boone.training

Class BackpropTrainer

java.lang.Object

boone.Trainer

boone.training.BackpropTrainer

All Implemented Interfaces:

Storable, java.lang.Cloneable

Direct Known Subclasses:

AdamTrainer, RMSpropTrainer, RpropTrainer

public class BackpropTrainer
extends Trainer

The plain back-propagation trainer.

Version:

$Id$

Author:

August Mayer, Helmut A. Mayer

Field Summary

Fields inherited from class boone.Trainer

currentEpoch, epochs, learnRate, net, shuffle, stepMode, testData, trainingData, trainingSignalGenerator

Constructor Summary

Constructors

Constructor and Description
BackpropTrainer()

Method Summary

Modifier and Type	Method and Description
protected double	calcAdaptation(BrainPart part) Calculates the adaptation value being simply the gradient.
protected void	endPattern() Resets the error signals of all neurons.
protected void	endTrain() Does nothing.
protected void	reset() Resets all neurons and links of the net for a fresh training procedure.
protected void	train(Link link) Calculates the error signal, back-propagates it to the source neuron, and accumulates the gradient information of the link.
protected void	train(java.util.List<java.lang.Double> input, java.util.List<java.lang.Double> target) Performs training of a single pattern with the following steps:
protected void	train(Neuron neuron) Accumulates the gradient of the neuron.

Methods inherited from class boone.Trainer

clone, endBatch, fromXML, getBatchSize, getEpochs, getLearnRate, getNetwork, getRank, getTestData, getTrainingData, getTrainingSignalGenerator, getWinningNeuron, setBatchSize, setEpochs, setLearnRate, setNetwork, setShuffle, setStepMode, setTestData, setTrainingData, setTrainingSignalGenerator, test, test, toString, toXML, train, trainPattern, updateLinks, updateNeurons

Methods inherited from class java.lang.Object

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

Constructor Detail

BackpropTrainer

public BackpropTrainer()

Method Detail

reset

protected void reset()

Resets all neurons and links of the net for a fresh training procedure.

Specified by:

reset in class Trainer

train

protected void train(java.util.List<java.lang.Double> input,
                     java.util.List<java.lang.Double> target)

Performs training of a single pattern with the following steps:

• apply input to net
• calculate the error signal using a TrainingSignalGenerator
• run through neurons according to tick list
• train each neuron and its preceding links
• reset the error signals of all neurons

Specified by:

train in class Trainer

Parameters:

input - an input pattern

target - a target pattern

endPattern

protected void endPattern()

Resets the error signals of all neurons.

endTrain

protected void endTrain()

Does nothing.

Specified by:

endTrain in class Trainer

train

protected void train(Neuron neuron)

Accumulates the gradient of the neuron. The gradient of the neuron bias is calculated by the product of the error signal and the derivative of the activation function (times 1, which is the constant input to the bias link, if you think of the bias as a link with a weight, where the weight is the bias).

Parameters:

neuron - the neuron (bias) to be trained

train

protected void train(Link link)

Calculates the error signal, back-propagates it to the source neuron, and accumulates the gradient information of the link. The link gradient is calculated by the product of the error signal, the derivative of the activation function, and the output from the source (pre-synaptic) neuron. The weighted error signal is back-propagated to the source neuron, where the error signals are accumulated.

Parameters:

link - the link to be trained

calcAdaptation

protected double calcAdaptation(BrainPart part)

Calculates the adaptation value being simply the gradient.

Specified by:

calcAdaptation in class Trainer

Parameters:

part - the brain part to be adapted

Returns:

the adaptation value