Learning Regression Function / Neural Networks

[Alistair MacLeod]

Hi,

I have a (multi)function of three variables and two outputs that I
need to be learnt from examples. The inputs are are slightly
correlated with each other. The outputs will also be. The function is
not at all smooth - there are many irregular peaks. Training data will
be generated as the system is running so it should be possible to
update the function as new data becomes available. It must be fast in
use and preferably fast to train incrementally so that we can use all
training data as it is produced.

I suspect a simple backpropagating multilayer perceptron should do ok
for this but I have not really used these in anger before.

My questions for the perlmongers:

1. Is a perceptron the best approach? How else could I do this?

2. Are there any good modules to implement this? It seems everything
on CPAN is either just a toy for the author to learn about NNs or is a
very early version and has not been maintained for three years.

Thanks,
Alistair