Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!Germany.EU.net!EU.net!sun4nl!news.nic.surfnet.nl!utciva.civ.utwente.nl!usenet
From: "Albert van Breemen" <s9101799@mail.student.utwente.nl>
Subject: Have you had problems with GA??
Message-ID: <68815.s9101799@mail.student.utwente.nl>
X-Minuet-Version: Minuet1.0_Beta_17A
Sender: usenet@Utwente.NL ((UT-utciva)USENET News System)
X-Popmail-Charset: English
Reply-To: <a.j.n.vanbreemen@student.utwente.nl>
Organization: University of Twente, Enschede, The Netherlands
Date: Mon, 5 Dec 1994 21:54:30 GMT
Lines: 79

On Mon, 5 Dec 1994 14:10:16 GMT, 
A.J. Swann  <A.J.Swann@csc.liv.ac.uk> wrote:

>I don't know much, and I'll watch your thread with interest so thanks for
>starting it. For what it's worth, I've heard that a genetic algorithm is
>what you use when you don't know what you're doing!

Here lies one of my problems. I think many people uses GA because 'it's 
simple'. Well it isn't. You have to understand the problem. Specialy when 
you want to code/decode your problem. Most of the time one will find some 
solution, but if the problems are bigger will they still find a solution?? 


>That is, if the problem is easy enough to use another optimizer, use it; 

I agree! However if you have to choose between learning Pontryagin and 
Riccati or GA one will choose for GA because it's simple. That's a big 
problem. It's just like ANN. If one has a problem then using ANN is a lot 
easier then trying analysing the problem. One big reason is the lack of 
probability knowledge. For example: if you have to built a spotdetector 
then using ANN is easier because "if it don't work then the ANN cannot 
learn it, I've done my best". Blame it on the ANN! Well I don't believe in 
that. Using probability math work very good for this problem and you have 
some proof that it will work rather then just trying to something that will 
work.

>if you have domain knowledge, build an optimizer out of that [although that might incorporate
>a GA if the problem is still hard eough]; and for hard problems that are
>very old, there are usually better optimizers than the GA. 
>I'm afraid I have no references for examples of wrong-headed GA approaches 
>using e.g. a bad representation. Ah, I tell a lie; there is one. See
>
>
>"Macro-cell and module placement by genetic adaptive search with
>bitmap-represented chromosome." Heming Chan, P. Mazumder and K. Shahookar,
>In "INTEGRATION, the VLSI journal" 12 (1991) 49-77 (Elsevier). At the end
>they write,
>
>        "Our implementation suffered from non-linear interactions between the genes,
>and the resultant degrading of search efficiency, and we need to find a better
>genetic coding for the placement problem. [sic]
>
>We conclude that the genetic algorithm is a promising searching method for the
>cell placement problem that warrants further investigation."
>
>
>They only tell you this after you've waded through the article! Typical,
>I'm afraid. There's little incentive to shout about negative results. 
>

Well here lies a problem. Most of the time one will not tell what the 
effort was to find a solution with GA. How many times went something wrong? 
I think that we can learn more if everyone tell us about their problems. 
Until now I haven't done any simulation because I wanna understand GA 
better. Although I hope soon to list some results of simulation. I want to 
analyse different code-systems (binairy, gray random), place of building 
block, dependency of buildingblocks, correlation between the offsping's 
performance and the parents performance, ect.) for optimizing an polynomal 
fitness function. 


>Hope that helps!
>
>
>Swanny.
>

I hope too!



================================================
A.J.N. van Breemen
lipperkerkstr 128
7511 DD  ENSCHEDE, The Netherlands
tel: 053-317819
email: a.j.n.vanbreemen@student.utwente.nl
================================================
       
