com.imsl.datamining.neural
Class QuasiNewtonTrainer

java.lang.Object
  com.imsl.datamining.neural.QuasiNewtonTrainer

All Implemented Interfaces:

Trainer, Serializable

public class QuasiNewtonTrainer
extends Object
implements Trainer, Serializable

Trains a network using the quasi-Newton method, MinUnconMultiVar.

The Java Logging API can be used to trace the performance training. The name of this logger is com.imsl.datamining.QuasiNewtonTrainer Accumulated levels of detail correspond to Java's FINE, FINER, and FINEST logging levels with FINE yielding the smallest amount of information and FINEST yielding the most. The levels of output yield the following:

Level	Output
FINE	A message on entering and exiting method train, and any exceptions from and the exit status of MinUnconMultiVar
FINER	All of the messages in FINE, the input settings, and a summary report with the statistics from Network.computeStatistics(), the number of function evaluations and the elapsed time.
FINEST	All of the messages in FINER, and a table of the computed weights and their gradient values.

See Also:

Feed Forward Class Example 1, MinUnconMultiVar, Serialized Form

Nested Class Summary

protected class	QuasiNewtonTrainer.BlockGradObjective
protected class	QuasiNewtonTrainer.BlockObjective
static interface	QuasiNewtonTrainer.Error Error function to be minimized by trainer.
protected class	QuasiNewtonTrainer.GradObjective The Objective class is passed to the optimizer.
protected class	QuasiNewtonTrainer.Objective The Objective class is passed to the optimizer.

Field Summary

static QuasiNewtonTrainer.Error

SUM_OF_SQUARES Compute the sum of squares error.

Constructor Summary

QuasiNewtonTrainer()
Constructs a QuasiNewtonTrainer object.

Method Summary

protected Object	clone() Clones a copy of the trainer.
QuasiNewtonTrainer.Error	getError() Returns the function used to compute the error to be minimized.
double[]	getErrorGradient() Returns the value of the gradient of the error function with respect to the weights.
int	getErrorStatus() Returns the error status from the trainer.
double	getErrorValue() Returns the final value of the error function.
static Formatter	getFormatter() Returns the logging formatter object.
static Logger	getLogger() Returns the Logger object.
int	getTrainingIterations() Returns the number of iterations used during training.
boolean	getUseBackPropagation() Returns the use back propagation setting.
protected void	setEpochNumber(int num) Sets the epoch number for the trainer.
void	setError(QuasiNewtonTrainer.Error error) Sets the function used to compute the network error.
void	setGradientTolerance(double gradientTolerance) Set the gradient tolerance.
void	setMaximumStepsize(double maximumStepsize) Sets the maximum step size.
void	setMaximumTrainingIterations(int maximumTrainingIterations) Sets the maximum number of iterations to use in a training.
protected void	setParallelMode(ArrayList[] allLogRecords) Sets the trainer to be used in multi-threaded EpochTainer.
void	setStepTolerance(double stepTolerance) Sets the scaled step tolerance.
void	setUseBackPropagation(boolean flag) Sets whether or not to use the back propagation algorithm for gradient calculations during network training.
void	train(Network network, double[][] xData, double[][] yData) Trains the neural network using supplied training patterns.

Methods inherited from class java.lang.Object

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

Field Detail

SUM_OF_SQUARES

public static final QuasiNewtonTrainer.Error SUM_OF_SQUARES

Compute the sum of squares error. The sum of squares error term is .

This is the default Error object used by QuasiNewtonTrainer.

Constructor Detail

QuasiNewtonTrainer

public QuasiNewtonTrainer()

Constructs a QuasiNewtonTrainer object.

Method Detail

clone

protected Object clone()

Clones a copy of the trainer.

Overrides:

clone in class Object

getError

public QuasiNewtonTrainer.Error getError()

Returns the function used to compute the error to be minimized.

Returns:

The Error object containing the function to be minimized.

getErrorGradient

public double[] getErrorGradient()

Returns the value of the gradient of the error function with respect to the weights.

Specified by:

getErrorGradient in interface Trainer

Returns:

A double array whose length is equal to the number of network weights, containing the value of the gradient of the error function with respect to the weights. Before training, null is returned.

getErrorStatus

public int getErrorStatus()

Returns the error status from the trainer.

Specified by:

getErrorStatus in interface Trainer

Returns:

An int representing the error status from the trainer. Zero indicates that no errors were encountered during training. Any non-zero value indicates that some error condition arose during training. In many cases the trainer is able to recover from these conditions and produce a well-trained network.

Error Status	Condition
0	No error occurred during training.
1	The last global step failed to locate a lower point than the current error value. The current solution may be an approximate solution and no more accuracy is possible, or the step tolerance may be too large.
2	Relative function convergence; both the actual and predicted relative reductions in the error function are less than or equal to the relative function convergence tolerance.
3	Scaled step tolerance satisfied; the current point may be an approximate local solution, or the algorithm is making very slow progress and is not near a solution, or the step tolerance is too big.
4	MinUnconMultiVar.FalseConvergenceException thrown by optimizer.
5	MinUnconMultiVar.MaxIterationsException thrown by optimizer.
6	MinUnconMultiVar.UnboundedBelowException thrown by optimizer.

See Also:

MinUnconMultiVar.FalseConvergenceException, MinUnconMultiVar.MaxIterationsException, MinUnconMultiVar.UnboundedBelowException

getErrorValue

public double getErrorValue()

Returns the final value of the error function.

Specified by:

getErrorValue in interface Trainer

Returns:

A double representing the final value of the error function from the last training. Before training, NaN is returned.

getFormatter

public static Formatter getFormatter()

Returns the logging formatter object. Logger support requires JDK1.4. Use with earlier versions returns null.

The returned Formatter is used as input to Handler.setFormatter(java.util.logging.Formatter) to format the output log.

Returns:

The Formatter object, if present, or null.

getLogger

public static Logger getLogger()

Returns the Logger object. This is the Logger used to trace this class. It is named com.imsl.datamining.neural.QuasiNewtonTrainer.

Returns:

The Logger object, if present, or null.

getTrainingIterations

public int getTrainingIterations()

Returns the number of iterations used during training.

Returns:

An int representing the number of iterations used during training.

See Also:

MinUnconMultiVar.getIterations()

getUseBackPropagation

public boolean getUseBackPropagation()

Returns the use back propagation setting.

Returns:

a boolean specifying whether or not back propagation is being used for gradient calculations.

setEpochNumber

protected void setEpochNumber(int num)

Sets the epoch number for the trainer.

Parameters:

num - An int array containing the epoch number.

setError

public void setError(QuasiNewtonTrainer.Error error)

Sets the function used to compute the network error.

Parameters:

error - The Error object containing the function to be used to compute the network error. The default is to compute the sum of squares error, SUM_OF_SQUARES.

setGradientTolerance

public void setGradientTolerance(double gradientTolerance)

Set the gradient tolerance.

Parameters:

gradientTolerance - A double specifying the gradient tolerance. Default: cube root of machine precision.

See Also:

MinUnconMultiVar.setGradientTolerance(double)

setMaximumStepsize

public void setMaximumStepsize(double maximumStepsize)

Sets the maximum step size.

Parameters:

maximumStepsize - A nonnegative double value specifying the maximum allowable step size in the optimizer.

See Also:

MinUnconMultiVar.setMaximumStepsize(double)

setMaximumTrainingIterations

public void setMaximumTrainingIterations(int maximumTrainingIterations)

Sets the maximum number of iterations to use in a training.

Parameters:

maximumTrainingIterations - An int representing the maximum number of training iterations. Default: 100.

See Also:

MinUnconMultiVar.setMaxIterations(int)

setParallelMode

protected void setParallelMode(ArrayList[] allLogRecords)

Sets the trainer to be used in multi-threaded EpochTainer.

Parameters:

allLogRecords - An ArrayList array containing the log records.

setStepTolerance

public void setStepTolerance(double stepTolerance)

Sets the scaled step tolerance.

The second stopping criterion for MinUnconMultiVar, the optimizer used by this Trainer, is that the scaled distance between the last two steps be less than the step tolerance.

Parameters:

stepTolerance - A double which is the step tolerance. Default: 3.66685e-11.

See Also:

MinUnconMultiVar.setStepTolerance(double)

setUseBackPropagation

public void setUseBackPropagation(boolean flag)

Sets whether or not to use the back propagation algorithm for gradient calculations during network training.

By default, the quasi-newton algorithm optimizes the network using numerical gradients. This method directs the quasi-newton trainer to use the back propagation algorithm for gradient calculations during network training. Depending upon the data and network architecture, one approach is typically faster than the other, or is less sensitive to finding local network optima.

Parameters:

flag - boolean specifies whether or not to use the back propagation algorithm for gradient calculations. Default value is true.

train

public void train(Network network,
                  double[][] xData,
                  double[][] yData)

Trains the neural network using supplied training patterns.

Each row of xData and yData contains a training pattern. The number of rows in these two arrays must be at least equal to the number of weights in the network.

Specified by:

train in interface Trainer

Parameters:

network - The Network to be trained.

xData - An input double matrix containing training patterns. The number of columns in xData must equal the number of nodes in the input layer.

yData - An output double matrix containing output training patterns. The number of columns in yData must equal the number of perceptrons in the output layer.