de.htwdd.rosenkoenig.ga
Class NetworkFitness

java.lang.Object
  de.htwdd.rosenkoenig.ga.NetworkFitness

All Implemented Interfaces:

de.htwdd.ga.FitnessFunction

public class NetworkFitness
extends java.lang.Object
implements de.htwdd.ga.FitnessFunction

A FitnessFunction to calculate the fitness of a neural net which could be applied to an artificial player.
To calculate the fitness, the neural network has to be trained, which is a very time consuming process. NetworkFitness does not use realistic numbers of training cycles (e.g. 1000 or more) but trains the network for about 25 cycles. So this fitness function can only approximate if a chromosome could be appropriate for artificial players. NetworkFitness does not calculate a mathematically correct fitness!

See the Rosenkoenig online documentation for more detailed information.

Field Summary

private Backpropagation	backprop
private java.util.ArrayList<Pattern>	ffPatterns Training patterns to train a Feed Forward Neural Network.
private int	index Index of the FitnessFunction.
private int[]	layerSizes
private org.apache.log4j.Logger	log
private static int	numFunctions Static member that holds the number of instances of NetworkFitness present in the current process.
private java.util.ArrayList<double[]>	somPatterns Training patterns for SOM training.
private boolean	useSom

Constructor Summary

NetworkFitness()
Creates a new fitness function and initializes the training data.

Method Summary

double

computeFitness(java.util.BitSet chromosome) Calculates the fitness for a given chromosome.

Methods inherited from class java.lang.Object

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

Field Detail

numFunctions

private static int numFunctions

Static member that holds the number of instances of NetworkFitness present in the current process.

index

private int index

Index of the FitnessFunction.

log

private org.apache.log4j.Logger log

somPatterns

private java.util.ArrayList<double[]> somPatterns

Training patterns for SOM training.

ffPatterns

private java.util.ArrayList<Pattern> ffPatterns

Training patterns to train a Feed Forward Neural Network.

layerSizes

private int[] layerSizes

useSom

private boolean useSom

backprop

private Backpropagation backprop

Constructor Detail

NetworkFitness

public NetworkFitness()

Creates a new fitness function and initializes the training data. The files used for training are specified by GeneticRosenkoenig.FILENAMES.

Method Detail

computeFitness

public double computeFitness(java.util.BitSet chromosome)

Calculates the fitness for a given chromosome. Extracts the chromosome and determines the number of layers, the number of neurons per layer. The first bit of the chromosome determines if a SOM should be used.
After that a neural network is created and trained for a small number of cycles. The lower the error on the training set, the higher is the chromosome's fitness.

Specified by:

computeFitness in interface de.htwdd.ga.FitnessFunction