Package libai.nn.unsupervised

Class Competitive

java.lang.Object

libai.nn.NeuralNetwork

libai.nn.unsupervised.UnsupervisedLearning

libai.nn.unsupervised.Competitive

All Implemented Interfaces:

java.io.Serializable

public class Competitive
extends UnsupervisedLearning

Competitive Learning is an unsupervised network where "the winner takes all". A pattern is presented to each neuron, the closest neuron wins the right to be updated. The update makes this neuron fittest for that pattern in the future. Finally, the network learns a set of descriptive patterns or centroids (if is compared with a clustering algorithm). In general, the output for that network will be a binary codified class, when only one bit on (the winner). One of the problems with the competitive learning is the unable to know beforehand, the order of the output. For example, suppose we have 3 possible outputs (100, 010, 001). For the first random initialization, the winner for the first pattern is the neuron 0, so the output is 100. But, if retrain with a new random initialization, the winner for the same pattern will be the neuron 2, so the output is 001. That is not necessary incorrect because the relative order could be maintained between the inputs. Therefore, before calculate any error metric we need to label the patterns with the respective answers, in a similar way as Kohonen does.

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field	Description
protected int	ins
protected int	outs
protected Matrix	W

Fields inherited from class libai.nn.NeuralNetwork

plotter, progress, random

Constructor Summary

Constructors

Constructor	Description
Competitive(int in, int out)	Constructor.
Competitive(int in, int out, java.util.Random rand)	Constructor.

Method Summary

Modifier and Type	Method	Description
double	error(Column[] patterns, Column[] answers, int offset, int length)	Calculate the error using the average distance between the closest neuron.
Column	simulate(Column pattern)	Calculates the output for the pattern.
void	simulate(Column pattern, Column result)	Calculate the output for the pattern and left the result in result.
protected int	simulateNoChange(Matrix pattern)
void	train(Column[] patterns, double alpha, int epochs, int offset, int length)	Train the network using "the winner takes all".

Methods inherited from class libai.nn.NeuralNetwork

error, euclideanDistance2, euclideanDistance2, gaussian, getDefaultRandomGenerator, getPlotter, getProgressBar, initializeProgressBar, open, open, open, save, setPlotter, setProgressBar, train, train

Methods inherited from class java.lang.Object

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

Methods inherited from class libai.nn.unsupervised.UnsupervisedLearning

train, validatePreconditions

Field Detail

W

protected Matrix W

ins

protected int ins

outs

protected int outs

Constructor Detail

Competitive

public Competitive(int in,
                   int out)

Constructor. Creates a network with the specified number of inputs and outputs.

Parameters:

in - Number of inputs.

out - Number of outputs.

Competitive

public Competitive(int in,
                   int out,
                   java.util.Random rand)

Constructor. Creates a network with the specified number of inputs and outputs.

Parameters:

in - Number of inputs.

out - Number of outputs.

rand - Random generator used for creating matrices

Method Detail

train

public void train(Column[] patterns,
                  double alpha,
                  int epochs,
                  int offset,
                  int length)

Train the network using "the winner takes all". For each neuron the Euclidean distance between the pattern and the neuron is calculated. The position with the lowest distance is updated with the rule:
Ww = Ww + alpha.(pattern - Ww)

Specified by:

train in class UnsupervisedLearning

Parameters:

patterns - The patterns to be learned.

alpha - The learning rate.

epochs - The maximum number of iterations

offset - The first pattern position

length - How many patterns will be used.

simulate

public Column simulate(Column pattern)

Description copied from class: NeuralNetwork

Calculates the output for the pattern.

Specified by:

simulate in class NeuralNetwork

Parameters:

pattern - Pattern to use as input.

Returns:

The output for the neural network.

simulate

public void simulate(Column pattern,
                     Column result)

Calculate the output for the pattern and left the result in result. The result will be a row matrix fill with 0 except for the winner position.

Specified by:

simulate in class NeuralNetwork

Parameters:

pattern - Pattern to use as input.

result - The output for the input.

simulateNoChange

protected int simulateNoChange(Matrix pattern)

error

public double error(Column[] patterns,
                    Column[] answers,
                    int offset,
                    int length)

Calculate the error using the average distance between the closest neuron. Less distance means less error and vice versa.

Overrides:

error in class NeuralNetwork

Parameters:

patterns - The array with the patterns to test

answers - The array with the expected answers for the patterns.

offset - The initial position inside the array.

length - How many patterns must be taken from the offset.

Returns:

The average distance between the pattern and the winner for that pattern.