drwnSontag08Inference Class Reference

Implements the incremental tightening of the LP MAP inference algorithm from Sontag et al., UAI 2008. More...

Inheritance diagram for drwnSontag08Inference:

Public Member Functions
	drwnSontag08Inference (const drwnFactorGraph &graph)
void	clear ()
	Clear internally cached data (e.g., computation graph)
std::pair< double, double >	inference (drwnFullAssignment &mapAssignment)
	Run inference (or resume for iterative algorithms). Algorithms may initialize from mapAssignment if not empty. Returns an upper and lower bound (if available) of the minimum energy. The upper bound is the same as the energy of the best solution found (i.e., same as graph.getEnergy(mapAssignment)). More...
Public Member Functions inherited from drwnGEMPLPInference
	drwnGEMPLPInference (const drwnFactorGraph &graph)
Public Member Functions inherited from drwnMessagePassingMAPInference
	drwnMessagePassingMAPInference (const drwnFactorGraph &graph)
Public Member Functions inherited from drwnMAPInference
	drwnMAPInference (const drwnFactorGraph &graph)
	drwnMAPInference (const drwnMAPInference &inf)

Static Public Attributes
static unsigned	WARMSTART_ITERATIONS = 20
	maximum number of iterations < after adding clusters
static unsigned	MAX_CLIQUES_TO_ADD = 5
	number of cliques to add per cycle
Static Public Attributes inherited from drwnMessagePassingMAPInference
static unsigned	MAX_ITERATIONS = 1000
	maximum number of iterations
static double	DAMPING_FACTOR = 0.0
	damping factor for updating messages

Protected Member Functions
void	buildComputationGraph ()
virtual void	findCliqueCandidates (std::map< drwnClique, std::vector< int > > &cliqueCandidateSet)
	Generates the set of clique candidates to test after each GEMPLP iteration. Derived classes can override this to implement different clique candidate strategies.
Protected Member Functions inherited from drwnGEMPLPInference
void	initializeMessages ()
void	decodeBeliefs (drwnFullAssignment &mapAssignment)
int	findSeparatorIndex (const drwnClique &cliqueA, const drwnClique &cliqueB)
void	addMessageUpdate (int cliqueId, const drwnClique &cliqueVars, const drwnTableFactor *psi=NULL)

Protected Attributes
std::vector< drwnClique >	_additionalCliques
Protected Attributes inherited from drwnGEMPLPInference
std::vector< drwnClique >	_separators
std::vector< drwnEdge >	_edges
std::vector< std::set< int > >	_cliqueEdges
std::vector< std::set< int > >	_separatorEdges
double	_lastDualObjective
unsigned	_maxIterations
Protected Attributes inherited from drwnMessagePassingMAPInference
std::vector< drwnTableFactor * >	_forwardMessages
std::vector< drwnTableFactor * >	_backwardMessages
std::vector< drwnTableFactor * >	_oldForwardMessages
std::vector< drwnTableFactor * >	_oldBackwardMessages
std::vector< drwnTableFactor * >	_intermediateFactors
std::vector < drwnFactorOperation * >	_computations
std::vector < drwnTableFactorStorage * >	_sharedStorage
Protected Attributes inherited from drwnMAPInference
const drwnFactorGraph &	_graph
	reference to initial clique potentials

Detailed Description

Implements the incremental tightening of the LP MAP inference algorithm from Sontag et al., UAI 2008.

Member Function Documentation

pair< double, double > drwnSontag08Inference::inference ( drwnFullAssignment &  mapAssignment )

virtual

Run inference (or resume for iterative algorithms). Algorithms may initialize from mapAssignment if not empty. Returns an upper and lower bound (if available) of the minimum energy. The upper bound is the same as the energy of the best solution found (i.e., same as graph.getEnergy(mapAssignment)).

Todo:

add this to computation graph

Reimplemented from drwnGEMPLPInference.

---

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

• drwnMapInference.h
• drwnMapInference.cpp