Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!delmarva.com!newsfeed.internetmci.com!news.sprintlink.net!simtel!harbinger.cc.monash.edu.au!aggedor.rmit.EDU.AU!matilda.vut.edu.au!rhh
From: rhh@matilda.vut.edu.au (Robert Hinterding)
Subject: Re: Urgent:Advice Needed on GA application
Message-ID: <DECyup.F4C@matilda.vut.edu.au>
Organization: Victoria University of Technology
References: <41qlos$m49@infoserv.rug.ac.be>
Date: Mon, 4 Sep 1995 01:59:13 GMT
Lines: 73

"Park, TAE YUN" <tp@elptrs7.rug.ac.be> writes:

>Hi everyone,

>The reason why I write this message to you is that I need your 
>help and advice for my Ph.D. work.  Please take some miniutes to read 
>this message if it doesn't disturb your work too much. 

>First of all, perhaps it would be much better if I explain you briefly 
>about myself and my Ph.D. problem.  As you can see at the end of this 
>message, I'm a Ph.D. student of State University of Ghent, Belgium, 
>in the field of chemical engineering. My current research interest is to
>develop a kinetic model which can describe a complex catalytic reaction 
>system.
	[snip]
>Thank you so much to read my message on my problem so far, and now I'd 
>like to ask some questions, as follows:


>          1. This is a bit fundamental, but I have to ask you since I have only 
>             4 months experience of GA even it was quite intensive: Do you 
>             think GA can solve my problem ? If not, what about ES then? (this 
>             is because somebody whom I discuss my problem through internet 
>             told me that I should consider ES rather than GA since my problem 
>             is real-valued based)

Both GAs and ESs will be able to solve the problem.  For the GA especially,
choosing the right operators is crucial. 

>          2. If the answer of question I-1 is positive, what is the 
>             optimal operator and condition of GA for a minimization of 
>             a objective function with multi-dimension and the existance of a 
>             lot of local minimum? More exactly, which initialization method? 
>             which crossover method? how much crossover rate? which mutation 
>             method? how much mutation rate? what population size?.......and...
>             why??
The best parameter settings can generally only be determined empirically.
But I would suggest that you Gaussian mutation as bit-flip mutation has a lot
of problems for difficult functions.  It does not give a good spread of
changes to the function variables, and the representation length (number of
bits you allocate to each variable) determines the mutation strength.


>          3. What is the reason of the failed solution of my problem using 
>             genesis 5.0 package or simple GA. Do you agree with my opinion 
>             written above on the reason why I had got wrong solutions?? 

Genesis and SGA are suuitable for simple functions and experimentation.
You need to use a GA that used Gaussian mutation, or Evolutionary Strategies.
Try "genocop".  
>          4. What do you think about self-adaptive GA for the probabilities 
>             of corssover and mutation? Is this really possible to enhence 
>             the GA ability? (This is because one of the packages 
>             obtained from internet have ability to adapt these probabilities 
>             during running the GA. However, this package doesn't work at 
>             all.)  
I have implemented self-adaption in GAs and IT DOES WORK.  The problem is
that if the problem is deceptive, it can be mislead.  It works very well
on problems such as the multi-dimension sphere function which are "smooth"
and not deceptive.  Contact me if you want a preprint of a paper where this is
demonstrated.


>Thank you again, and I'm HIGHLY looking forward to getting ANY news from you 
>soon. 

Hope that helps,

-- 
Robert Hinterding
VICTORIA UNIVERSITY OF TECHNOLOGY            Email: rhh@matilda.vut.edu.au 
P.O. Box 14428, Melb Mail Centre             Fax:   +61 3 688 4050
AUSTRALIA 3000                               Phone: +61 3 688 4686                                 
