Re: Announcing Breeder 1.0

[James Joseph Clark]

A couple of years back I worked on using a GA to learn FSMs. I was
using these nice reprogrammable PLA chips by Lattice Semi, which take
in their programming data as a serial stream, perfect for crossover
and mutation at the bit level. The upshot was that I found that
evolving FSMs to do a particular function was quite difficult using
GAs, mainly due to the "spikiness" of the objective function. For
example, I tried evolving a FSM that would repetitively count to 100
from zero by one. As I recall, after hundreds of thousands of generations,
the best I cam up with was a FSM that had a string of 72 consecutive
increments by one before jumping by more than one. I found that real
careful attention needs to be paid to the design of the objective
function to get good results.

Well, good luck. Let me know if you need any more of what little
insight I have on this matter.

Jim Clark
address@hidden

> Hello,
> 
> I'm going to be working on a varient of this. I am going to be using a GA 
> to evolve pseudo finite state machines. I haven't gotten to this yet, I 
> first need to finish my environment (I only recently shifted from a class
> 3 lurker to a class 4...). 
> 
> I am also currently working on a web page for my project which 
> (hopefully) will explain how I'm thinking about going about this.
> I may be using the Breeder lib (chromosome part) for this. I haven't
> dug into the GA code it provides yet, but it sounds like it will
> at least be a good starting point (thanks JJ!!).
> 
> John Eikenberry
> address@hidden