DiscriminativeLearner Class Reference

Discriminative learning algorithms (see "Discriminative Training of Markov Logic Networks", Singla and Domingos, 2005 and "Efficient Weight Learning for Markov Logic Networks", Lowd and Domingos, 2007). More...

#include <discriminativelearner.h>

List of all members.

Public Member Functions
	DiscriminativeLearner (const Array< Inference * > &inferences, const StringHashArray &nonEvidPredNames, IndexTranslator *const &idxTrans, const bool &lazyInference, const bool &withEM, const bool &rescaleGradient, const int &method, const double &lambda, const bool &preconditionCG, const double &maxLambda)
	Constructor.
	~DiscriminativeLearner ()
void	setMeansStdDevs (const int &arrSize, const double *const &priorMeans, const double *const &priorStdDevs)
void	setMLNWeights (double *const &weights)
void	setLogOddsWeights (double *weights, int numWeights)
void	getVariance (Array< double > &variance, int numWeights)
const Array< double > *	getHessianVectorProduct (const Array< double > &d)
double	computeQuadraticStepLength (double *gradient, const Array< double > &d, const Array< double > *Hd, double lambda)
void	learnWeights (double *const &weights, const int &numWeights, const int &maxIter, const double &maxSec, const double &learningRate, const double &momentum, bool initWithLogOdds, const int &mwsMaxSubsequentSteps, bool periodicMLNs)
Static Public Attributes
static const int	SIMPLE = 0
static const int	DN = 2
static const int	CG = 3

---

Detailed Description

Discriminative learning algorithms (see "Discriminative Training of Markov Logic Networks", Singla and Domingos, 2005 and "Efficient Weight Learning for Markov Logic Networks", Lowd and Domingos, 2007).

Definition at line 131 of file discriminativelearner.h.

---

Constructor & Destructor Documentation

DiscriminativeLearner::DiscriminativeLearner	(	const Array< Inference * > &	inferences,
		const StringHashArray &	nonEvidPredNames,
		IndexTranslator *const &	idxTrans,
		const bool &	lazyInference,
		const bool &	withEM,
		const bool &	rescaleGradient,
		const int &	method,
		const double &	lambda,
		const bool &	preconditionCG,
		const double &	maxLambda
	)			[inline]

Constructor.

Various variables are initialized, relevant clauses are determined and weights and inference procedures are initialized.

Parameters:

	inferences	Array of inference procedures to be used for inference in each domain.
	nonEvidPredNames	Names of non-evidence predicates. This is used to determine the relevant clauses.
	idxTrans	IndexTranslator needed when multiple dbs are used and they don't line up.
	lazyInference	If true, lazy inference is used.
	withEM	If true, EM is used to fill in missing values.
	rescaleGradient	If true, use per-weight learning rates
	method	Determines how direction and step size are chosen
	lambda	Initial value of lambda for SMD or CG
	preconditionCG	Whether or not to use a preconditioner with scaled conjugate gradient
	maxLambda	Maximum value of lambda for CG

Definition at line 154 of file discriminativelearner.h.

References Array< Type >::append(), Array< Type >::growToSize(), and Array< Type >::size().

    : domainCnt_(inferences.size()), idxTrans_(idxTrans),
      lazyInference_(lazyInference), rescaleGradient_(rescaleGradient),
      method_(method),
      // HACK: for now, we'll use the SMD lambda value for CG, even
      // though the two represent *very* different things!
      cg_lambda_(lambda), preconditionCG_(preconditionCG),
      maxBacktracks_(1000), backtrackCount_(0),
      cg_max_lambda_(maxLambda), withEM_(withEM) 
  { 
    cout << endl << "Constructing discriminative learner..." << endl << endl;

    inferences_.append(inferences);
    logOddsPerDomain_.growToSize(domainCnt_);
    clauseCntPerDomain_.growToSize(domainCnt_);
    
    for (int i = 0; i < domainCnt_; i++)
    {
      clauseCntPerDomain_[i] =
        inferences_[i]->getState()->getMLN()->getNumClauses();
      logOddsPerDomain_[i].growToSize(clauseCntPerDomain_[i], 0);
    }

    totalTrueCnts_.growToSize(domainCnt_);
    totalFalseCnts_.growToSize(domainCnt_);
    defaultTrueCnts_.growToSize(domainCnt_);
    defaultFalseCnts_.growToSize(domainCnt_);
    relevantClausesPerDomain_.growToSize(domainCnt_);
    //relevantClausesFormulas_ is set in findRelevantClausesFormulas()

    findRelevantClauses(nonEvidPredNames);
    findRelevantClausesFormulas();

      // Initialize the clause wts
    initializeWts(nonEvidPredNames);
    
      // Initialize the inference / state
    for (int i = 0; i < inferences_.size(); i++)
      inferences_[i]->init();
  }

---

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

• src/learnwts/discriminativelearner.h

---