Newsgroups: comp.ai.genetic
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!newsflash.concordia.ca!news.nstn.ca!ott.istar!istar.net!van.istar!west.istar!n1van.istar!van-bc!nntp.portal.ca!newsfeed.direct.ca!hunter.premier.net!netnews.worldnet.att.net!ix.netcom.com!news2.noc.netcom.net!noc.netcom.net!netcom.net.uk!xara.net!agate.xara.net!Aladdin!rmplc!yama.mcc.ac.uk!loki.cf.ac.uk!news
From: Marc Bright <brightms@cf.ac.uk>
Subject: Parent Selection (long question)
Sender: news@cf.ac.uk (Usenet News user)
Message-ID: <31F7C9DE.69B@cf.ac.uk>
Date: Thu, 25 Jul 1996 19:24:14 GMT
X-Nntp-Posting-Host: d164.elsy.cf.ac.uk
Content-Transfer-Encoding: 7bit
Content-Type: text/plain; charset=us-ascii
Mime-Version: 1.0
X-Mailer: Mozilla 2.02 (Win16; I)
Organization: University Of Wales Cardiff
Lines: 27

This a question related to the best method of parent selection in search spaces with non-continous fitness 
profiles.

My fitness function is (A*A)*B, so A has a much larger effect on the fitness than B. In my GA initial 
improvements in fitness are made through A being increased, later improvements, usually after A has reached it 
best value, are made by increasing B.

The problem is, B is incremented in very small increments. The best value of A can yield a fitness at many 
different locations in the search space, but only one of these will give the maximum best fitness through an 
increase in B. Currently I have the problem of getting stuck in local maxima. Even with large population sizes 
the population congregates around the local maxima provided by the best A value, and consequently is too far 
away (in solution space terms) from the real maxima which will be provided by an increase in B.

I am using roulette wheel selection and was wondering if anyone can suggest an alternative method, specifically 
to reduce the problem of population saturation.

I am using problem specific GA strings (non-binary) and also problem specific operators (no random muation or 
crossover) so I cannot just diversify the population by increasing the mutation rate. Also, the fitness 
function represents a real world quantity, so the scaling of B is not a particularly useful option.


-- 
--------------------------------
MARC BRIGHT , CARDIFF, WALES, UK
BRIGHTMS@CARDIFF.AC.UK
HOMEPAGE: http://vlsi2.elsy.cf.ac.uk/bright/home.html
JOHN CARPENTER PAGE : http://vlsi2.elsy.cf.ac.uk/bright/carp.html
