edu.udo.cs.yale.operator.learner.kernel.evosvm
Class EvoOptimization

java.lang.Object
  edu.udo.cs.yale.tools.math.optimization.ec.es.ESOptimization
      edu.udo.cs.yale.operator.learner.kernel.evosvm.EvoOptimization

All Implemented Interfaces:

Optimization

public class EvoOptimization
extends ESOptimization

Evolutionary Strategy approach for SVM optimization. Currently only classification problems are supported.

Version:

$Id: EvoOptimization.java,v 1.28 2006/09/30 00:05:30 ingomierswa Exp $

Author:

Ingo Mierswa

Field Summary

private double	c The parameter C indicating the weight of errors for regression.
private double	epsilon This parameter indicates the weight of errors for regression.
private ExampleSet	exampleSet The training example set.
private static double	IS_ZERO Number smaller than this number are regarded as zero.
private Kernel	kernel The used kernel function.
private OptimizationFunction	optimizationFunction This function is to maximize.
private double[]	ys The label values.

Fields inherited from class edu.udo.cs.yale.tools.math.optimization.ec.es.ESOptimization

BOLTZMANN_SELECTION, CUT_SELECTION, GAUSSIAN_MUTATION, INIT_TYPE_MAX, INIT_TYPE_MIN, INIT_TYPE_RANDOM, MUTATION_TYPES, NO_MUTATION, NON_DOMINATED_SORTING_SELECTION, POPULATION_INIT_TYPES, RANK_SELECTION, ROULETTE_WHEEL, SELECTION_TYPES, SPARSITY_MUTATION, STOCHASTIC_UNIVERSAL, SWITCHING_MUTATION, TOURNAMENT_SELECTION, UNIFORM_SELECTION, VALUE_TYPE_DOUBLE, VALUE_TYPE_INT

Constructor Summary

EvoOptimization(ExampleSet exampleSet, Kernel kernel, double c, int initType, int maxIterations, int generationsWithoutImprovement, int popSize, int selectionType, double tournamentFraction, boolean keepBest, int mutationType, double crossoverProb, boolean showPlot, RandomGenerator random)
Creates a new evolutionary SVM optimization.

Method Summary

private static double	determineC(double _c, Kernel kernel, ExampleSet exampleSet)
PerformanceVector	evaluateIndividual(Individual individual) Subclasses must implement this method to calculate the fitness of the given individual.
private EvoSVMModel	getModel(double[] alphas) Returns a model containing all support vectors, i.e. the examples with non-zero alphas.
PerformanceVector	getOptimizationPerformance() Delivers the fitness of the best individual as performance vector.
EvoSVMModel	train() Trains the SVM.

Methods inherited from class edu.udo.cs.yale.tools.math.optimization.ec.es.ESOptimization

getBestFitnessEver, getBestFitnessInGeneration, getBestPerformanceEver, getBestValuesEver, getGeneration, getMax, getMin, getValueType, nextIteration, optimize, setMax, setMin, setValueType

Methods inherited from class java.lang.Object

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

Field Detail

IS_ZERO

private static final double IS_ZERO

Number smaller than this number are regarded as zero.

See Also:

Constant Field Values

exampleSet

private ExampleSet exampleSet

The training example set.

kernel

private Kernel kernel

The used kernel function.

epsilon

private double epsilon

This parameter indicates the weight of errors for regression.

c

private double c

The parameter C indicating the weight of errors for regression.

ys

private double[] ys

The label values.

optimizationFunction

private OptimizationFunction optimizationFunction

This function is to maximize.

Constructor Detail

EvoOptimization

public EvoOptimization(ExampleSet exampleSet,
                       Kernel kernel,
                       double c,
                       int initType,
                       int maxIterations,
                       int generationsWithoutImprovement,
                       int popSize,
                       int selectionType,
                       double tournamentFraction,
                       boolean keepBest,
                       int mutationType,
                       double crossoverProb,
                       boolean showPlot,
                       RandomGenerator random)

Creates a new evolutionary SVM optimization.

Method Detail

determineC

private static final double determineC(double _c,
                                       Kernel kernel,
                                       ExampleSet exampleSet)

evaluateIndividual

public PerformanceVector evaluateIndividual(Individual individual)

Description copied from class: ESOptimization

Subclasses must implement this method to calculate the fitness of the given individual. Please note that null might be returned for non-valid individuals. The fitness will be maximized.

Specified by:

evaluateIndividual in class ESOptimization

train

public EvoSVMModel train()
                  throws OperatorException

Trains the SVM. In this case an evolutionary strategy approach is applied to determine the best alpha values.

Throws:

OperatorException

getOptimizationPerformance

public PerformanceVector getOptimizationPerformance()

Delivers the fitness of the best individual as performance vector.

getModel

private EvoSVMModel getModel(double[] alphas)

Returns a model containing all support vectors, i.e. the examples with non-zero alphas.