PRTools Contents

PRTools User Guide



Back-propagation trained feed-forward neural net classifier


 A Dataset
 UNITS Array indicating number of units in each hidden layer.  Default is a single hidden layer. Its size is the half of the  number of objects in A divided by feature size plus class size  (roughly half of the number of parameters to be optimised) with  a maximum of 100;
 ITER Number of iterations to train (default: inf)
 W_INI Weight initialisation network mapping (default: [], meaning  initialisation by Matlab's neural network toolbox). W_INI can  be the result of a previous training of LMNC, BPXNC or NEURC.
 T Tuning set (default: [], meaning use A)
 FID File descriptor to report progress to (default: 0, no report)

 W Trained feed-forward neural network mapping
 HIST Progress report (see below)


A feed-forward neural network classifier with length(N) hidden layers with  N(I) units in layer I is computed for the dataset A. Training is stopped  after ITER epochs (at least 50) or if the iteration number exceeds twice  that of the best classification result. This is measured by the labeled  tuning set T. If no tuning set is supplied A is used. W_INI is used, if  given, as network initialisation. Use [] if the standard Matlab  initialisation is desired.

An early stopping of the network optimisation is controlled by PRTIME.

The entire training sequence is returned in HIST (number of epochs,  classification error on A, classification error on T, MSE on A, MSE on T).

This routine escapes to KNNC if any class has less than 3 objects.

Uses the Mathwork's Neural Network toolbox.

See also

mappings, datasets, lmnc, neurc, rnnc, rbnc, knnc,

PRTools Contents

PRTools User Guide

This file has been automatically generated. If badly readable, use the help-command in Matlab.