Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!gatech!howland.reston.ans.net!EU.net!uknet!newsfeed.ed.ac.uk!ainews!dave
From: dave@aisb.ed.ac.uk (Dave Corne)
Subject: Re: Right poulation size
Message-ID: <D6zJ8r.CD@aisb.ed.ac.uk>
Sender: news@aisb.ed.ac.uk (Network News Administrator)
Organization: Dept AI, Edinburgh University, Scotland
References: <3k2qtf$npj$2@mhadg.production.compuserve.com> <D6v <D6vp92.BnB@freenet.buffalo.edu>
Date: Thu, 13 Apr 1995 17:58:03 GMT
Lines: 51

  Salvatore R. Mangano wrote, some time back:
>>>>In general, you should pick the largest pop size that you can 
>>>>afford for the problem at hand. 

  I answered:
>>>  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

 Ralley David Thomas chipped in:

>>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 

  Yes, I'm familiar with that -- that's not quite the point though:
 Goldberg, Deb, & Clark's sizing work tries to come up with an
 estimate of the population size needed to give you an adequate measure of 
 confidence in your building-block sampling. This, of course, does not 
 have much to do with _best_ population size -- such depends on lot more,
 and is probably going to be an empirical thing for a good while yet. 
 It's not even well-defined anyway. 
  For example, if GDC-sizing gives you a pop of 100 for a certain
 problem, this means that less-than-100 would probably be very risky,
 but it has little to say about how much better greater-than-100 might
 be. In the sense of quality/time tradeoffs re the given problem and 
 one's particular implementation, the best population size here 
 might actually be 728.

   There's much left to do with G,D,& C's sizing work anyway, 
 of course. Notice how it's rather unlikely to come out with `1' in 
 those (many ) cases where it turns out that SH would be better than
 a GA. This is, again, mainly because it doesn't relate to ideal
 population size {\em as far as eventual solution quality in 
 reasonable time is concerned} , but ideal population size {\em as
 far as ironing out sampling-noise to an acceptable flatness is
 concerned}. Not the same thing, unfortunately.

----------------------------------------------------------------------
   _______       __   Dave W. Corne,
 /   ------.   / ._`_          Department of Artificial Intelligence
|  /         ~--~    \                University of Edinburgh,
| |             __    `.____________________ _^-----^ Forrest Hill,  
| |  I=|=======/--\=========================| o o o | Edinburgh,
  |  I=|=======\__/=========================|_o_o_o_| EH1 2QL, U.K.
 \|                   /                       ~    ~
   \       .---.    .   031 650 3084 - 0324 613945 - Dave.Corne@ed.ac.uk
     -----'     ~~''
----------------------------------------------------------------------
