Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!delmarva.com!newsfeed.internetmci.com!howland.reston.ans.net!EU.net!uknet!newsfeed.ed.ac.uk!ainews!andrewt
From: andrewt@aisb.ed.ac.uk (Andrew Tuson)
Subject: Re: Urgent:Advice Needed on GA application
Message-ID: <DE67K9.71H@aisb.ed.ac.uk>
Sender: news@aisb.ed.ac.uk (Network News Administrator)
Organization: Dept AI, Edinburgh University, Scotland
References: <41qlos$m49@infoserv.rug.ac.be> <stevem-3008952200400001@newlight.iea.com>
Date: Thu, 31 Aug 1995 10:24:09 GMT
Lines: 38

In article <stevem-3008952200400001@newlight.iea.com> stevem@comtch.iea.com (Steve McGrew) writes:

[Useful advice about GAs deleted!]

>        4. I think self-adaptive GA's are an interesting but impractical
>           idea at this stage.  The only reason to have a self-adaptive GA for
>           a particular problem is if you don't really understand the problem
>           and you want the GA to try to understand it for you.
>           Unfortunately, the self- adaptive GA's that have been described in
>           the literature simply aren't smart enough.

It`s interesting that you brought this up - I have studied adaptation of
operator probabilities for the last 5 months as my MSc project. I found that
simply encoding the operator probabilities onto each string made performance
(measured in terms of both solution quality and speed to solution) worse!

This appears to be for two reasons: first, adaptation is unrealiable - the
information about which operators to use is indirect, therefore the selection
pressure upon the encoded probabilities is low and noisy; second, the roles
that crossover and mutation play do, in some respects, depend upon the problem.
By this I mean that sometimes the limiting factor is the speed which improved
solutions can be found - in which case the most productive (by this I mean the
ability to produce improved children) operator should be used the most. However
there are situations where although say crossover is producing all the
improvements, it relies upon diversity in the population to do so - so mutation
shold be used widely. Adaptation by co-evolution allows the user no control
over this!

I think that you have hit the nail on the head with the statement above.
Different problems pose different difficulties for a GA - the key to good GA
design is to find out what these problems are and inform the GA (by changing
the representation, operators, etc.) so it can avoid them.
-- 
Andrew Tuson (andrewt@aisb.ed.ac.uk)

Department of Artificial Intelligence, University of Edinburgh, Scotland, U.K.
An expert is a person who avoids the small errors while sweeping on to the
grand fallacy..........:-)
