Back-propagation trained feed-forward neural net classifier
[W,HIST,UNITS] = BPXNC (A,UNITS,ITER,W_INI,T,FID)
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. Consequently it is not available under Octave.