
Genetic Algorithms Digest    Wednesday, 11 October 1989    Volume 3 : Issue 15

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

Today's Topics:
	- Call for Papers
	- Re: Partial Matching
	- Convergence of GAs

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

Date: Wed, 4 Oct 89 14:49:48 BST
From: Steve Renals <sjr%eusip.edinburgh.ac.uk@NSFnet-Relay.AC.UK>
Subject: Call for Papers
Organisation: Department of Physics, Edinburgh University, Scotland.
Phone: (44) +31 225 8883 x278

[All correspondance to Ken Sharman, kenshar@vme.glasgow.ac.uk]

                       **** CALL FOR PAPERS ****
 
 
         GENETIC ALGORITHMS, SIMULATED ANNEALLING, NEURAL NETS
 
                                  in
 
              SIGNAL/IMAGE PROCESSING and COMMUNICATIONS
 
 
                  University of Glasgow 8-9 May 1990
 
 
     The  INSTITUTE OF ELECTRONIC AND ELECTRICAL  ENGINEERS  UK&RI
     Section (Joint Chapter in Acoustics, Speech,  Signal Process-
     ing and Communications) is running a two day workshop in  the
     areas  mentioned  above.   The workshop will be held  in  the
     Wolfson  Conference Centre at Glasgow University on 8'th  and
     9'th May 1990.  
 
 
     PAPER PROPOSALS are invited from prospective authors  working
     in  the  areas  of  genetic  algorithms,   neural   networks,
     simulated annealing, and variants thereof applied to problems
     in signal and image processing (including speech),  and  com-
     munications systems.
 
 
     DEADLINES:    Interested authors should submit a 400 word ab-
     stract of their proposed paper to Dr.  K.  C.  Sharman at the
     address given below BEFORE 20'th DECEMBER 1989.  After review
     by  a technical committee the successful  (and  unsuccessful)
     authors will be notified by 31'st January 1990.   Full papers
     (up  to 8 A4 pages) should be 'camera ready'  by  1'st  April
     1990.
 
 
     COSTS:  The costs of the workshop will be as follows (includ-
     ing lunch and refreshments):
 
          Authors:                 70 pounds
 
          Members of the IEEE:    115
 
          Non-Members             175
 
          Students                 25
 
 
     Note:  Attendance at the workshop (including authors) will be
     limited to 140.
 
 
 
     FURTHER DETAILS:   Please contact the workshop organiser:
 
          Dr. Ken Sharman
          Dept. Electronics and Electrical Eng.
          The University
          Glasgow G12 8QQ
          (44) +41 339 8855 x 4902
 
          email: kenshar@uk.ac.glasgow.vme

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

Date: Tue, 03 Oct 89 17:51:38 EDT
From: Stewart Wilson <wilson@Think.COM>
Subject: Re: partial matching

Lashon replies to the question about partial matching in part as follows:

> The only objections to partial matching I'm aware of
> argue that it is an unnecessary complication to the
> basic execution cycle. The claim is that "create"
> operators are adequate for managing situations where
> no rule matches exactly. I obviously disagree, and I
> think I've shown that partial matching can be used
> in a limited way that doesn't complicate the basic cycle.

Though "exact" matching is simpler technically than partial matching,
that is certainly not the main argument for the former.

Exact matching of a condition to a message means that all specified
(non-#) bits in the condition must occur with the same values in the message.
Partial matching allows mismatches of one or more of the specified
bits and computes a corresponding match score.  Lashon's scoring
formulas always give an exact match a significantly higher score
than a partial match.

The argument against partial matching is that in many cases, the
particular feature represented by a specified condition bit may be very
significant in determining the appropriateness of the underlying
classifier.  To quote "Induction" (Holland, Holyoak, Nisbett, and
Thagard, 1986, page 24), 

	...allowing rules to fire when only some of their conditions
	are matched can have treacherous consequences.  Suppose the 
	system has the rule "If the goal is to cross a body of water,
	and it is about a mile wide, and you are a strong swimmer, then
	decide to swim across."  Now imagine that our system, a 
	lamentably weak swimmer, wants to cross a mile-wide lake.
	A partial match will be found...(two of three conditions in
	the most relevant rule are satisfied).  The result, sad to
	say, is a dramatic decrease in our system's life expectancy.

In contrast, the argument against exact matching is that often, because
exact matching is so picky, the system will be stuck with no matching
classifiers when in fact classifiers are present that would respond
appropriately to the situation if only the match rule were slightly 
relaxed.  A second argument, which depends on the application of the
GA to match sets instead of to whole populations, is that partial
matching gives the GA better guidance than exact matching since under
the latter the match set will always be smaller and may be empty.

Booker's basic answer to the Holland argument is to rate exact matches
highly over inexact ones, as already mentioned.

Exact match fans' basic answer to Booker's critique is to have plenty
of classifiers, including many with lots of #'s, so exact matching 
rarely fails.  

Both answers are in turn open to further questions.  My personal opinion
is that the debate is somewhat problem-dependent.  In cases where
classifiers that match actual problem states occupy only a small fraction 
of the space of possible classifiers, partial matching will increase the 
performance of the system and greatly aid the GA in searching the space.  
The necessary alternative under exact matching, as Booker states, is
some form a classifier "create" or "cover" operator, but the value of
this in sparse classifier spaces is not yet clear--for example, if
"create" occurs a lot, the system may just churn.  Similarly, the
"many classifier, many #" approach has problems with rampant 
overgeneralization, in my experience.  However, in cases where most
possible classifiers will match some state or other of the actual problem,
exact matching is appropriate and efficient, since most classifiers 
that are generated will eventually be evaluated.  The BOOLE system
on the multiplexer problems is an example.

Stewart Wilson

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

Date: Tue, 10 Oct 89 11:20:09 -0500
From: androula@helium.ecn.purdue.edu (Ioannis Androulakis)
Subject: Convergence of GAs


  I would like to know if there is any expression for the time
  bound that GAs would need to converge in. I am also interested
  to know the kind of relation that exists between this bound
  and the basic parameters of Genetic search (mutation and crossover
  rate, population size for example).

  Thank you,
  
  Ioannis P. Androulakis
  androula@helium.ecn.purdue.edu

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

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

