Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!oitnews.harvard.edu!purdue!lerc.nasa.gov!magnus.acs.ohio-state.edu!math.ohio-state.edu!cs.utexas.edu!news.sprintlink.net!simtel!harbinger.cc.monash.edu.au!yarrina.connect.com.au!labtam!labtam!chris
From: chris@labtam.labtam.oz.au (Chris Taylor)
Subject: Re: Bloating the population size
Message-ID: <chris.803613436@labtam>
Organization: Labtam Australia Pty. Ltd., Melbourne, Australia
References: <ksteinhoff-0906951209300001@james12.hac.com> <3rheaf$af4@hawk.ee.port.ac.uk> <3rqoe7$j10@pith.uoregon.edu>
Date: Tue, 20 Jun 1995 01:57:16 GMT
Lines: 58

ben@chinook.uoregon.edu (Ben Marcotte) writes:

>Simon Thompson (sgt@sis.port.ac.uk) wrote:
>: In article <ksteinhoff-0906951209300001@james12.hac.com>, ksteinhoff@msmail3.hac.com (Karl Steinhoff) says:

>: I think that the only consensus on population size is that in general, the
>: bigger the better. But the only reason why bigger is better is that if you
>: generate the population randomly, or at least with a search space spanning
>: heuristic, then you will introduce more variety, ie. more genes, into the
>: genepool. 

>: Basically I think that GA's can be improved in two ways with regard to 
>: populations.

>: 1) Increase the size for the whole shibang.

>	Not neccesarily.  And if my previous discussion wasn't confusing, 
>this one will be!

>	The quick way to put it is that two individuals with "good" 
>fitnesses have a harder time "finding" each other for mating in a larger
>population.


One obvious way to address that is to make selection favour pairing
according to fitness. (i.e the 'good-looking' parents tend to seek each
other out for mating - sounds familiar)


My small experience with GA's is that population size is not
as major a factor as you might think. 
Naively you expect that more individuals means more parallelism
and therefore convergence should be correspondingly faster.
But more individuals also means more redundancy and less chance
of good individuals meeting for mating - and more computation.

Population size seems to be have most effect at start-up time,
where a greater variety means better chance of randomly choosing 
good individuals and a wider range of genes to assess.

After the first few generations only the best of these should tend to
survive and so a degree of redundancy starts to creep in.
Probably one should start off with a big initial population and then
reduce this after a few generations.
Variety should be maintained in the smaller population by mutation so one
shouldn't need to maintain a big population for this purpose.

My understanding is that an advantage of a GA with population size N
over just performing something like Simulated Annealing N times
(computationally similar) is that the GA incorporates sharing of information
which transcends the "N times" speed up that you naively expect.
The "N times" effect is useful initially to increase the chance of a good
random choice but the sharing of information effect is not directly
proportional to N. The sharing of information effect is similar for a
range of (smallish) populations and may even become less efficient for
larger populations.

These are just my preliminary impressions however.
