/*
* File: OnlineVotedPerceptron.java
* Authors: Justin Basilico
* Company: Sandia National Laboratories
* Project: Cognitive Foundry Learning Core
*
* Copyright October 20, 2010, Sandia Corporation.
* Under the terms of Contract DE-AC04-94AL85000, there is a non-exclusive
* license for use of this work by or on behalf of the U.S. Government. Export
* of this program may require a license from the United States Government.
*
*/
package gov.sandia.cognition.learning.algorithm.perceptron;
import gov.sandia.cognition.annotation.PublicationReference;
import gov.sandia.cognition.annotation.PublicationType;
import gov.sandia.cognition.learning.algorithm.AbstractSupervisedBatchAndIncrementalLearner;
import gov.sandia.cognition.learning.algorithm.ensemble.WeightedBinaryEnsemble;
import gov.sandia.cognition.learning.function.categorization.LinearBinaryCategorizer;
import gov.sandia.cognition.math.matrix.Vector;
import gov.sandia.cognition.math.matrix.VectorFactory;
import gov.sandia.cognition.math.matrix.VectorFactoryContainer;
import gov.sandia.cognition.math.matrix.Vectorizable;
import gov.sandia.cognition.util.DefaultWeightedValue;
/**
* An online version of the Voted-Perceptron algorithm. It is similar to the
* typical Perceptron algorithm except that it creates multiple Perceptrons,
* and combines them together in a weighted vote. Whenever a mistake is made,
* a new Perceptron is created by modifying the previous one and given a weight
* of 1. When it gets an example correct, it simply increments the weight on
* the most recent one.
*
* @author Justin Basilico
* @since 3.1
*/
@PublicationReference(
title="Large Margin Classification Using the Perceptron Algorithm",
author={"Yoav Freund", "Robert E. Schapire" },
year=1999,
type=PublicationType.Journal,
publication="Machine Learning",
pages={277, 296},
url="http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.48.8200")
public class OnlineVotedPerceptron
extends AbstractSupervisedBatchAndIncrementalLearner<Vectorizable, Boolean, WeightedBinaryEnsemble<Vectorizable, LinearBinaryCategorizer>>
implements VectorFactoryContainer
{
/** The factory to create weight vectors. */
protected VectorFactory<?> vectorFactory;
/**
* Creates a new {@code OnlinePerceptron}.
*/
public OnlineVotedPerceptron()
{
this(VectorFactory.getDenseDefault());
}
/**
* Creates a new {@code OnlinePerceptron} with the given vector factory.
*
* @param vectorFactory
* The vector factory to use to create the weight vectors.
*/
public OnlineVotedPerceptron(
final VectorFactory<?> vectorFactory)
{
super();
this.setVectorFactory(vectorFactory);
}
@Override
public WeightedBinaryEnsemble<Vectorizable, LinearBinaryCategorizer> createInitialLearnedObject()
{
return new WeightedBinaryEnsemble<Vectorizable, LinearBinaryCategorizer>();
}
@Override
public void update(
final WeightedBinaryEnsemble<Vectorizable, LinearBinaryCategorizer> target,
final Vectorizable input,
final Boolean output)
{
if (input != null && output != null)
{
this.update(target, input.convertToVector(), (boolean) output);
}
}
/**
* The {@code update} method updates an object of {@code ResultType} using
* the given a new supervised input-output pair, using some form of
* "learning" algorithm.
*
* @param target
* The object to update.
* @param input
* The supervised input vector to learn from.
* @param actual
* The supervised output label to learn from.
*/
public void update(
final WeightedBinaryEnsemble<Vectorizable, LinearBinaryCategorizer> target,
final Vector input,
final boolean actual)
{
// Predict the output as a double (negative values are false, positive
// are true).
final double prediction = target.evaluateAsDouble(input);
// The computation that we do is based on using the last member in
// the ensemble.
final DefaultWeightedValue<LinearBinaryCategorizer> lastMember =
getLastMember(target);
// Make an update if there was an error.
final boolean correct =
(actual && prediction > 0.0)
|| (!actual && prediction < 0.0);
if (correct)
{
// There was no error made, so increase the weight on the latest
// member of the ensemble.
// Note: It should never reach here when lastMember is null because
// then the prediction has to be zero.
lastMember.setWeight(lastMember.getWeight() + 1.0);
}
else
{
final LinearBinaryCategorizer next;
if (lastMember == null)
{
// This is the very first data point we've seen, so we need
// to create an initial categorizer.
next = new LinearBinaryCategorizer(
this.getVectorFactory().createVector(
input.getDimensionality()), 0.0);
}
else
{
// Clone the previous member.
next = lastMember.getValue().clone();
}
if (actual)
{
// An error with the true (positive) category.
next.getWeights().plusEquals(input);
next.setBias(next.getBias() + 1.0);
}
else
{
// An error with the false (negative) category.
next.getWeights().minusEquals(input);
next.setBias(next.getBias() - 1.0);
}
// Add the new member to the ensemble.
target.add(next, 1.0);
}
}
/**
* Gets the last member in the ensemble. This is the one used by the
* algorithm.
*
* @param ensemble
* The ensemble to get the last member from.
* @return
* The last member in the ensemble, or null if it is empty.
*/
public static DefaultWeightedValue<LinearBinaryCategorizer> getLastMember(
final WeightedBinaryEnsemble<Vectorizable, LinearBinaryCategorizer> ensemble)
{
final int ensembleSize = ensemble.getMembers().size();
if (ensembleSize <= 0)
{
return null;
}
else
{
return (DefaultWeightedValue<LinearBinaryCategorizer>)
ensemble.getMembers().get(ensembleSize - 1);}
}
/**
* Gets the VectorFactory used to create the weight vector.
*
* @return The VectorFactory used to create the weight vector.
*/
public VectorFactory<?> getVectorFactory()
{
return this.vectorFactory;
}
/**
* Sets the VectorFactory used to create the weight vector.
*
* @param vectorFactory The VectorFactory used to create the weight vector.
*/
public void setVectorFactory(
final VectorFactory<?> vectorFactory)
{
this.vectorFactory = vectorFactory;
}
}