Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!cs.utexas.edu!utnut!nott!cunews!freenet.carleton.ca!FreeNet.Carleton.CA!av574
From: av574@FreeNet.Carleton.CA (Tim Sallans)
Subject: Re: Fitness in a competitive population
Message-ID: <D3p6uF.HDy@freenet.carleton.ca>
Sender: av574@freenet3.carleton.ca (Tim Sallans)
Reply-To: av574@FreeNet.Carleton.CA (Tim Sallans)
Organization: The National Capital FreeNet
References: <3hap0l$fl2@news.bu.edu> <3h3nu4$5k1@news.bu.edu> <D3o1n9.734@nbn.com>
Date: Wed, 8 Feb 1995 20:13:26 GMT
Lines: 48


In a previous posting, Karl Meissner (meissner@space.bu.edu) writes:
> Ron Macken (rmacken@calon.com)
>

<assorted conversation removed>
 
> : strategies to different groups.  Each group plays a single elimination
> : tournament.  Each group tournament winner survives until the next 
> : generation.
> 
> : This method of selection is fairly brutal and many good strategies will
> : be eliminated by having the bad luck of facing a tough opponent.  But the
> : best strategies will eventually rise up and it may help your problem of
> : convergence.
> 
> This is a pretty good idea.  It sounds a lot like the "islands" 
> method.  The idea is to divide your population of size N into 
> K subpopulations, K << N.  The strings can only breed and compete with 
> members on their own "island".  Each island will evolve its own
> local population when the algorithm is run for a number of generations.
> At intervals, all the islands "come together" and the entire population
> competes and breeds.  Then they are randomly split into islands again.
>

I originally heard of this method under the name of GAPE -- genetic
algorithms with puncuated equilibria.  If I remember the implementation
correctly (a quick glance through my library couldn't find it), it involved
each sub-population competing and breeding within itself, until there
was a certain level of genetic similarity in each sub-population.  Then
the populations would cross pollinate by exchanging members based on
a selection based on their local fitness.  The important bits (again
if I remeber this correctly) were to choose the correct sub to total
population ratio, and to only evaluate fitness within the sub-populations
(no global criterion) to get good results.  After the sub populations
cross pollinated, all of the members would start with the same fitness
in their respective sub pops and begin anew.
 
Don't know if this helps at all, but your post triggered a memory
that I had to get out of my head... :)

					--Tim


--
Tim Sallans               "Scientists follow their probability instincts
Ottawa, Ontario, Canada    as their hominid forefathers followed theirs.
av574@freenet.carleton.ca  Scientists just know more math." -- Bart Kosko
