Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!EU.net!uunet!hobbes!earth.armory.com!jon
From: jon@armory.com (Jon Shemitz)
Subject: Re: large random space
Organization: Midnight Beach
Date: Wed, 2 Nov 1994 17:55:53 GMT
Message-ID: <CynJ57.rA@armory.com>
References: <1994Oct27.163228@ebi.ac.uk>
Sender: news@armory.com (Usenet News)
Nntp-Posting-Host: deepthought.armory.com
Lines: 47

I'm a little reluctant to post answers to this newsgroup where so many 
know much more than I, but I don't see any other answers, and my r)eply 
bounced, so ....
 
notredame@ebi.ac.uk asks:
 
> I am trying to apply GA to the search of a huge space, obviously non 
> linear. 
> 
> Initialy, I am dealing with an object which is made of X subobjects, 
> each of them can be in Xi configurations. Thus the total of possible 
> configurations for my object is the product ( Xi). Given a set of Xi, it
> is possible to generate the object and then to score its fitness by an 
> objective function. I have designed a very basic GA, where each 
> individual is made of X genes, each coding for the value of one 
> subobject, and thus, each chromosome coding for one object only ( the 
> subojects cannot be permuted). Of course, this doesn't work. The space 
> is huge and the size of each gene is about 200 bits, for ten subobjects 
> ( 2000 bits for the whole chromosome). Practicly, I cannot handle more 
> than a hundred individuals per Generation. So I d like to know if anyone
> has had the same kind of problem. I beleve that a poor coding ( N-> one 
> state) is responsible of the pour efficiency. Are there any systematic 
> ways, for making a clever coding, and try to bring the problem in a more
> linear space ( at least less stochastic)?
> 
> Any idea is welcome especialy if you have a good reason to tell me that 
> this kind of approach is hopeless for a space too large and too random.
 
I am not totally sure that I understand your problem description here.
 
Why does each gene take 200 bits?  Are there 2^200 possible 
configurations for each object?  Or are there 200?
 
I'm not sure this addresses your question or not, but if there are Xi 
possible configurations, then your gene need only take Lg(Xi) bits.  (If 
the number of possible object configurations is not a power of two, you 
just decode this as a base i number.  For example, if you have seven 
possible configurations and 3 objects, the decimal number 123 
corresponds to the base seven number 234 or object 1 in configuration 2, 
object 2 in configuration 3, and object 3 in configuration 4.)  
 
In general, every extra bit in the gene doubles the search space, so you 
want to keep the bit count down.

-- 

http://www.armory.com/~jon
