
Genetic Algorithms Digest   Monday, 22 October 1990   Volume 4 : Issue 25

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

Today's Topics:

	- Updating the GA Mailing List
	- Re: Evolver
	- GA workshop abstract: GAs FOR REAL PARAMETER OPTIMIZATION
	- New GA book info and change of address
	- Some answers to classifier system questions

******************************************************************************

CALENDAR OF GA-RELATED ACTIVITIES: (with GA-List issue reference)

Conference on Simulation of Adaptive Behavior, Paris (v4n17)  Sep 24-28, 1990
Workshop Parallel Prob Solving from Nature, W Germany (v4n18) Oct 1-3,   1990
2nd Intl Conf on Tools for AI, Washington, DC (v4n6)          Nov 6-9,   1990
4th Intl. Conference on Genetic Algorithms (v4n17)            Jul 14-17, 1991

(Send announcements of other activities to GA-List@aic.nrl.navy.mil)

******************************************************************************

Date: 22 October, 1990
From: Alan C. Schultz (GA-List moderator)
Subject: Updating the GA Mailing List

   We are once again trying to update our mailing list.  The mailing list
   serves two purposes.  One, it allows us to have a land address so that
   conference announcements and other relevant materials can be sent to you.
   Two, it allows us to make sure you have a valid email address, and also
   should the email fail, we can contact you to try to resolve the problem.
   This information will only be used for the above purposes.

   We will be processing your responses electronically, so please follow the
   instructions carefully.

   Copy the lines between the two "CUT HERE" lines to a file, and fill in
   each field according to the following list.  Do NOT remove the keys at the
   beginning of the lines.  Leave a space between the key and your response.
   Then mail back the information, with the subject line of your response
   containing "RESPONSE" and nothing else.  Also, do not put ANY other
   material in your mail message, including signature lines.

   Mail the response to ga-list-request@aic.nrl.navy.mil.

List of keys and responses:
	%FN 	First name (you could also put middle initial).
	%LN	Last name.
	%AD	Address line. Use as many as you need, adding them if
		neccessary.  Remove those you do not need.
		Make sure to include your country.
	%EM	Email address from the point of view of the internet.
	%PH	Complete phone number.

Example:

%FN Alan C.
%LN Schultz
%AD Code 5510
%AD Naval Research Laboratory
%AD Washington, DC  20375-5000
%AD USA
%EM schultz@aic.nrl.navy.mil
%PH (202) 767-2684

%++++++++++  CUT HERE ++++++++++++++++++++
%FN
%LN
%AD
%AD
%AD
%AD
%AD
%EM
%PH
%++++++++++  CUT HERE ++++++++++++++++++++

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

Date: Thu, 27 Sep 90 10:50:13 -0400 (EDT)
From: Leslie Burkholder <lb0q+@andrew.cmu.edu>
Subject: Evolver

   > Date: Tue, 4 Sep 90 13:34:22 EDT
   > From: Rick_Riolo@um.cc.umich.edu
   > Subject: Evolver by Axcelis, Inc?
   >    Does anyone know anything about a commercial package called
   >    Evolver, by Axcelis Inc?  There was a brief article about
   >    it in the New York Times 29 Aug 1990.  The article starts:
   >	 A Seattle company has developed a personal computer
   >     program that "evolves" the best solution amoung
   >     various models run on spreadsheets by financial 
   >     planners for businesses. ... 

   Not quite a reply to the query but further information about it can be
   obtained from
   Axcelis, 1406 Western Ave, Seattle WA 98101; phone 1 800 Axcelis.
   The first available version is for WingZ on the Mac.
   Leslie Burkholder

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

Date: Fri, 28 Sep 90 10:29:42 MDT
From: cs_ahw%umt01.dnet@deimos
Subject: GA workshop abstract

[Ed. Note:  This is an abtract from the FOGA workshop this past summer.]

     GENETIC ALGORITHMS FOR REAL PARAMETER OPTIMIZATION

	   This paper is concerned with the application of genetic algorithms
	to optimization problems over several real parameters.  It is
	shown that k-point crossover (for k small relative to the number
	of parameters) can be viewed as a crossover operation on the
	vector of parameters plus perturbations of some of the parameters.
	Mutation can also be considered as a perturbation of some of the
	parameters.  This suggests a genetic algorithm that uses real
	parameter vectors as chromosomes, real parameters as genes, and
	real numbers as alleles.  Schemata are defined for this algorithm,
	and it is shown that Holland's Schema theorem holds.  Experimental
	results are also given that indicate that this algorithms
	sometimes does much better than a binary-coded genetic algorithm.
  	   One advantage of using a real-encoding over a binary encoding
	for real-paramter problems is that it is more efficient--real
	encoding avoids any necessity for converting from binary or Gray
	codes into floating-point numbers.
	   Real-encoding also allows for a better understanding of how
	genetic algorithms work in real-parameter spaces.  This intuition
	can lead to alternative mutation and crossover operators, and to a
	better understanding of the strengths and weaknesses of genetic
	algorithms.

   Alden Wright, Computer Science, University of Montana, Missoula, MT 59812
   umt!cs_ahw@apple.com

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

Date: Fri, 31 Aug 90 8:33:36 EDT
From: Dave Davis <ddavis@BBN.COM>
Subject:  New GA book info and change of address

  I would like to put these two items of information on the GA-list:

  1. The Handbook of Genetic Algorithms is scheduled to come out late this
  year.  It is a 400-page book about applying genetic algorithms to
  real-world problems.  There is a GA tutorial by me and there are thirteen
  chapters about real-world applications.  The book is being published by Van
  Nostrand Reinhold, and will cost ~$50.  A computer diskette will also be
  available optionally for ~$50 plus postage.  The diskette will contain John
  Grefenstette's current, updated version of GENESIS, a Common Lisp system
  called OOGA (the Object-Oriented Genetic Algorithm) that works out the
  examples in the tutorial, and documentation. 

  2. I am leaving Bolt Beranek and Newman today to begin working on my own as
  a consultant in the genetic algorithm field.  This is a market niche I hope
  will support a business.  I will be off the net for a while, and would like
  to let you know my new address and phone number:

  Lawrence (David) Davis
  Tica Associates
  36 Hampshire St.
  Cambridge, MA 02139
  (617) 864-2292

  ("Tica" comes from geneTIC Algorithms).

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

Date: Thu, 4 Oct 90 07:37:08 EDT
From: Rick_Riolo@um.cc.umich.edu
Subject: Some answers to classifier system questions
 
 Here are my answers to some questions that were asked in ga list recently.
 

   I. Thomas Antalffy asked about crossing lions and antelopes.

      Though I don't agree with his characterization of the search space
      as smooth vs. erratic, I do think he raises a good point with
      respect to classifier systems.  Here are some references
      that may help:

      This point has been raised in the context of GA's
      by Goldberg and Richardson (in the 1987 ICGA's proceedings,
      a paper on multimodal function optimization).

      This has also been raised in the context of classifier systems
      by Robertson and Riolo (1988 special issue of Machine Learning).


      I would guess it has been discussed by other classifier system
      researchers as well (e.g., Booker's papers on Gofer).


   II. P. Herrera-Boyer askes several questions about classifier systems.

      Q1. Bids are not restricted to 0 to 1 (though perhaps they should be).

	  The detectors must assign a value to each message which corresponds
	  to the bid made to post other messages.  This can be done in a lot
	  of ways: (a) assign a constant; (b) assign a value that is the
	  average of bids on last step (or some multiple there of, or use
	  the average over recent history, etc.); (c) have a subsystem
	  that decides what value to assign based on extra-classifier system
	  factors (eg. for a robot or organism, assign a larger value
	  to moving things, to large things, or whatever is thought a priori
	  to be important).
	  The point is just that you raised: the value assigned can be
	  used to bias activity to or from rules that process detector messages.

	  Note that the role of support in bids is not a closed book.
	  For example, if bids do range over positive reals, and if
	  support for rule A is defined as the sum of bids made to 
	  post messages matching A, and if the bid is calculated as
	      bid = k * b * S * x
	  where k is constant, b is specificity, S is strength, and x
	  is support, then bids could be greater than strenghts!
	  Thus something probably should be done to keep x in the range 0..1,
	  e.g., normalize it somehow.  Of course that operation will
	  not be completely parallelizable on a SIMD machine.


      Q2. First I am confused about your premise ("If my system posts
	  only 1 message after competition,...").
	  Is that a constraint you have imposed, or are you just
	  talking about a situation where that happens to be the case?
	  I would suggest that in general you should allow more than
	  one message to be posted at one time: that allows the system
	  to process information in parallel, as I think Holland intends.

	  At any rate, if I understand your proposal, I think one problem is
	  that in order to use it, you must virtually post all messages
	  from all matched rules.  Otherwise, you won't know which rules get
	  supported at the next step.
	  But maybe that cost is ok if there are other reasons supporting
	  your proposal.

	  Can you give an example (in terms of some application)
	  where support from rules that have lost the competition should
	  bias the activity at the next step?

	  One way to think of competition is as cross inhibition.
	  How does your proposal look when thought of in those
	  (neural network) terms?


      Q3. The usual way to select the winner(s) is probabilistically.
	  I know in my experience, choosing the max *often*
	  leads to the system getting stuck on "false peaks".

	  If you have mutliple "channels" do the same for each, perhaps.

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