Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!udel!gatech!newsfeed.pitt.edu!dsinc!ub!freenet.buffalo.edu!ah818
From: ah818@freenet.buffalo.edu (Phillip M. Fries)
Subject: Re: Right poulation size
Message-ID: <D6vp92.BnB@freenet.buffalo.edu>
Sender: nntp@acsu.buffalo.edu
Nntp-Posting-Host: freenet.buffalo.edu
Reply-To: ah818@freenet.buffalo.edu (Phillip M. Fries)
Organization: State University of New York At Buffalo, NY (USA)
References: <3me2nh$55s@vixen.cso.uiuc.edu> <3k0g0l$scm@rover.ucs.ualberta.ca> <3k2qtf$npj$2@mhadg.production.compuserve.com> <D6v
Date: Tue, 11 Apr 1995 16:17:25 GMT
Lines: 58


In a previous article, ralley@uxa.cso.uiuc.edu (ralley david thomas) says:

>In article <D6vG33.u7@aisb.ed.ac.uk> dave@aisb.ed.ac.uk (Dave Corne) writes:
>>
>> Salvatore R. Mangano <72053.2032@CompuServe.COM> writes:
>>>
>>>In general, you should pick the largest pop size that you can 
>>>afford for the problem at hand. 
>
>>  a very long way down the line!  Generally, we've always found
>>  a cutoff point -- be it 200, 500, or whatever, after which the
>>  benefits of increasing the population size become perishingly 
>>  small
>
>Am I missing the point here? You both seem to be guessing, and you should
>be able to determine the population size needed based on the size of
>the problem and the noise inherent in the evaluation signal?
>
>You might check out Goldberg's "Genetic algorithms, noise and the sizing 
>of populations" available from the IlliGaL web site, 
>http://gal4.ge.uiuc.edu/orderform.html
>
>David
>
>

One can also try an expanding/contracting population size. I use 
tournament selection that always precludes the replacement of the "best 
fitness score" and the insertion of duplicate chromes. This approach 
tends to "fill" the population with chromes of the same fitness value the 
longer a GA works inside a local optimum. Each time tournament selection 
attempts to utilize a "best fitness score" chrome as a chrome to be 
replaced I do a quick frquency analysis on the number of "best score 
chromes" in the population. If I have 50% saturation, I double the size 
of the pool with randomly created new chromes, and I adjust the rate of 
mutation upwards a bit. This helps the GA in breaking out of local 
optimums, if possible.

I generally start with a population size of 50 chromes. The
population size might expand to 250-300 chromes in attempting to break out
of a local optimum, but seldom goes beyond that.  When it does go beyond
that, you know you've probably got a very good local optimum, and it might
be time to quit the search.

Oh, yes, I also sort the pool on fitness score and reduce it back to 50
chromes whenever I break out of a local optimum, which I generally define
as finding a better fitness score than the existing "best fitness score"
upon entering the tournament process. 

There's not as much overhead in this approach as one might think at first
glance, and what overhead there is is made up by the fact that the
population remains small so long as its members remain fertile, so to
speak. 

-- 
Phillip M. Fries
ah818@freenet.buffalo.edu
