             ? nbVar__
                             : rd__->variablesSequence().pos(
                                DG2__->node(currentSituation.DG2Node())->nodeVar());
 
     short int* instNeeded
        = static_cast< short int* >(ALLOCATE(sizeof(short int) * nbVar__));
 
     for (Idx i = 0; i < nbVar__; i++) {
       instNeeded[i] = dg1NeededVar[i] + dg2NeededVar[i];
     }
 
     double curSitKey = currentSituation.key(instNeeded);
 
     if (explorationTable__.exists(curSitKey)) {
       DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
 
       return explorationTable__[curSitKey];
     }
 
     // ====================================================
 
     NodeId origDG1 = currentSituation.DG1Node(),
            origDG2 = currentSituation.DG2Node();
 
     const MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy >* leaddg
        = nullptr;
     NodeId leadNodeId = 0;
     Idx    leadVarPos = rd__->variablesSequence().size();
     typedef void (O4DGContext::*SetNodeFunction)(const NodeId&);
     SetNodeFunction leadFunction = nullptr;
 
     bool sameVar = false;
 
     if (!DG1__->isTerminalNode(currentSituation.DG1Node())) {
       if (currentSituation.varModality(dg1CurrentVarPos) != 0) {
         // If var associated to current node has already been instanciated, we
         // have to jump it
         currentSituation.setDG1Node(
            DG1__->node(currentSituation.DG1Node())
               ->son(currentSituation.varModality(dg1CurrentVarPos) - 1));
 
         newNode = compute__(currentSituation, lastInstVarPos);
         explorationTable__.insert(curSitKey, newNode);
         currentSituation.setDG1Node(origDG1);
         currentSituation.setDG2Node(origDG2);
 
         DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
 
         return newNode;
       }
 
       leaddg       = DG1__;
       leadNodeId   = currentSituation.DG1Node();
       leadVarPos   = dg1CurrentVarPos;
       leadFunction = &O4DGContext::setDG1Node;
     }
 
     if (!DG2__->isTerminalNode(currentSituation.DG2Node())) {
       if (currentSituation.varModality(dg2CurrentVarPos) != 0) {
         // If var associated to current node has already been instanciated, we
         // have to jump it
         currentSituation.setDG2Node(
            DG2__->node(currentSituation.DG2Node())
               ->son(currentSituation.varModality(dg2CurrentVarPos) - 1));
 
         newNode = compute__(currentSituation, lastInstVarPos);
         explorationTable__.insert(curSitKey, newNode);
         currentSituation.setDG1Node(origDG1);
         currentSituation.setDG2Node(origDG2);
 
         DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
 
         return newNode;
       }
 
       if (leadVarPos == dg2CurrentVarPos) { sameVar = true; }
 
       if (leadVarPos > dg2CurrentVarPos) {
         leaddg       = DG2__;
         leadNodeId   = currentSituation.DG2Node();
         leadVarPos   = dg2CurrentVarPos;
         leadFunction = &O4DGContext::setDG2Node;
       }
     }
 
     // ====================================================
     // Anticipated Exploration
 
     // Before exploring nodes, we have to ensure that every anticipated
     // exploration is done
     for (Idx varPos = lastInstVarPos + 1; varPos < leadVarPos; ++varPos) {
       if (instNeeded[varPos]) {
         const DiscreteVariable* curVar  = rd__->variablesSequence().atPos(varPos);
         NodeId*                 sonsIds = static_cast< NodeId* >(
            ALLOCATE(sizeof(NodeId) * curVar->domainSize()));
 
         for (Idx modality = 0; modality < curVar->domainSize(); modality++) {
           currentSituation.chgVarModality(varPos, modality + 1);
 
           sonsIds[modality] = compute__(currentSituation, varPos);
         }
 
         newNode = rd__->manager()->addInternalNode(curVar, sonsIds);
 
         explorationTable__.insert(curSitKey, newNode);
         currentSituation.chgVarModality(varPos, 0);
         currentSituation.setDG1Node(origDG1);
         currentSituation.setDG2Node(origDG2);
 
         DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
 
         return newNode;
       }
     }
 
     // ====================================================
     // Terminal Exploration
     if (sameVar && DG1__->node(origDG1)->nodeVar() == targetVar__) {
       GUM_SCALAR newVal = neutral__;
       for (Idx targetModa = 0; targetModa < targetVar__->domainSize();
            ++targetModa)
         newVal = project__(
            newVal,
            combine__(DG1__->nodeValue(DG1__->node(origDG1)->son(targetModa)),
                      DG2__->nodeValue(DG2__->node(origDG2)->son(targetModa))));
       newNode = rd__->manager()->addTerminalNode(newVal);
       explorationTable__.insert(curSitKey, newNode);
       DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
       return newNode;
     }
     if (DG1__->isTerminalNode(origDG1)) {
       if (DG2__->node(origDG2)->nodeVar() == targetVar__) {
         GUM_SCALAR newVal = neutral__;
         for (Idx targetModa = 0; targetModa < targetVar__->domainSize();
              ++targetModa)
           newVal = project__(
              newVal,
              combine__(DG1__->nodeValue(origDG1),
                        DG2__->nodeValue(DG2__->node(origDG2)->son(targetModa))));
         newNode = rd__->manager()->addTerminalNode(newVal);
         explorationTable__.insert(curSitKey, newNode);
         DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
         return newNode;
       }
     } else {
       if (DG1__->node(origDG1)->nodeVar() == targetVar__
           && DG2__->isTerminalNode(origDG2)) {
         GUM_SCALAR newVal = neutral__;
         for (Idx targetModa = 0; targetModa < targetVar__->domainSize();
              ++targetModa)
           newVal = project__(
              newVal,
              combine__(DG1__->nodeValue(DG1__->node(origDG1)->son(targetModa)),
                        DG2__->nodeValue(origDG2)));
         newNode = rd__->manager()->addTerminalNode(newVal);
         explorationTable__.insert(curSitKey, newNode);
         DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
         return newNode;
       }
     }
 
     // ====================================================
     // Normal Exploration
 
     // If only one of the current node is terminal,
     // we have to pursue deeper on the other diagram
     if (sameVar) {
       // If so - meaning it's the same variable - we have to go
       // down on both
       const InternalNode* dg1Node = DG1__->node(origDG1);
       const InternalNode* dg2Node = DG2__->node(origDG2);
 
       const DiscreteVariable* curVar = dg1Node->nodeVar();
       Idx                     varPos = rd__->variablesSequence().pos(curVar);
       NodeId*                 sonsIds
          = static_cast< NodeId* >(ALLOCATE(sizeof(NodeId) * curVar->domainSize()));
 
       for (Idx modality = 0; modality < curVar->domainSize(); modality++) {
         currentSituation.chgVarModality(varPos, modality + 1);
         currentSituation.setDG1Node(dg1Node->son(modality));
         currentSituation.setDG2Node(dg2Node->son(modality));
 
         sonsIds[modality] = compute__(currentSituation, varPos);
       }
 
       newNode = rd__->manager()->addInternalNode(curVar, sonsIds);
 
       explorationTable__.insert(curSitKey, newNode);
       currentSituation.chgVarModality(varPos, 0);
       currentSituation.setDG1Node(origDG1);
       currentSituation.setDG2Node(origDG2);
 
       DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
 
       return newNode;
     }
     // ====================================================
     else {
       const InternalNode* leaddgNode = leaddg->node(leadNodeId);
 
       const DiscreteVariable* curVar = leaddgNode->nodeVar();
       NodeId*                 sonsIds
          = static_cast< NodeId* >(ALLOCATE(sizeof(NodeId) * curVar->domainSize()));
 
       for (Idx modality = 0; modality < curVar->domainSize(); modality++) {
         currentSituation.chgVarModality(leadVarPos, modality + 1);
         (currentSituation.*leadFunction)(leaddgNode->son(modality));
 
         sonsIds[modality] = compute__(currentSituation, leadVarPos);
       }
 
       newNode = rd__->manager()->addInternalNode(curVar, sonsIds);
 
       explorationTable__.insert(curSitKey, newNode);
       currentSituation.chgVarModality(leadVarPos, 0);
       currentSituation.setDG1Node(origDG1);
       currentSituation.setDG2Node(origDG2);
 
       DEALLOCATE(instNeeded, sizeof(short int) * nbVar__);
 
       return newNode;
     }
   }

Here is the call graph for this function:

◆ establishVarOrder__()

template<typename GUM_SCALAR , template< typename > class COMBINEOPERATOR, template< typename > class PROJECTOPERATOR, template< typename > class TerminalNodePolicy>

INLINE void gum::Regress< GUM_SCALAR, COMBINEOPERATOR, PROJECTOPERATOR, TerminalNodePolicy >::establishVarOrder__ ( )

private

Computes an order for the final Decision graph that will minimize the number of re exploration.

Definition at line 141 of file regress_tpl.h.

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

                               {
     SequenceIteratorSafe< const DiscreteVariable* > fite
        = DG1__->variablesSequence().beginSafe();
     SequenceIteratorSafe< const DiscreteVariable* > site
        = DG2__->variablesSequence().beginSafe();
 
     while (fite != DG1__->variablesSequence().endSafe()
            && site != DG2__->variablesSequence().endSafe()) {
       // Test : if var from first order is already in final order
       // we move onto the next one
       if (rd__->variablesSequence().exists(*fite)) {
         ++fite;
         continue;
       }
 
       // Test : if var from second order is already in final order
       // we move onto the next one
       if (rd__->variablesSequence().exists(*site)) {
         ++site;
         continue;
       }
 
       // Test : is current var of the first order present in the second order.
       // if not we add it to final order
       if (!DG2__->variablesSequence().exists(*fite)
           && !primedVars__->exists(*fite)) {
         rd__->add(**fite);
         ++fite;
         continue;
       }
 
       // Test : is current var of the second order present in the first order.
       // if not we add it to final order
       if (!DG1__->variablesSequence().exists(*site)
           && !primedVars__->exists(*site)) {
         rd__->add(**site);
         ++site;
         continue;
       }
 
       // Test : is current var of the second order present in the first order.
       // if not we add it to final order
       if (*fite == *site) {
         rd__->add(**fite);
         ++fite;
         ++site;
         continue;
       }
 
       // Test : if chosing first order var cost less in terms or re exploration,
       // we chose it
       rd__->add(**fite);
       ++fite;
     }
 
     // Whenever an iterator has finished its sequence,
     // the other may still be in the middle of its one.
     // Hence, this part ensures that any variables remaining
     // will be added to the final sequence if needed.
     if (fite == DG1__->variablesSequence().endSafe()) {
       for (; site != DG2__->variablesSequence().endSafe(); ++site)
         if (!rd__->variablesSequence().exists(*site)) rd__->add(**site);
     } else {
       for (; fite != DG1__->variablesSequence().endSafe(); ++fite)
         if (!rd__->variablesSequence().exists(*fite)) rd__->add(**fite);
     }
 
     // Various initialization needed now that we have a bigger picture
     nbVar__ = rd__->variablesSequence().size();
 
     if (nbVar__ != 0) {
       default__ = static_cast< short int* >(ALLOCATE(sizeof(short int) * nbVar__));
       for (Idx i = 0; i < nbVar__; i++)
         default__[i] = (short int)0;
     }
   }

Here is the call graph for this function:

◆ findRetrogradeVariables__()

template<typename GUM_SCALAR , template< typename > class COMBINEOPERATOR, template< typename > class PROJECTOPERATOR, template< typename > class TerminalNodePolicy>

INLINE void gum::Regress< GUM_SCALAR, COMBINEOPERATOR, PROJECTOPERATOR, TerminalNodePolicy >::findRetrogradeVariables__	(	const MultiDimFunctionGraph< GUM_SCALAR, TerminalNodePolicy > *	dg,
		HashTable< NodeId, short int * > &	dgInstNeed
	)

private

Establish for each node in both function graph if it has retrograde variables beneath it.

Definition at line 229 of file regress_tpl.h.

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

                                                         {
     HashTable< NodeId, short int* > nodesVarDescendant;
     Size                            tableSize = Size(nbVar__ * sizeof(short int));
 
     for (auto varIter = dg->variablesSequence().rbeginSafe();
          varIter != dg->variablesSequence().rendSafe();
          --varIter) {
       Idx varPos = rd__->variablesSequence().pos(*varIter);
 
       const Link< NodeId >* nodeIter = dg->varNodeListe(*varIter)->list();
       while (nodeIter != nullptr) {
         short int* instantiationNeeded
            = static_cast< short int* >(ALLOCATE(tableSize));
         dgInstNeed.insert(nodeIter->element(), instantiationNeeded);
         short int* varDescendant = static_cast< short int* >(ALLOCATE(tableSize));
         nodesVarDescendant.insert(nodeIter->element(), varDescendant);
         for (Idx j = 0; j < nbVar__; j++) {
           instantiationNeeded[j] = (short int)0;
           varDescendant[j]       = (short int)0;
         }
 
 
         varDescendant[varPos] = (short int)1;
         for (Idx modality = 0; modality < dg->node(nodeIter->element())->nbSons();
              ++modality) {
           if (!dg->isTerminalNode(dg->node(nodeIter->element())->son(modality))) {
             short int* sonVarDescendant
                = nodesVarDescendant[dg->node(nodeIter->element())->son(modality)];
             for (Idx varIdx = 0; varIdx < nbVar__; varIdx++) {
               varDescendant[varIdx] += sonVarDescendant[varIdx];
               if (varDescendant[varIdx] && varIdx < varPos)
                 instantiationNeeded[varIdx] = (short int)1;
             }
           }
         }
         nodeIter = nodeIter->nextLink();
       }
     }
 
     for (auto varIter = dg->variablesSequence().beginSafe();
          varIter != dg->variablesSequence().endSafe();
          ++varIter) {
       const Link< NodeId >* nodeIter = dg->varNodeListe(*varIter)->list();
       while (nodeIter != nullptr) {
         for (Idx modality = 0; modality < dg->node(nodeIter->element())->nbSons();
              ++modality) {
           NodeId sonId = dg->node(nodeIter->element())->son(modality);
           if (!dg->isTerminalNode(sonId)) {
             for (Idx varIdx = 0; varIdx < nbVar__; ++varIdx) {
               if (dgInstNeed[nodeIter->element()][varIdx]
                   && nodesVarDescendant[sonId][varIdx]) {
                 dgInstNeed[sonId][varIdx] = (short int)1;
               }
             }
           }
         }
         nodeIter = nodeIter->nextLink();
       }
     }
 
     for (HashTableIterator< NodeId, short int* > it = nodesVarDescendant.begin();
          it != nodesVarDescendant.end();
          ++it) {
       DEALLOCATE(it.val(), tableSize);
     }
 
     nodesVarDescendant.clear();
   }

Here is the call graph for this function: