Neural nets was: Re[2]: Info on Smalltalk DSLs or Metaprogramming...

Herbert König herbertkoenig at gmx.net
Mon Sep 4 20:44:51 UTC 2006


Hello Rich,


RW> I don't really know anything about FANN. I was planning on just  
RW> implementing my own (they're not that hard). I'm torn because FANN

no they aren't, so I ended up doing my own too.

RW> Portability is probably more important than speed at this point, so I
RW> may just implement my own.

There is a Squeak implementation of a two layered perceptron with
backpropagation learning by Luciano Notarfrancesco, I decided against
it for it was too slow. Much better Smalltalk and OO though.

I needed speed in computing the outputs and FloatArray is fast in #*
and #sum.

RW> The project I'm working on is actually an aLife simulation. The

I'll take a look, when I'm online again.

RW> neural nets would only be a small portion (the brains of the agents).
RW> I like the idea of using them, since the weights can be learned  
RW> genetically (over several generations) as well as modified within a
RW> single generation (using reinforcement learning).

This (reinforcement learning) is said to be slow. What number of
inputs and how many neurons would such a brain have? How many agents?

With 500 epochs of 400 samples training of a single Perceptron of 64
hidden and 16 output neurons took over an hour on a 1.8GHz Pentium M.
It had 140 inputs.

Do you have any pointers on how to use genetic algorithms on neural
nets? More practical, I'm an EE not a CS person :-)



Cheers,

Herbert                            mailto:herbertkoenig at gmx.net









More information about the Squeak-dev mailing list