Package com.tibco.patterns.learn.training

Class COScoreChange

java.lang.Object

com.tibco.patterns.learn.training.COIterationCount

com.tibco.patterns.learn.training.COScoreChange

All Implemented Interfaces:

ConvergenceObserver, TrainingObserver

public class COScoreChange
extends COIterationCount

Stops the training when the maximum change in prediction scores compared with the previous iteration is below the given threshold. Always does at least minIterations.

Constructor Summary

Constructors

Constructor	Description
COScoreChange()	Creates object with default parameters.
COScoreChange(double targetScoreChange, int minIterations, int maxIterations)	Creates the object with given parameters.

Method Summary

Modifier and Type	Method	Description
void	beginIteration()	Prepares to accumulate statistics for a new training iteration.
void	endIteration()	If printing is enabled, outputs the maximum score change of the iteration that has just ended.
void	evaluatePrediction(VectorExample e)	Updates the current maximum score change based on the prediction score for the given example.
static int	getDftMinIterations()
static double	getDftTargetScoreChange()
double	getProgressEstimate()	Returns an estimate of training progress, which can be used to indicate how close the desired convergence is.
boolean	isConverged()	Returns whether the trained model has converged.
boolean	isPerfectResult()
boolean	needStopTraining()	Returns whether the training has to be stopped after current iteration.
COScoreChange	setPrint(boolean value)	Enables printing of iteration results to console.
java.lang.String	toString()

Methods inherited from class com.tibco.patterns.learn.training.COIterationCount

getDftMaxIterations, getMaxIterations, getNIterations, hasIterations, setMaxIterations, verifyIterationBegun, verifyIterationEnded

Methods inherited from class java.lang.Object

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

Constructor Detail

COScoreChange

public COScoreChange(double targetScoreChange,
                     int minIterations,
                     int maxIterations)

Creates the object with given parameters.

Parameters:

targetScoreChange - - target change of prediction scores that is used as a threshold to stop the training.

minIterations - - minimum number of iterations that is always performed.

maxIterations - - maximum number of model training iterations to perform.

Throws:

java.lang.IllegalArgumentException - if max iterations is not positive, or target score change is not positive, or min iterations is not less than max iterations, or min iterations is negative.

COScoreChange

public COScoreChange()

Creates object with default parameters.

See Also:

getDftTargetScoreChange(), getDftMinIterations()

Method Detail

getDftTargetScoreChange

public static double getDftTargetScoreChange()

Returns:

default target score change (0.01).

getDftMinIterations

public static int getDftMinIterations()

Returns:

default minimum number of iterations to perform (0).

toString

public java.lang.String toString()

Overrides:

toString in class COIterationCount

setPrint

public COScoreChange setPrint(boolean value)

Enables printing of iteration results to console.

beginIteration

public void beginIteration()

Prepares to accumulate statistics for a new training iteration.

Specified by:

beginIteration in interface TrainingObserver

Overrides:

beginIteration in class COIterationCount

Throws:

java.lang.IllegalStateException - if an iteration has already begun.

endIteration

public void endIteration()

If printing is enabled, outputs the maximum score change of the iteration that has just ended.

Specified by:

endIteration in interface TrainingObserver

Overrides:

endIteration in class COIterationCount

Throws:

java.lang.IllegalStateException - if iteration has not begun.

evaluatePrediction

public void evaluatePrediction(VectorExample e)

Updates the current maximum score change based on the prediction score for the given example. The examples in the dataset can be trained/evaluated in any order.

Specified by:

evaluatePrediction in interface TrainingObserver

Overrides:

evaluatePrediction in class COIterationCount

Parameters:

e - - must contain RLink prediction.

Throws:

java.lang.IllegalStateException - if iteration has not begun.

getProgressEstimate

public double getProgressEstimate()

Description copied from interface: ConvergenceObserver

Returns an estimate of training progress, which can be used to indicate how close the desired convergence is. The interpretation and range of the returned value can vary between subclasses. Typically, the return value approaches 0 as the model is converging. Should be called after an iteration, not in the middle of iteration.

Specified by:

getProgressEstimate in interface ConvergenceObserver

Overrides:

getProgressEstimate in class COIterationCount

Returns:

maximum change in prediction scores compared with previous iteration.

Throws:

java.lang.IllegalStateException - if iteration has not ended or no iterations were performed.

isConverged

public boolean isConverged()

Description copied from interface: ConvergenceObserver

Returns whether the trained model has converged. Depending on implementation, this can be based on one or several iterations of evaluating the verification set.

Specified by:

isConverged in interface ConvergenceObserver

Overrides:

isConverged in class COIterationCount

Returns:

true if minimum number of iterations were done and the maximum change in prediction scores compared with previous iteration is below the target score change.

Throws:

java.lang.IllegalStateException - if iteration has not ended or no iterations were performed.

needStopTraining

public boolean needStopTraining()

Description copied from interface: ConvergenceObserver

Returns whether the training has to be stopped after current iteration. This may mean that the model has converged, based on all score evaluations that were performed. Or the model may have failed to converge and another termination condition was reached.

Specified by:

needStopTraining in interface ConvergenceObserver

Overrides:

needStopTraining in class COIterationCount

Returns:

true if maximum number of iterations was performed or convergence criteria are satisfied.

Throws:

java.lang.IllegalStateException - if iteration has not ended or no iterations were performed.

See Also:

isConverged()

isPerfectResult

public boolean isPerfectResult()

Specified by:

isPerfectResult in interface TrainingObserver

Overrides:

isPerfectResult in class COIterationCount

Returns:

true if there was no change in any prediction score compared with the previous iteration

Throws:

java.lang.IllegalStateException - if iteration has not ended or no iterations were performed.