Genetic Algorithms Digest    Monday, 9 December 1985    Volume 1 : Issue 2

 - Send submissions to GA-List@NRL-AIC.ARPA
 - Send administrative requests to GA-List-Request@NRL-AIC.ARPA

Today's Topics:

	Request for References
	Dewdney Article
	Proposal for a Comprehensive GA Bibliography
	Some GA and Classifier System References

------------------------------

From jpo%computer-science.nottingham.ac.uk@ucl-cs.ARPA  Thu Nov  7 15:28:04 1985
Return-Path: <jpo%computer-science.nottingham.ac.uk@ucl-cs.ARPA>
Received: from computer-science.nottingham.ac.uk by 44d.Cs.Ucl.AC.UK 
           via Janet with NIFTP  id a000902; 6 Nov 85 11:28 GMT
To: GA-List@nrl-aic.arpa
Subject: Re: GA-List Digest V1 #1
In-Reply-To: Your message of Mon, 28 Oct 85 16:24:55 est.
	     <8510282124.AA22809@nrl-aic>
Date: 06 Nov 85 11:27:02 GMT (Wed)
Message-Id: <205.500124422@robin>
From: Julian Onions <jpo%computer-science.nottingham.ac.uk@ucl-cs.arpa>
Sender: jpo%computer-science.nottingham.ac.uk@ucl-cs.arpa
Status: R


As an introduction to this area, could someone with sufficient knowledge
in this area put together a list of references to Genetic Algorithms.
The quoted references mentioned don't appear in our library and so
I'd be interested in other sources of information on these Algorithms,
or a pointer to an electronic copy if such things exists.

Julian.


[ A number of the members of this list have references in machine-readable
  form and I hope they will respond.  However, see discussion below.
]

------------------------------

Date:	Friday 22 Nov 85
From:	GA-List Moderator
Subject: Dewdney Article

Well, I finally got a chance to sit down and read Dewdney's article
on GAs in the November issue of Scientific American.  I'm left with
mixed feelings.  The discussion seemed very uneven and hastily done
at times.  On the positive side, it was nice to get the kind of
exposure implied by Scientific American.  From a technical point of
view, I was pleasantly surprised at the treatment of crossover.  In
addition to nice visual graphics, its role and importance relative to
mutation was reasonably well stressed.  However, I am concerned
that the article per se can leave readers with a somewhat unbalanced
view of GAs.  I've seen this happen before, so here are some thoughts
and/or observations:

  1.  In choosing simple finite state automata as examples of how
GAs work and what kinds of problems they might be good for, one must
be careful not to make the space to be searched so small that
exhaustive search might be perceived to be a more effective strategy.
In Dewdney's case N-state flibs define a search space of size (2N)**2N,
so even 4-state flibs yield search spaces of size 2**24.  However,
a more subtle "mistake" can be made by choosing a large search space
but failing to note that solution points are highly dense.  Hence,
random or systematic search may again be a perfectly viable alternative.
In Dewdney's case, there are lots of 4-state flibs that can predict
short period environments.  But as the period length increases, the
number of perfect 4-state predictors decreases dramatically.  This
has lead many of us to use random search times as a raw measure of the
difficulty of a search problem (rather than search space size) and,
as a baseline measure of performance, evaluate GAs in terms of their speedup
relative to random search.

  2.  Population size is a very non-linear parameter with respect to
the performance of GAs.  There is considerable evidence that populations
smaller than 20 suffer badly from the effects of too few hyperplane
samples and are prone to premature convergence.  In Dewdney's case,
population sizes of 10 were used and his graphs show the classic
symptoms of premature convergence after only 200-300 samples.  The moral
here is that when scaling back a GA for explanatory purposes or for running
on your PC, don't skimp on population size!  Sizes in the 40-60 range
generate much more robust and interesting behavior.

  3.  In one's eagerness to get GAs to find "the best" solution,
there is a strong temptation to implement very selective mating
schemes in which only the strongest are mated and the weakest deleted.
Dewdney allowed only the current best to mate with a random partner
to produce via crossover an offspring which replaced the weakest.
There is a very delicate balance here.  Too much emphasis on the
current best turns GAs into hill climbers.  Reread Holland's discussion
of the importance of reproduction in proportion to observed payoff.

  4.  Finally, don't short change the theory behind GAs.  When trying
to convey GAs to others, it's all too easy to get caught up in the many
implementation issues, leaving the impression that GAs are defined by
an implementation rather by a strong body of theory.

So, where do we go from here?  I am planning to send a note to Dewdney
thanking him for the article and pointing out in a constructive tone
some of these issues.  Should we suggest a followup column in which
we provide him with additional "interesting" material?  Should we
take a more aggressive posture and go for a full blown article?

Comments and suggestions encouraged!

	Ken

------------------------------

From booker  Fri Dec  6 17:15:08 1985
Return-Path: <booker>
Date: Fri, 6 Dec 85 17:15:08 est
From: Lashon Booker <booker>
Message-Id: <8512062215.AA06740@nrl-aic>
To: GA-List
Subject: bibliography
Status: R

I am volunteering to put together a comprehensive bibliography of
research about genetic algorithms.  There are undoubtedly many
technical reports, dissertations (from places other than U. of Michigan),
and published papers that I'm not aware of.  Please send your list of
non-obvious references to booker@nrl-aic.arpa.  I will merge the results
and send them to this list.

Lashon

------------------------------

From @MIT-MULTICS.ARPA:Richard_K._Belew@UMich-MTS.Mailnet  Mon Dec  9 04:23:13 1985
Return-Path: <@MIT-MULTICS.ARPA:Richard_K._Belew@UMich-MTS.Mailnet>
Received: from UMich-MTS.Mailnet by MIT-MULTICS.ARPA with Mailnet id <2680419942884526@MIT-MULTICS.ARPA>; 09 Dec 1985 04:05:42 est
Date: Sun, 8 Dec 85 16:36:06 EST
From: Richard_K._Belew%UMich-MTS.Mailnet@MIT-MULTICS.ARPA
To: ga-list@nrl-aic.arpa
Message-Id: <1041026@UMich-MTS.Mailnet>
Status: R


     Well, here's my attempt at getting something going on this digest.
 I finally got together all the Genetic Algorithm and Classifier System
 references I know of, and they're listed below. These are articles I've
 bumped into, plus some references John Grefenstette passed around about
 six months ago, plus all the articles in the GA Conference Proceedings,
 plus the recent Dewdney column in Scientific American.

     As some of you may know, my thesis investigates learning as applied
 to the Information Retrieval task. For that reason, I've become
 familiar with bibliographic software systems; these references all
 came out of that system. I mention this for a couple of reasons. First,
 if any of you have references I don't have listed below, I'd appreciate
 getting them. The other reason is that it is not too difficult for
 me to spit these citations out in other forms, if any of you might
 prefer something else. I chose a real austere format which won't
 pass any editorial reviews but meant least characters flowing through
 the net. But if any of you could use some other form (eg, %REFER
 or BIBTeX database formats), that wouldn't be too much of a problem.


     Rik Belew

     Rik%UMICH-MTS.MAILNET@MIT-MULTICS.ARPA

     c/o Computer Science Dept.
         2045 East Engineering Bldg.
         University of Michigan
         Ann Arbor, MI 48109
         313 / 764 -8504
         313 / 996 -5640 (home)


[ As moderator, I have taken the liberty to delete Rik's rather lengthy
set of references which appeared at this point.  Lashon is in the
process of incorporating them into a larger list which will appear as
a special issue in the near future.  If you need something quickly,
contact Rik or Lashon directly.

I would like to see some discussion as to the format of this bibliographic
database.  I tend to favor a simple REFER-like format which allows
for the inclusion of keywords, comments, summaries, etc.  For example,

	%T  the title
	%A  the author(s)
	%K  keywords
	.
	.
	.
]

------------------------------

End of Genetic Algorithms Digest
********************
