GAlib and neural networks

[Michael Z Katz-Hyman]

I just recently integrated GAs into NNs for use with polymer physics.
Right now the GA is used to itialize the weights using the M.S.E. as the
objectuve function.  It works quite well, many times the backpropagation
algo. doesn't even have to be run, I can under 1% error on some things
like XOR by just running 50 populations over 500 generations (or something
similar).   

One problem I have with Galib is that the matrix implementations are
static, meaning that if you do a 3D matrix let's say to represent the
weight matrices, and if your first weight matrix and second weight matrix
are different sizes then your have some "extra" space in the
representation.  Right now we are running a two (weight) layer matrix with
it being represented in the GA by one matrix whose number of columns equal
the number of hidden nodes and the first n rows belong to the first weight
matrix and the next m rows belong to the second matrix, this has worked
well so far.

As for geometry, we haven't tried that yet but it  is planned.

As for software, we just wrote our own basing it off a matrix class
(CAMClass)....

Hope this helps.

Michael

KATZ-HYMAN, MICHAEL Z.                         m.z.katz-hyman at LaRC.NASA.GOV

    Mail Stop 226
    8 Taylor Street      	               Advanced Materials 
    NASA Langley Research Center               Building 1293A, Room 122
    Hampton, VA 23681-0001                     Phone +1 757 864-4235


On Wed, 9 Aug 2000, Bram Stolk wrote:

> Hello,
> 
> 
> Has someone on this list tried using GAlib on neural networks?
> 
> What I had in mind was, using the NN weights as the genome.
> And to determine the fitness of a NN, simply have the NN do
> some conventional NN-learning, and then testing the NN's fault-rate.
> 
> A more advanced aproach would also encode the geometry of the
> NN (nr of neurons and the connections) in the genome.
> 
> I am looking for NN software that has a similar ease of use
> as GAlib, comes as C++ source, and interfaces nicely with GA.
> Any suggestions?
> 
> Many thanks,
> 
>    Bram Stolk
> 
> -- 
> ------------------------------------------------------------------------------
>  Bram Stolk, VR Specialist.
>  SARA Academic Computing Services Amsterdam, PO Box 94613, 1090 GP  AMSTERDAM
>  email: bram at sara.nl   Phone +31-20-5923059  Fax +31-20-6683167
> 
>  "I heard if you play the NT-4.0-CD backwards, you get a satanic message."
>  "Thats nothing, if you play it forward, it installs NT-4.0"
> ------------------------------------------------------------------------------
> 
>