tuffy.learn
Class DNLearner

java.lang.Object
  tuffy.main.Infer
      tuffy.learn.Learner
          tuffy.learn.DNLearner

public class DNLearner
extends Learner

Learner instance using diagonal Newton with dynamic step size. Method used here follows: Lowd, Daniel and Domingos, Pedro (2007). Efficient Weight Learning for Markov Logic Networks.

Author:

Ce Zhang

Field Summary

double	alpha Step size.
java.util.HashMap<java.lang.String,java.lang.Double>	currentD Map from clause ID to current H^(-1)g value.
java.util.HashMap<java.lang.String,java.lang.Double>	currentGradient Map from clause ID to current gradient value.
double	lambda Lambda used to control the step size.
int	nCall Number of invocations of DNLearner#updateWeight(MCSAT).
java.util.HashMap<java.lang.String,java.lang.Double>	oldD Map from clause ID to H^(-1)g value in last iteration.
double	oldDG D'g value of last iteration.
double	oldDHD D'HD value of last iteration.
java.util.HashMap<java.lang.String,java.lang.Double>	oldG Map from clause ID to gradient value in last iteration.
java.util.HashMap<java.lang.String,java.lang.Double>	oldWeight Map from clause ID to weight in last iteration.

Fields inherited from class tuffy.learn.Learner

_oldWeight, backtrackCount_, backtracked, currentWeight, finalWeight, isHardMappings, negativeWeightSatisfication, negativeWeightViolation, odds, oriWeight, positiveWeightSatisfication, positiveWeightViolation, trainingSatisification, trainingViolation

Fields inherited from class tuffy.main.Infer

db, dmover, grounding, mln, options

Constructor Summary

DNLearner()

Method Summary

void	getGradientAndD(MRF mcsat) Calculate the gradient and H^(-1)g by filling in currentD and currentGradient.
void	loadingTrainingData(MRF _mcsat) Load training data's truth value into MRF.
boolean	updateWeight(MRF mcsat) Update Learner.currentWeight using Diagonal Newton method.

Methods inherited from class tuffy.learn.Learner

calcCurrentTrainingViolation, dumpAnswers, fillInCurrentWeight, run

Methods inherited from class tuffy.main.Infer

cleanUp, ground, loadMLN, setUp

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

oldG

public java.util.HashMap<java.lang.String,java.lang.Double> oldG

Map from clause ID to gradient value in last iteration.

oldD

public java.util.HashMap<java.lang.String,java.lang.Double> oldD

Map from clause ID to H^(-1)g value in last iteration.

oldWeight

public java.util.HashMap<java.lang.String,java.lang.Double> oldWeight

Map from clause ID to weight in last iteration. It is used for backtracking.

currentGradient

public java.util.HashMap<java.lang.String,java.lang.Double> currentGradient

Map from clause ID to current gradient value. This will be filled after the invocation of DNLearner#getGradientAndD(MCSAT).

currentD

public java.util.HashMap<java.lang.String,java.lang.Double> currentD

Map from clause ID to current H^(-1)g value. This will be filled after the invocation of DNLearner#getGradientAndD(MCSAT).

oldDHD

public double oldDHD

D'HD value of last iteration. Here D = H^(-1)g, H is Hessian.

oldDG

public double oldDG

D'g value of last iteration. Here D = H^(-1)g.

lambda

public double lambda

Lambda used to control the step size.

alpha

public double alpha

Step size.

nCall

public int nCall

Number of invocations of DNLearner#updateWeight(MCSAT).

Constructor Detail

DNLearner

public DNLearner()

Method Detail

loadingTrainingData

public void loadingTrainingData(MRF _mcsat)

Load training data's truth value into MRF.

Specified by:

loadingTrainingData in class Learner

Parameters:

_mcsat - The MCSAT object to be filled in.

getGradientAndD

public void getGradientAndD(MRF mcsat)

Calculate the gradient and H^(-1)g by filling in currentD and currentGradient.

Parameters:

mcsat - MCSAT instance used to estimate the expectation of violations.

updateWeight

public boolean updateWeight(MRF mcsat)

Update Learner.currentWeight using Diagonal Newton method.

Specified by:

updateWeight in class Learner

Parameters:

mcsat - MCSAT instance after this iteration.

Returns:

true if it wants the learner to terminate iterations. This may happen when converge.