boone.training

Class NetError

java.lang.Object

boone.training.NetError

public class NetError
extends java.lang.Object

Collection of Neural network error measurements.

The difference to the TrainingSignalGenerator is that these methods calculate the errors for the whole network, not for each neuron.

Since:

2005-04-15

Version:

$Id$

Author:

August Mayer

Constructor Summary

Constructors

Constructor and Description
NetError()

Method Summary

Modifier and Type	Method and Description
static double[]	netErrors(NeuralNet net, PatternSet patterns, double[][][] statValues, double min_out, double max_out) Calculate SSE, MSE, NRMSE, SQEP, and classification winner error, and return them in an array of five doubles.
static double	sumSquareError(NeuralNet net, PatternSet patterns) Calculate the Square Sum Error on the network.

Methods inherited from class java.lang.Object

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

Constructor Detail

NetError

public NetError()

Method Detail

sumSquareError

public static double sumSquareError(NeuralNet net,
                                    PatternSet patterns)

Calculate the Square Sum Error on the network.

Parameters:

net - NeuralNet to calc the error for.

patterns - Patterns to calc the error with.

Returns:

Square Sum Error, that is, the sum of the squared error values.

netErrors

public static double[] netErrors(NeuralNet net,
                                 PatternSet patterns,
                                 double[][][] statValues,
                                 double min_out,
                                 double max_out)

Calculate SSE, MSE, NRMSE, SQEP, and classification winner error, and return them in an array of five doubles. This method calculates all error values in one loop, and is thus more efficient.

This method can also return a stats array with per-pattern square error sums, and neuron output and error values for every neuron and pattern.

Parameters:

net - NeuralNet to calc the errors for.

patterns - Patterns to calc the errors with.

statValues - If non-null and length >= 3, more detailed statistics are returned:

• statValues[0][0][pattern#] -- total squared error sum per pattern.
• statValues[1][pattern#][neuron#] -- neuron outputs for every pattern.
• statValues[2][pattern#][neuron#] -- neuron errors (non-squared) for every pattern.
• statValues[3][pattern#][neuron#] -- neuron target values for every pattern.

min_out - Minimum expected output (for squared error percentage); nothing bad happens if the real output happens to be larger.

max_out - Maximum expected output (for squared error percentage); nothing bad happens if the real output happens to be larger.

Returns:

double[5] with the SSE, MSE, NRMSE, SQEP, and winner classification error.