Genetic Algorithms Digest    Wednesday, 13 January 1988    Volume 2 : Issue 3

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

Today's Topics:
	- GA-List Host status
	- GA Research Poll
	- A Question about expected number of crossovers
	  needed to produce an given individual

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

Date: 13 Jan 1988 16:10:50 EST
From: John Grefenstette < gref@nrl-aic.arpa >
Subject: GA-List Host status

The 780 at NRL-AIC has finally been converted to 4.3.
We hope that this will eliminate many of the problems
we have had in distributing GA-List in a timely fashion.
Please keep those cards and letters coming.

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

Date: 28 Oct 1987 20:05-EST 
From: Michael.Mauldin@NL.CS.CMU.EDU
Subject: Re: GA Research Poll

    Project:	FERRET (Flexible Expert Retrieval of Relevant English Text)
		One aspect of my dissertation will be a learning component
		for a text classifying system based on Gerald DeJong's
		FRUMP text skimming parser.  The learning component will
		initially be based on a simple GA model, but with heuristic
		knowledge added to the structure generator.

Application:	Natural Language Understanding
       area	Text processing for information retrieval

    General:	Search a space of knowledge structures for
   approach	those which match user concepts about documents.
		(Simple GA with both random and heuristic
		generation of possible structures).

    GA Tool:	Custom software based on previous work of mine
		written up in AAAI-84.

    Results:	[ not yet available ]

   Problems:	There is a fine line between classic GAs and
		"simple" heuristic search, so calling this
		work a "genetic algorithm" is a bit iffy.  The
		biggest challenge will be a low amount of
		feedback from the environment about the utility
		of the structures generated.  There will also be
		a large amount of noise, since the payoff will
		involve user classification of system performance.

Michael L. Mauldin (Fuzzy)		Department of Computer Science
ARPA: Michael.Mauldin@NL.CS.CMU.EDU	Carnegie-Mellon University
Phone: (412) 268-3065			Pittsburgh, PA  15213-3890

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

Date: Thu, 29 Oct 87 10:42 EST
From: Steve Smith <smith@Think.COM>
Subject: Re: GA Research Poll

APPLICATION AREA: Optimization of search in n-dimensional space.
 Specifically using the Travelling Salesman Problem as test for
 the approach.

GENERAL APPROACH: Using a parallel architecture (Connection Machine)
 to run straight genetic algorithms with ARGOT (see Shaefer from the
 last proceedings) to use feedback to improve the manner in which the
 space is searched.

GA TOOL: Custom Software

RESULTS: ARGOT is doing all the right things and results are good on
 small TSP tours; haven't tried for tours larger than 30 cities
 yet though.

PROBLEMS: Not clear yet.

- stephen smith (the other one)
  Thinking Machines Corporation
  smith@think.com

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

Date: Thu, 29 Oct 87 10:42 EST
From: "Stephen Judd, ABD" <JUDD%cs.umass.edu@RELAY.CS.NET>
Subject: Re: GA Research Poll

Here's my response to the poll:
Area--- connectionist learning (Can a GA help me find `weights' for a node
	in a network so that it will learn to perform a given task?)
approach--- mostly analytical at this time.
	may become empirical if my theory involves approximations or
	probabilistic arguments
results--- well, er, ahh... no.
sj

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

Date:     Thu, 5 Nov 87 10:04 N
From: <MAEN@DHDMPI5V.BITNET>
Subject: Re: GA Research Poll

GA activities in Heidelberg:
R. Maenner, J. Hesser
Physical Institute, University of Heidelberg, West Germany

We try to construct minimal Steiner trees using GA's. A minimal tree is the
shortest connection of n fixed points in a plane. Its length can further be
reduced by a proper placement of up to n-2 additional points. This yields a
minimal Steiner tree. Applications are the optimization of networks and the
layout of PC boards and chips. The Steiner tree length is minimized applying
the genetic operators crossover, mutation, and dominant heredity. The GA's
are run on a IBM 3090 using custom software. For performance measurements,
a 3*3 rectangular grid of fixed points is used. With a population size of 50,
a solution is found in 10 - 20 iterations. Currently, modified genetic
operators are investigated for improved convergence.

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

Date: Mon, 2 Nov 87 08:34:20 EST
From: Rick_Riolo@um.cc.umich.edu
Subject: Re: GA Research Poll
Subject: Poll response

Appication:  Machine learning
Approach:    Classifier System
Tool:        CFS-C
Results:     indifferent - good
Problems:    Applying GA to classifier system with limited
             message list (which makes it hard to control
             premature convergence, and to establish chains
             of coupled classifiers);
             representation; parameters abound

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

Date: Thu, 29 Oct 87 17:40:24 CST
From: honavar@cs.wisc.edu (Jughead Jones III)
Subject: Re: GA Research Poll

I am currently not doing any work on genetic algorithms although I am 
thinking of using genetic algorithm like techniques as part of a learning
program that  would use a number of other techniques.

Thanks.

-- Vasant Honavar
Computer Sciences Dept.
University of Wisconsin-Madison
1210 W. Dayton St.
Madison, WI 53706

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

Date: Tue, 8 Sep 87 12:52:21 pdt
From: Tom Dietterich <tgd@mist.cs.orst.edu>
Subject: Question about genetic algorithms

I have a student who is doing some experiments with the genetic
algorithm (applied to concept learning), and I was wondering if either
of you could answer the following question:

Has anyone conducted either some experiments or some analytical work
to estimate the expected number of cross-over operations required to
generate particular organisms?  For example, suppose I have a randomly
generated population of N organisms (each a single classifier).  Each
organism has a genome of length l over the alphabet {0,1,#}.  Suppose
I generate a "goal" organism, also at random.  Now suppose I perform
crossover operations by randomly selecting two of the N organisms in
the usual fashion, and add the resulting organism to the population
(to give N+1 organisms).  What I want to know is, how many crossovers,
on the average, are required, and what factors influence this number
(e.g., if I keep the population size fixed by randomly deleting an
organism, what happens?  If I insist that each organism in the
population MUST have at least one allele in common with the "goal"
organism, what happens?).

My student and I are just getting started in doing GA-related
research.  We've read most of the papers in the 1985 GA conference
plus several of Holland's papers.  But we don't by any means have a
good grasp of the GA literature, so I hope you will be able to give us
some pointers to relevant articles.

Thanks in advance,

Tom Dietterich
Department of Computer Science
Oregon State University
Corvallis, OR 97331
503-754-3273
tgd@cs.orst.edu


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

Date: Wed, 16 Sep 87 16:19:33 1987
From: John Grefenstette <gref@nrl-aic.arpa>
Subject: Re:  Question about genetic algorithms

Tom:

Thanks for your question about genetic algorithms. I think I need some
further information on your application in order to answer your
questions.

You ask about the expected number of crossovers required to generate a
given organism.  The answer depends on the payoff function, since
organisms receiving higher payoffs are likely to generate a larger
number of similar offspring than organisms receiving little payoff do.
You don't say what your payoff function is, so it's hard to be more
specific.

In general, there has been little work done on predicting the number of
generations needed before a given structure is evolved.  The
theoretical work [Holland 75, De Jong 75] mostly shows that GA make
efficient use of the information gathered, but no one has provided
predictive measures for practical systems.  The experimental studies
usually compare GA's with other search techniques.  On the other hand,
there have been some theoretical works (Bethke's 1981 Ph.D., Goldberg's
recent study of "Minimal Deceptive Problems") that attempt to
characterize "GA-Hard" problems for which GA's are not well suited.

There is also the point of view that the result of a GA is the
information distribution in the entire population, not just the best
single individual.  Like natural evolution, GA's pay little attention
to individual survival.  This point of view is evident in Stewart
Wilson's paper that has been accepted in the Machine Learning Journal.
I recommend that paper as a source of ideas on concept learning with
GA's.

As far as additional GA literature goes, I recommend the Proceedings of
the 1987 Conference on GA's, available from

	Lawrence Erlbaum Associates,
	Publishers 365 Broadway
	Hillsdale, NJ 07642
	(201) 666-4110


I'd be interested in more details about the approach you are taking.
Keep in touch.

-- John

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

Date: Wed, 16 Sep 87 13:42:56 pdt
From: Tom Dietterich <tgd@mist.cs.orst.edu>
To: gref@nrl-aic.arpa
Subject:  Question about genetic algorithms

John,

Thanks for your answer to my message.  

RE: Payoff function

For starters, I would like to do this analysis using a flat payoff
function in which organisms are chosen uniformly for cross-over.  I'd
even be curious to know the minimum number of cross-overs required,
given a randomly-generated population (which I suppose would
correspond to the IDEAL payoff function that gave maximum payoff
precisely to those individuals that would yield the minimum-length
sequence of cross-overs). In short, I want to investigate cases where
the payoff function is helping, rather than causing any problems.

RE: population vs. individual

The view that the whole population should be considered is
interesting, but probably a lot harder to analyze.  I'll have to give
some more thought to that approach.

Thank-you for the pointers to the literature.  I'll order the
proceedings today.

Cheers,

--Tom

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

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


