org.encog.neural.networks.training.lma
Class LevenbergMarquardtTraining

java.lang.Object
  org.encog.neural.networks.training.BasicTraining
      org.encog.neural.networks.training.lma.LevenbergMarquardtTraining

All Implemented Interfaces:

Train

public class LevenbergMarquardtTraining
extends BasicTraining

Trains a neural network using a Levenberg Marquardt algorithm (LMA). This training technique is based on the mathematical technique of the same name. http://en.wikipedia.org/wiki/Levenberg%E2%80%93Marquardt_algorithm The LMA training technique has some important limitations that you should be aware of, before using it. Only neural networks that have a single output neuron can be used with this training technique. The entire training set must be loaded into memory. Because of this an Indexable training set must be used. However, despite these limitations, the LMA training technique can be a very effective training method. References: - http://www-alg.ist.hokudai.ac.jp/~jan/alpha.pdf - http://www.inference.phy.cam.ac.uk/mackay/Bayes_FAQ.html

Field Summary

static double	LAMBDA_MAX The max amount for the LAMBDA.
static double	SCALE_LAMBDA The amount to scale the lambda by.

Constructor Summary

LevenbergMarquardtTraining(BasicNetwork network, NeuralDataSet training)
Construct the LMA object.

Method Summary

void	calculateHessian(double[][] jacobian, double[] errors) Calculate the Hessian matrix.
BasicNetwork	getNetwork() Get the current best network from the training.
boolean	isUseBayesianRegularization()
void	iteration() Perform one iteration.
void	setUseBayesianRegularization(boolean useBayesianRegularization) Set if Bayesian regularization should be used.
static double	trace(double[][] m) Return the sum of the diagonal.
double	updateWeights() Update the weights.

Methods inherited from class org.encog.neural.networks.training.BasicTraining

addStrategy, finishTraining, getCloud, getError, getIteration, getStrategies, getTraining, isTrainingDone, iteration, postIteration, preIteration, setCloud, setError, setIteration, setTraining

Methods inherited from class java.lang.Object

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

Field Detail

SCALE_LAMBDA

public static final double SCALE_LAMBDA

The amount to scale the lambda by.

See Also:

Constant Field Values

LAMBDA_MAX

public static final double LAMBDA_MAX

The max amount for the LAMBDA.

See Also:

Constant Field Values

Constructor Detail

LevenbergMarquardtTraining

public LevenbergMarquardtTraining(BasicNetwork network,
                                  NeuralDataSet training)

Construct the LMA object.

Parameters:

network - The network to train. Must have a single output neuron.

training - The training data to use. Must be indexable.

Method Detail

trace

public static double trace(double[][] m)

Return the sum of the diagonal.

Parameters:

m - The matrix to sum.

Returns:

The trace of the matrix.

calculateHessian

public void calculateHessian(double[][] jacobian,
                             double[] errors)

Calculate the Hessian matrix.

Parameters:

jacobian - The Jacobian matrix.

errors - The errors.

getNetwork

public BasicNetwork getNetwork()

Description copied from interface: Train

Get the current best network from the training.

Returns:

The trained network.

isUseBayesianRegularization

public boolean isUseBayesianRegularization()

Returns:

True, if Bayesian regularization is used.

iteration

public void iteration()

Perform one iteration.

setUseBayesianRegularization

public void setUseBayesianRegularization(boolean useBayesianRegularization)

Set if Bayesian regularization should be used.

Parameters:

useBayesianRegularization - True to use Bayesian regularization.

updateWeights

public double updateWeights()

Update the weights.

Returns:

The sum squared of the weights.