
Genetic Algorithms Digest   Wednesday, December 18 1991   Volume 5 : Issue 38

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

Today's Topics:
	- Re: negative rewards and optimal ML length in LCS  (2 messages)
	- NETtalk task using GAs?
	- Genetic Algorithm Program in Smalltalk
	- TCGA method of payment
	- New Technical Report Available
	- Call for Bin Packing Benchmarks

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

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

 Canadian AI Conference, Vancouver, (CFP 1/7)                 May 11-15, 1992
 COGANN, Combinations of GAs and NNs, @ IJCNN-92 (v5n31)      Jun 6,     1992
 10th National Conference on AI, San Jose, (CFP 1/15)         Jul 12-17, 1992
 FOGA-92, Foundations of Genetic Algorithms, Colorado (v5n32) Jul 26-29, 1992
 ECAI 92, 10th European Conference on AI (v5n13)              Aug  3-7,  1992
 Parallel Problem Solving from Nature, Brussels, (v5n29)      Sep 28-30, 1992

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

**********************************************************************
----------------------------------------------------------------------

From: Rick.Riolo@um.cc.umich.edu
Date: Wed, 27 Nov 91 08:55:36 EST
Subject: Re: negative rewards and optimal ML length in LCS

   I have played a little with negative rewards, and for many problems they
   can help.  Carl Simon (an economist at um) did tests of lots of reward
   settings in cfsc1/letseq1 on a simple letter sequence problem or two and
   as I recall he found mildly negative rewards (for wrong guesses) improved
   the learning rates.  Still, I suspect whether negative rewards helps or
   not will depend a lot on the problem and on the representation (i.e., the
   detector and effector setup, etc.).  If you punish things that are "wrong"
   by some external definition but those things have "building blocks" that
   are useful for the "right" rules, you are shooting yourself in the foot,
   since negative rewards can quickly cause those rules to perish, taking
   their building blocks with them.

   Which brings up one related problem: how can an LCS "remember" things its
   *not* supposed to do?  Typically rules that suggest things not to be done
   get low S, but then they are replaced by the GA.  Then the system has to
   relearn that its not supposed to do whatever that rule suggested!  Its
   true that other rules may have larger S, and so will suggest doing
   something else (and so as a side effect imply not doing something), but
   that seems to me not quite as good as having rules that say "don't do A",
   or having rules that indicate a negative reward can be expected if the
   system does A.

   Re Message List length: I haven't done anything systematic with this. For
   a range of modest list sizes, the length didn't seem to have a significant
   effect, at least for some problems.  For other problems, a minimum list
   size was necessary (eg if chains are being coupled by non-effector rules,
   or if there are multiple level default hierarchies involved.)  For some
   problems, though, substantially longer lists (eg >20) seem to cause
   problems, I think because they allow "parasites" and overly general rules
   to post messages and survive.

   I can believe the ML size should be related to classifier list size. 
   A couple of related issues: 
   a. Do you allow duplicate rules? 
      If not, this is not too expensive to control 
      (at rule creation time). 
      Note that (many) classifier systems use a GA that is like the 
      "steady state" GA that people have been using in GA work lately, 
      and (I think) those people are finding that not allowing duplicate 
      rules is helpful. 
   b. Do you allow duplicate messages? 
      This can be more expensive to control, and more importantly 
      (to my mind), it raises questions about how to allocate credit. 
      That is, if you are using the Holland-cfsc1-type 
      bucket brigade, where specific rules pay specific other rules, 
      and a rule wins but its trying to post a message already on the 
      new message list, what do you do?  Just treat it like it lost? 
      Somehow that doesn't seem right. 
   I must confess its not clear to me what is best to do about 
   duplicate rules and messages. 

     - Rick Riolo

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

From: willamu!levenick@uunet.UU.NET  (Jim Levenick)
Date: Tue, 26 Nov 91 10:25:38 PST
Subject: Re: negative rewards and optimal ML length in LCS

   You pose two questions:
   >   Is there anyone that used punishments (negative rewards)?

   I've done some of this (although unpublished).  Yes, it does
   make the learning go faster (I was starting with all possible
   rules in a very simple situation and trying to select just
   an optimal subset of the rules; there were just two types of
   objects, good and bad and the rover was supposed to learn to
   get the good ones - it worked very well but seemed too trivial
   to try to publish -- how could the results have been different?).
   I also tried to get it to "pick blackberries", where the system
   had to learn to endure slight punishments to gain rewards; a blackberry
   bush looked like
   ---
   -$-
   ---
   The problem was to retain rules that approached -'s (even though
   -'s got a negative reward) so as to get the $'s which gave a large
   reward.  I again started with all possible rules and wanted the
   classifier system to select and retain the correct set.  Unfortunately
   I couldn't develop a robust scheme, even with much fiddling;
   in testing over many parameter settings and initial situations
   the correct set was seleted and retained roughly half the time.
   Thus, I gave up in disgust (no, really other interests took
   precedence; I've always been skeptical of the robustness of
   classifier systems and my experiments tended to support this 
   bias).

   >   2) It seems to me that nobody cares about message list optimal
   length.

   I'm not aware of any results on this.  From my perspective it is a 
   difficult and poorly defined problem, symptomatic of the state of the
   field.  Since there is no base of theory it is difficult to derive
   results wrt things like optimal message list length.  I'd guess
   the answer is a function of the particular application
   and the particular inplementation
   of the bucket brigade.  This would be a useful area for someone to
   do a careful analysis and attempt to establish definitions and methods
   to compare different systems; but it is not my cup of tea (as the
   English would say).

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

From: Mark.Kantrowitz@GLINDA.OZ.CS.CMU.EDU
Date: Mon, 25 Nov 91 11:06:26 EST
Subject: NETtalk task using GAs?

   Has anybody implemented the NETtalk task using GAs? NETtalk was a 
   connectionist system for producing pronunciations for English words.
   An implementation of NETtalk using memory-based reasoning was used to
   compare memory-based reasoning with neural nets. The NETtalk task
   seems like a reasonably fair way of comparing different machine
   learning algorithms, which is why I'm interested in it.

   Please send replies to mkant@cs.cmu.edu, as I don't read this list.

   Thank you.

   Mark Kantrowitz
   <mkant+@cs.cmu.edu>

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

From: SNYDERR@randd.abbott.com
Date: Mon, 25 Nov 1991 12:51 CDT
Subject: Genetic Algorithm Program in Smalltalk

   Does anyone have the source for a GA program written in Smalltalk?  How about
   a GA written in an object-oriented language like C++ that I might try
   converting over to Smalltalk?  Thanks in advance.

   Robert W. Snyder
   <snyderr@randb.abbott.com>

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

From: <@ua1ix.ua.edu:rob@galab2.mh.ua.edu>
Date: Mon, 25 Nov 91 20:38:59 CST
Subject: TCGA method of payment.

   TCGA paper requestors:
   When requesting a dissertations or theses from The Clearinghouse
   for Genetic Algorithms, please include 
   a check or money order for $9.00 per dissertation or thesis
   ($12.00 overseas) made payable to The University of Alabama
   to defray the costs of copying and shipping these large documents.

   The Clearinghouse for Genetic Algorithms
       Department of Engineering of Mechanics
       Room 210 Hardaway Hall
       The University of Alabama
       Box 870278
       Tuscaloosa, Alabama 35487
   <<email>> @ua1ix.ua.edu:rob@galab2.mh.ua.edu 
   <<phone>> (205) 348-1618
   <<fax>> (205) 348-8573    

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

From: Nick Radcliffe <njr@epcc.edinburgh.ac.uk>
Date: Tue, 26 Nov 91 16:26:29 GMT
Subject: New Technical Report Available

   New Technical Report Available through anonymous ftp.

	Genetic Set Recombination and its Application to
	      Neural Network Topology Optimisation

	Abstract
	--------
	Forma analysis is applied to the task of optimising the
	connectivity of a feed-forward neural network with a single
	layer of hidden units.   This problem is reformulated as a
	multiset optimisation problem and techniques are developed
	to allow principled genetic search over fixed- and variable-size
	sets and multisets.   These techniques require a further
	generalisation of the notion of gene, which is presented.
	The result is a non-redundant representation of the neural network
	topology optimisation problem together with recombination operators
	which have carefully designed and well-understood properties.
	The techniques developed have relevance to the application
	of genetic algorithms to constrained optimisation problems.

   This paper is now available in PostScript form through anonymous ftp
   at ftp.epcc.ed.ac.uk [129.215.56.29].   Some older papers are also
   available at this site.   The following details are relevant:

	Genetic Set Recombination and its Application to
	      Neural Network Topology Optimisation
	Nicholas J Radcliffe
	EPCC Technical Report EPCC-TR-91-21.
	Available as tr9121.ps.Z in pub/tr/91.

	Equivalence Class Analysis of Genetic Algorithms
	Nicholas J Radcliffe
	Complex Systems, 5:2 1991.
	Available as tr9003.ps.Z in pub/tr/90.

	Forma Analysis and Random Respectful Recombination
	Nicholas J Radcliffe
	Proc. Fourth Int'l Conf. on Genetic Algorithms (1991)
	Available as tr9102.ps.Z in pub/tr/91.

   The following sample session shows the use of the ftp site.

   % ftp ftp.epcc.ed.ac.uk
   Connected to crystal.epcc.ed.ac.uk.
   +=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=

		 Welcome to the Edinburgh Parallel Computing Centre


   For information on the material contained in this archive, contact
   L.Chantler@ed.ac.uk.

   For ftp access login as: anonymous
		  password: your email address

   +=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=
   220 crystal FTP server (Version 5.64 Tue Jun 18 12:54:32 BST 1991) ready.
   Name (ftp.epcc.ed.ac.uk:njr): anonymous
   331 Guest login ok, send ident as password.
   Password: <njr@epcc.ed.ac.uk>
   230 Guest login ok, access restrictions apply.
   ftp> cd pub/tr/91
   250 CWD command successful.
   ftp> get tr9121.ps.Z
   200 PORT command successful.
   150 Opening ASCII mode data connection for tr9121.ps.Z (xxxx bytes).
   226 Transfer complete.
   local: tr9121.ps.Z remote: tr9121.ps.Z
   104210 bytes received in xx seconds (xx Kbytes/s)
   ftp> ^D
   Goodbye.
   -------

   Nick Radcliffe			Edinburgh Parallel Computing Centre
					King's Buildings
   email: njr@epcc.ed.ac.uk		University of Edinburgh
   fax:	+44 31 662 4712			EH9 3JZ
   phone: +44 31 650 5020		Scotland

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

From: Pierre Lecocq CRIF <M3730%eurokom.ie@pucc.PRINCETON.EDU>
Date: Tue, 26 Nov 1991 19:56 GMT
Subject: Call for Bin Packing Benchmarks

     I have written a GA for the Bin Packing Problem (unidimensional - as
   defi- ned in [1]) which seems to fare well on the randomly generated
   examples I feed it. It beats the First Fit Descendig ([1]) heuristic, but
   admittedly that is not a very tough competition.
     I would therefore like to make the following 'Call for Bin Packing
   Bench- marks': can anybody point out (a reference to) a means of comparing
   our GA to other (not necessarily GA) solutions of the problem
     Fitted with acyclic precedence constraints, the Bin Packing becomes the
   important industrial problem of Line Balancing ([2]). Again, our GA seems
   to perform well, but I even more lack a comparison, as the First Fit
   Descending breaks down with the additional constraints. So here too a
   benchmark would be welcome.

    Real world instances of BP and/or LB would be most appreciated.

     Thank you.

   [1] Garey Michael R. and Johnson David S. "Computers and Intractability -
   A Guide to the Theory of NP-completeness", W.H.Freeman Co., San Francisco,
   1979.

   [2] Falkenauer Emanuel and Delchambre Alain "A Genetic Algorithm for Bin
   Packing and Line Balancing", submitted to the IEEE 1992 Int'l Conference on
   Robotics and Automation (RA92), May 10-15, 1992, Nice, France.

   E.Falkenauer
   CRIF - Research Centre for the
     Belgian Metalworking Industry
   CP 106 - P4
   50, av. F.D.Roosevelt
   B-1050 Brussels
   Belgium
   Email: PIERRE_LECOCQ_CRIF@eurokom.ie

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