Newsgroups: comp.ai.alife
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!hookup!swrinde!pipex!sunic!news.chalmers.se!news.gu.se!gd-news!d6217.shv.hb.se!sa209
From: sa209@utb.shv.hb.se (Claes Andersson)
Subject: Re: Evolvable Fitness Formula
Message-ID: <sa209.81@utb.shv.hb.se>
Keywords: Evolvable fitness formula, genetic algorithms, neural networks
Sender: usenet@gdunix.gd.chalmers.se (USENET News System)
Nntp-Posting-Host: d6217.shv.hb.se
Organization: Department of Scocial Science
References: <3bcu1o$sa8@belfort.daimi.aau.dk> <3bdg3g$mhs@gap.cco.caltech.edu>
Date: Mon, 28 Nov 1994 23:10:15 GMT
Lines: 40

In article <3bdg3g$mhs@gap.cco.caltech.edu> brown@altair.krl.caltech.edu (C. Titus Brown) writes:

>In article <3bcu1o$sa8@belfort.daimi.aau.dk>,
>Henrik Hautop Lund <hhl@daimi.aau.dk> wrote:
>>In reference to the ongoing discussion and broad interest on the 
>>newsgroup comp.ai.genetic, I would like to propose the following to
>>the ALife community, too, and invite people to join the discussion.
>>
>>EVOLVABLE FITNESS FORMULA.

>[ munch ]

>Could you define fitness formula in the context in which you're using it?
>Precisely what relation does it bear to the fitness landscape, in particular?

>It also sounds like you're using it as an observable, not as a calculation
>that is fed back in - GAs usually make the calculation and then use it to
>base the next generation's population structure.  Is this what you mean?

 It think it is important to make a distinction between simulation of life 
and utilizing lifelike systems for optimizations. The entire landscape model 
falls when the fitness formula is emergent from the animates traits. This is 
because all animates is located at one place each in a landscape that is 
defined by where it is itself and where all others is. The landscape 
determines the landscape.. the hen and the egg.. or a paradox. Well, not 
quite but still meaningless.

 But who says that the landscape model is divine?.. the landscape is 
good for visualizing the hillclimbing or to map the maximas of simple GA/EP'
s.. however, if one use many genes it becomes totally without use. Therefor 
the loss of the landscape is no loss if one use evolvable fitness formulas 
since these are only of any use in complex lifelike simulations and in 
these, the landscape in of no interest. On the contrary, when one wish to 
optimize a clearly defined algorithm, evolvable fitness would only produce 
undesired selection pressures. The animates would co-adapt and communication 
between them etc. would present a selection pressure. Why would one want 
that? If they didn't interact, they'd have no environment except from the 
defined one.

Claes Andersson. University of Bors. Sweden
