tuffy.learn
Class Learner

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

Direct Known Subclasses:

DNLearner, NaiveDNLearner, NaiveGDLearner

public abstract class Learner
extends Infer

The abstract class of learning a weight of MLN. A runnable learner should extend this abstract class by specifying some ad-hoc functions. This class extends Infer class, because it uses inference as subroutines. NOTE: this class has static variables. Parallell running of multiple learner instances may cause problems.

Author:

Ce Zhang

Field Summary

java.util.HashMap<java.lang.String,java.lang.Double>	_oldWeight Map from clause name to clause weight learned in last iteration.
int	backtrackCount_ Number of past backtracked steps.
boolean	backtracked Whether current step is a backtracked step.
static java.util.HashMap<java.lang.String,java.lang.Double>	currentWeight Map from clause name to current clause weight.
static java.util.HashMap<java.lang.String,java.lang.Double>	finalWeight Map from clause name to final weight.
static java.util.HashMap<java.lang.String,java.lang.Boolean>	isHardMappings Map from clause name to whether it is assigned to hard weight clause while learning.
(package private) java.util.HashMap<java.lang.String,java.lang.Long>	negativeWeightSatisfication Map from clause name to training data satisfaction of negative weight.
(package private) java.util.HashMap<java.lang.String,java.lang.Long>	negativeWeightViolation Map from clause name to training data violation of negative weight.
double	odds FOR JUNIT TEST ONLY.
static java.util.HashMap<java.lang.String,java.lang.Double>	oriWeight Map from clause name to the clause weight read originally from MLN program.
(package private) java.util.HashMap<java.lang.String,java.lang.Long>	positiveWeightSatisfication Map from clause name to training data satisfaction of positive weight.
(package private) java.util.HashMap<java.lang.String,java.lang.Long>	positiveWeightViolation Map from clause name to training data violation of positive weight.
java.util.HashMap<java.lang.String,java.lang.Long>	trainingSatisification Map from clause name to current training data satisfaction.
java.util.HashMap<java.lang.String,java.lang.Long>	trainingViolation Map from clause name to current training data violation.

Fields inherited from class tuffy.main.Infer

db, dmover, grounding, mln, options

Constructor Summary

Learner()

Method Summary

void	calcCurrentTrainingViolation() Calculate current training violations according to current sign of weight.
void	dumpAnswers(java.lang.String fout) Dump the learning result to file CommandOptions.fout.
void	fillInCurrentWeight(MRF _mcsat) Initialize weight according to the log odd of training data.
abstract void	loadingTrainingData(MRF _mcsat) Reading from training data and fill it into MRF.atoms.
void	run(CommandOptions opt) run the learner
abstract boolean	updateWeight(MRF mcsat) Update currentWeight to new weights according to information provided by mcsat instance, e.g., MCSAT#expectationOfViolation; This is a virtual function, each instance of Learner should materialize an adhoc version of this function.

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

currentWeight

public static java.util.HashMap<java.lang.String,java.lang.Double> currentWeight

Map from clause name to current clause weight.

oriWeight

public static java.util.HashMap<java.lang.String,java.lang.Double> oriWeight

Map from clause name to the clause weight read originally from MLN program.

isHardMappings

public static java.util.HashMap<java.lang.String,java.lang.Boolean> isHardMappings

Map from clause name to whether it is assigned to hard weight clause while learning.

_oldWeight

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

Map from clause name to clause weight learned in last iteration.

finalWeight

public static java.util.HashMap<java.lang.String,java.lang.Double> finalWeight

Map from clause name to final weight. Here by final weight, it means the the average of the weights in each iteration.

odds

public double odds

FOR JUNIT TEST ONLY.

backtracked

public boolean backtracked

Whether current step is a backtracked step.

backtrackCount_

public int backtrackCount_

Number of past backtracked steps.

trainingViolation

public java.util.HashMap<java.lang.String,java.lang.Long> trainingViolation

Map from clause name to current training data violation.

trainingSatisification

public java.util.HashMap<java.lang.String,java.lang.Long> trainingSatisification

Map from clause name to current training data satisfaction.

positiveWeightViolation

java.util.HashMap<java.lang.String,java.lang.Long> positiveWeightViolation

Map from clause name to training data violation of positive weight.

positiveWeightSatisfication

java.util.HashMap<java.lang.String,java.lang.Long> positiveWeightSatisfication

Map from clause name to training data satisfaction of positive weight.

negativeWeightViolation

java.util.HashMap<java.lang.String,java.lang.Long> negativeWeightViolation

Map from clause name to training data violation of negative weight.

negativeWeightSatisfication

java.util.HashMap<java.lang.String,java.lang.Long> negativeWeightSatisfication

Map from clause name to training data satisfaction of negative weight.

Constructor Detail

Learner

public Learner()

Method Detail

run

public void run(CommandOptions opt)
         throws java.sql.SQLException

run the learner

Parameters:

opt - Command line Options

Throws:

java.sql.SQLException

dumpAnswers

public void dumpAnswers(java.lang.String fout)

Dump the learning result to file CommandOptions.fout. The format of this file is consistent with inference part.

updateWeight

public abstract boolean updateWeight(MRF mcsat)

Update currentWeight to new weights according to information provided by mcsat instance, e.g., MCSAT#expectationOfViolation; This is a virtual function, each instance of Learner should materialize an adhoc version of this function. This function should return whether the learner thinks this iteration should terminate.

Parameters:

mcsat - MCSAT instance after this iteration.

Returns:

Whether terminate.

loadingTrainingData

public abstract void loadingTrainingData(MRF _mcsat)

Reading from training data and fill it into MRF.atoms. This function should be materialized by an instance of abstract class Learner.

Parameters:

_mcsat - The MCSAT object to be filled in.

calcCurrentTrainingViolation

public void calcCurrentTrainingViolation()

Calculate current training violations according to current sign of weight. This function will fill trainingViolation and trainingSatisification.

fillInCurrentWeight

public void fillInCurrentWeight(MRF _mcsat)

Initialize weight according to the log odd of training data.

Parameters:

_mcsat - MCSAT instance containing the vio/sat informaiton.