org.apache.commons.math3.ml.neuralnet.sofm

Class KohonenUpdateAction

java.lang.Object

org.apache.commons.math3.ml.neuralnet.sofm.KohonenUpdateAction

All Implemented Interfaces:

UpdateAction

public class KohonenUpdateAction
extends java.lang.Object
implements UpdateAction

Update formula for Kohonen's Self-Organizing Map.
The update method modifies the features w of the "winning" neuron and its neighbours according to the following rule: wnew = wold + α e(-d / σ) * (sample - wold) where

• α is the current learning rate,
• σ is the current neighbourhood size, and
• d is the number of links to traverse in order to reach the neuron from the winning neuron.

This class is thread-safe as long as the arguments passed to the constructor are instances of thread-safe classes.
Each call to the update method will increment the internal counter used to compute the current values for

• the learning rate, and
• the neighbourhood size.

Consequently, the function instances that compute those values (passed to the constructor of this class) must take into account whether this class's instance will be shared by multiple threads, as this will impact the training process.

Since:

3.3

Field Summary

Fields

Modifier and Type	Field and Description
private DistanceMeasure	distance Distance function.
private LearningFactorFunction	learningFactor Learning factor update function.
private NeighbourhoodSizeFunction	neighbourhoodSize Neighbourhood size update function.
private java.util.concurrent.atomic.AtomicLong	numberOfCalls Number of calls to update(Network,double[]).

Constructor Summary

Constructors

Constructor and Description
KohonenUpdateAction(DistanceMeasure distance, LearningFactorFunction learningFactor, NeighbourhoodSizeFunction neighbourhoodSize)

Method Summary

Methods

Modifier and Type	Method and Description
private double[]	computeFeatures(double[] current, double[] sample, double learningRate) Computes the new value of the features set.
private Neuron	findAndUpdateBestNeuron(Network net, double[] features, double learningRate) Searches for the neuron whose features are closest to the given sample, and atomically updates its features.
long	getNumberOfCalls() Retrieves the number of calls to the update method.
void	update(Network net, double[] features) Updates the network in response to the sample features.
private void	updateNeighbouringNeuron(Neuron n, double[] features, double learningRate) Atomically updates the given neuron.

Methods inherited from class java.lang.Object

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

Field Detail

distance

private final DistanceMeasure distance

Distance function.

learningFactor

private final LearningFactorFunction learningFactor

Learning factor update function.

neighbourhoodSize

private final NeighbourhoodSizeFunction neighbourhoodSize

Neighbourhood size update function.

numberOfCalls

private final java.util.concurrent.atomic.AtomicLong numberOfCalls

Number of calls to update(Network,double[]).

Constructor Detail

KohonenUpdateAction

public KohonenUpdateAction(DistanceMeasure distance,
                   LearningFactorFunction learningFactor,
                   NeighbourhoodSizeFunction neighbourhoodSize)

Parameters:

distance - Distance function.

learningFactor - Learning factor update function.

neighbourhoodSize - Neighbourhood size update function.

Method Detail

update

public void update(Network net,
          double[] features)

Updates the network in response to the sample features.

Specified by:

update in interface UpdateAction

Parameters:

net - Network.

features - Training data.

getNumberOfCalls

public long getNumberOfCalls()

Retrieves the number of calls to the update method.

Returns:

the current number of calls.

updateNeighbouringNeuron

private void updateNeighbouringNeuron(Neuron n,
                            double[] features,
                            double learningRate)

Atomically updates the given neuron.

Parameters:

n - Neuron to be updated.

features - Training data.

learningRate - Learning factor.

findAndUpdateBestNeuron

private Neuron findAndUpdateBestNeuron(Network net,
                             double[] features,
                             double learningRate)

Searches for the neuron whose features are closest to the given sample, and atomically updates its features.

Parameters:

net - Network.

features - Sample data.

learningRate - Current learning factor.

Returns:

the winning neuron.

computeFeatures

private double[] computeFeatures(double[] current,
                       double[] sample,
                       double learningRate)

Computes the new value of the features set.

Parameters:

current - Current values of the features.

sample - Training data.

learningRate - Learning factor.

Returns:

the new values for the features.