Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!europa.eng.gtefsd.com!newsxfer.itd.umich.edu!gatech!darwin.sura.net!news.Vanderbilt.Edu!NewsWatcher!user
From: goldenjb@ctrvax.vanderbilt.edu (jim golden)
Subject: Re: Choice of fitness function
Message-ID: <goldenjb-181094104459@129.59.170.62>
Followup-To: comp.ai.genetic
Sender: news@news.vanderbilt.edu
Nntp-Posting-Host: 129.59.170.62
Organization: vanderbilt
References: <364o52$e6h@newsbf01.news.aol.com>  <Pine.SUN.3.90.941006131909.18254B-100000@euston> <37181i$t4l@infa.central.susx.ac.uk>
Date: Tue, 18 Oct 1994 15:29:06 GMT
Lines: 31

In article <37181i$t4l@infa.central.susx.ac.uk>, inmanh@cogs.susx.ac.uk
(Inman Harvey) wrote:

> Dr R Summers (sg346@city.ac.uk) wrote:
> :   What has been published in the area of convergence of a GA and choice
> : of fitness function for a given optimisation problem?
> 
> :   We are analysing a large data set using ANNs whose topology has been
> : optimised by applying a GA - at the moment we are using a function of the
> : error, complexity of the network and learning time - is this optimal?
> 
> If you have a 'given optimisation problem' then presumably you have been
> 'given' a specification of what has to be optimised. If the decision is to
> optimise a function of the error, then use that as the determinant of the
> fitness function. If the decision is to optimise some specific function 'of
> the error, complexity of the network and learning time' then use that specific
> function as the determinant of the fitness function instead.
> 

This raises some interesting questions regarding optimizing neural
networks.  My committee and I are going around over the term
_optimization_.  What does it mean to optimize a networks?  Minimize the
error?  Maximize the correct calls in a classifier system?  Decrease
learning time?  I don't think anyone has answered this in a straight
forward way, and I think its problem dependent.  I would be very interested
in other opinions, especially from people attempting to _optimize_ a
network using GAs for large, real-world problems.  What cost function and
what constraints are you using?

Jim Golden
goldenjb@ctrvax.vanderbilt.edu
