Xref: utzoo comp.ai:1817 comp.ai.neural-nets:145
Path: utzoo!attcan!uunet!husc6!mit-eddie!mit-amt!bc
From: bc@mit-amt.MEDIA.MIT.EDU (bill coderre)
Newsgroups: comp.ai,comp.ai.neural-nets
Subject: Re: Genetic algorithms
Message-ID: <2520@mit-amt.MEDIA.MIT.EDU>
Date: 25 May 88 06:17:00 GMT
References: <317@mmlai.UUCP>
Reply-To: bc@media-lab.media.mit.edu.MEDIA.MIT.EDU (bill coderre)
Distribution: na
Organization: MIT Media Lab, Cambridge MA
Lines: 60

In article <317@mmlai.UUCP> barash@mmlai.UUCP (Rev. Steven C. Barash) writes:
>A while back someone posted an extended definition of "Genetic algorithms".

>I would also appreciate any pointers to literature in this area.


Well, let's start talking about it right here. Make a change from the
usual rhetoric.

The classic (Holland) Genetic Algorithm stuff involves a pool of rules
which look like ascii strings, the left side of which are
preconditions and the right which are assertions. Attached to each
rule is a probability of firing.

When the clock ticks, all the rules that match their left side are
culled, and one is probabilistically selected to fire.

There is also an "evaluator" that awards "goodness" to rules that are
in the chain of producing a good event. This goodness usually results
in greater probability of firing. (Of course, one could also use
punishment strategies.)

Last, there is a "mutator" that makes new rules out of old. Some
heuristics that are used:

* randomly change a substring (usually one element)

* "breed" two rules together, by taking the first N of one and the
last M-N of another.

The major claim is that this approach avoids straight hill-climbing's
tendency to get stuck on local peaks, by using some "wild" mutations,
like reversing substrings of rules. I'm not gonna guess whether this
claim is true.

I have met Stewart Wilson of the Rowland Institute here in Cambridge,
and he has made simple critters that use the above strategy. They
start out with random rulebases, and over the course of a few million
ticks develop optimal ones.


>>>>>>>>>>
What is particularly of interest to me is genetic-LIKE algorithms that 
use more sophisticated elements than ascii strings and simple numeric
scorings.

My master's research is an attempt to extend Genetic AI in just that
way. I wanna use genetic AI's ideas to cause a Society of Mind to
learn. 

It appears that using Lenat-like ideas is the right way to make the
mutator, but the evaluator seems like a difficult trick. My hunch is
to use knowledge frames ala Winston, but this is looking less likely. 


??????????
So does anybody know about appropriately similar research? 
Anybody got any good ideas?

appreciamucho....................................................bc