Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.mathworks.com!newshost.marcam.com!zip.eecs.umich.edu!umn.edu!news
From: "James Albert Larson" <larso171@maroon.tc.umn.edu>
Subject: Re: culling 'fitness function' use
To: dlw@isd.rrassoc.co.uk
Message-ID: <56306.larso171@maroon.tc.umn.edu>
X-Minuet-Version: Minuet1.0_Beta_16
Sender: news@news.cis.umn.edu (Usenet News Administration)
Nntp-Posting-Host: dialup-5-127.gw.umn.edu
X-Popmail-Charset: English
Organization: University of Minnesota, Twin Cities
Date: Mon, 3 Jul 1995 18:05:12 GMT
Lines: 34

On Thu, 22 Jun 95 19:55:51 GMT, 
Andrew Darlow  <dlw@isd.rrassoc.co.uk> wrote:

>
>     I've got a fitness function that takes about 1/2 hour to run (it uses
>     a SLOW simulator). Hence, it will take 10 hours to evaluate a even 
>     a small population of 20 individuals (solutions).  It will take a
>     week for the GA to do even a small number of iterations ( aprox. 16 ). 
>     It will take months to proces a larger population (approx. 40 
>     individuals) with more iterations (100+);
>

Goldberg's 1989 book on p. 138 reports on an experiment where two 
xray pictures of the same patient at slightly different points in time are 
being compared on a pixel by pixel basis.  The goal being to determine by 
how much to translate / rotate one image to exactly match the other 
position-wise (the patient probably has moved slightly between the two 
xrays).  They could sample 10000 pixel differences between the xrays.  But 
with a given solution time constraint, it was determined that the optimum 
was to compare only 10 pixel differences, but over 1000 times as many 
generations.  This increases the amount of search space explored compared 
at the expense of accuracy.  Pretty amazing result.

The upshot is that a lot of slightly or moderately inaccurate fitness 
values are worth a lot more than a few highly accurate one.  GA's also 
tolerate noisy data very well -- one can consider small or modest 
inaccuracies in measuring fitness values as noise.

One might tolerate speedy somewhat inaccurate methods for measuring 
fitnesses at the beginning of the run, and then use more accurate methods 
later in the run as one approaches the "fine-tuning" stage.

Jim Larson

