
Genetic Algorithms Digest   Wednesday, June 10 1992   Volume 6 : Issue 19

 - Send submissions to GA-List@AIC.NRL.NAVY.MIL
 - Send administrative requests to GA-List-Request@AIC.NRL.NAVY.MIL
 - anonymous ftp archive: FTP.AIC.NRL.NAVY.MIL (see v6n5 for details)

Today's Topics:
	- GAs compared to annealing
	- Re: Is Crossover Useful?
	- Re: Benefits of crossover
	- info on genetic algo. request
	- Papers on Routing and Scheduling Problems
	- Request for information on GA and discriminant functions.
	- SAB92 Reminder Notice
	- Biocomputation Workshop, new deadline for posters

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

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

 COGANN, Combinations of GAs and NNs, @ IJCNN-92 (v5n31)      Jun 6,     1992
 ARTIFICIAL LIFE III, Santa Fe, NM                            Jun 15-19, 1992
 Evolution as a computational process, Monterey (v6n9)        Jun 22-24, 1992
 ML-92, Machine Learning Conference, Aberdeen (v6n8)          Jul  1-3,  1992
 10th National Conference on AI, San Jose,                    Jul 12-17, 1992
 FOGA-92, Foundations of Genetic Algorithms, Colorado (v5n32) Jul 26-29, 1992
 COG SCI 92, Cognitive Science Conference, Indiana, (v5n39)   Jul 29-1,  1992
 ECAI 92, 10th European Conference on AI (v5n13)              Aug  3-7,  1992
 Parallel Problem Solving from Nature, Brussels, (v5n29)      Sep 28-30, 1992
 SAB92, From Animals to Animats, Honolulu (v6n6)              Dec  7-11, 1992

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

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

From: kcj@matilda.vut.edu.au (Kate Juliff)
Date: Thu, 21 May 92 10:20:50 EST
Subject: GAs compared to annealing

   Can anyone recommend any recent papers comparing GAs and simulated
   annealing, specifically in regard to complex problems?

   I am getting very poor results when I use annealing, whereas my GA gets
   almost perfect scores, on a packing/scheduling problem.

   Thanks in advance,

   Kate Juliff
   Victoria University
   P.O. BOX 14428,
   MELBOURNE MAIL CENTRE, AUSTRALIA 3000


[Editor's Note: The following issues of ga-list contained articles on GAs
and simulated annealing; v5n26, v5n36, v5n37, v5n40, v6n1, v6n2, and v6n6.
You might want to take a look at these.  As always, back issues are
available from ftp.aic.nrl.navy.mil.  -  Alan Schultz]

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

From: Pierre Lecocq CRIF <M3730@eurokom.ie>
Date: Mon, 25 May 1992 21:14 CET
Subject: Re: Is Crossover Useful?

       In "Cross-over vs Mutation" (GA-List v6n16), Mike Hobbs wrote:

  +          I have been using a steady-state, rank based genetic algorithm
  +  (loosely based on Whitley's Genitor) with a uniform cross-over and
  +  produced some pleasing results. However, on investigation I find that
  +  the results obtained with the cross-over operator disabled are only
  +  marginally worse than those using all the benefits of the GA paradigm.
  +
  +          Typical parameters are a population of 100, bitstring length
  +  80, 70/30 split between a crossover or mutation reproduction event
  +  (probability of mutation 0.06 per bit) and about 5000 individuals
  +  in a run.
  +
  +          I am quite prepared to believe that it is my implementation
  +  of the GA, the encoding of the problem, or some other mistake that can
  +  be laid at my door which can explain this apparent anomaly. But, if
  +  for the sake of argument, this is a real effect it implies that most
  +  of the real benefits of the GA, for this problem, can be obtained by
  +  the way in which strings are selected and removed, rather than the way
  +  in which they are generated.
  +
  +          Does this mean that the advantage gained by using a cross-over
  +  operator is an indicator of the effectiveness of the GA
  +  implementation, or simply that in some problems the advantages are
  +  small over a 'mutation only' reproduction technique?

  (end of citation of Mike Hobbs)


       I would like to make the following remarks:

  1.   The message gives a lot of details about the method used, but unfor-
  tunately fails to specify the most important thing, namely *the problem*
  to be solved.
       The point is, I think that there are problems for which the usual
  encoding/crossover combination cannot work well. Indeed, the power of a
  crossover lies with its capacity to *combine* parts of *each* of the two
  parents in the offspring, with the intention of getting together, in one
  individual, the 'good things' previously separated (of course, one can
  also get together the 'bad things', but the ensuing elimination handles
  that). However, in order to make the magic work, we must make sure that
  the 'things' above are meaningful, i.e. that the crossover indeed works
  with parts of solutions which actually can be combined.
       In practice, this means, on the one hand, that the encoding must use
  genes that map onto parts of solutions which, taken *isolated*, convey
  as much information as possible about the expected quality of the
  solution (i.e. chromosome) they are part of. Why do I insist on the word
  isolated? Because a crossover really combines *genes*, not whole
  solutions (it has been said so many times that a GA works with
  *representations* of solutions, rather than the solutions themselves -
  I just tried to say it in other words).
       On the other hand, the crossover must be adapted to the encoding
  *and* the problem. That is, it should combine the genes (i.e., as said
  before, *meaningful* parts of solutions) in such a way that they retain
  as much as possible of their meaning inside the new chromosome. Only then
  the 'good things' handed over from parents to the offspring stay *good*
  things.
       Also, the crossover must not be *too disruptive*. This obviously
  looks like a banality (Holland's 1975 seminal book already treated
  disruption, so this is really a 'deja vu'), but it is crucial to keep in
  mind that a crossover which is non-disruptive with respect to the
  *chromosomes* used can turn out to be catastrophically disruptive with
  respect to the *problem* being solved. That is, while a crossover can
  well transmit schemata from parents to offspring, if the schemata don't
  encode meaningful parts of solutions (here we are again) of the original
  problem, the transmitted schemata won't combine into anything useful.
       Another typical danger arises when the solutions must comply with
  some 'exotic' hard constraints. In that case, we are often tempted to
  'adapt' the genes from one parent to those in the other in order not to
  produce invalid individuals. This is o.k. as long as the adaptation
  leaves *most* of the incoming genes intact. Otherwise the information
  finally retained from one of the parents is so distorted that the
  'crossover' is in fact a mutation. Of course, one way out of this is to
  keep the invalid individuals (probably with some penalty in their
  fitness), but they will most probably be eliminated from the next
  generations, so it's like not having a crossover at all.

       In short, I think that the crossover is indeed a *very* powerful
  device in searching solution spaces (for me, it is The Method for
  exploiting regularities of search spaces - very few, if any, other
  methods have an equivalent), but these regularities must have a
  *structural counterpart* in the encoding, and they must be correctly
  combined by the crossover, i.e. the allimportant Thesis of Good Building
  Blocs must be complied with.
       Consequently, in situations like Mike Hobbs's, I would immediately
  check the encoding/crossover (he's probably right to blame the encoding):
  does it really avoid the caveats described above?
       Unfortunately, like most theses, the one above doesn't help much in
  choosing the right encoding/crossover for a particular problem. I am
  afraid we are (or at least I am) still nowhere near a system which would
  *automatically* adapt the encoding and the crossover to an arbitrary
  problem (to my opinion, that would be The Breakthrough in the GA
  paradigm, for I do not think that just using fixed-length strings with
  a 'tested' crossover will automatically yield a good method for an
  arbitrary combinatorial problem). Goldberg's Messy GAs constitute a very
  interesting  attempt to get just there, but I do not know whether it's
  conclusive yet.

       I have at least one example to illustrate the above difficulties.
  A very interesting class of problems which I think cannot be easily
  tackled by the usual encoding/crossover scheme are *grouping* problems,
  i.e. those where the aim is to group things together (find a good
  partition of a set). Maybe Mike Hobbs's is actually one of these.
       The straightforward encoding would map objects onto genes, yielding
  nice constant-length chomosomes, thus allowing any of the panoply of
  crossovers found in the literature. The advantage would be that the power
  of these crossovers is (in most cases) supported by the theory, so it
  should work out well.
       Nevertheless, when I tried to apply the classic GA to these prob-
  lems, it didn't work well. The reason is, in fact, quite simple: the
  object-to-gene encoding does not seem appropriate for these problems (and
  *not* that a GA couldn't be applied here). This is because the meaningful
  piece of information (part of a solution) which has to be manipulated by
  the crossover is a group, not an object. Consequently, even the least
  disruptive crossover, when applied to chromosomes mapping *objects* onto
  genes, is despairingly disruptive with respect to the original problem.

       This conclusion resulted in use of an encoding mapping groups onto
  genes. That in turn required a crossover adapted to the encoding and the
  problem (handling variable length chromosomes and performing a certain
  amount of interchromosomal adaptation), quite different from the usual
  'cut and concatenate' operators ([1],[2],[3] - sorry for the self-refs).
  Unfortunately, this means that I do not so far have a theory equivalent
  to Holland's to support what I dared to call the Grouping GA, except for
  the fact that I believe it does comply with the Thesis above. Fortunately
  however, it does seem to work well in practice (incidentally, I always
  use steady-state, rank based GAs, so I think that shouldn't be Mike
  Hobbs's problem).


  2.   Maybe the problem is actually simpler than those mentioned above:
  although it is an 'old' topic (see for example the discussion in this
  list at the end of the last year), I do think the uniform crossover is
  much too disruptive with respect to higher order schemata. That is, I
  think UC is good when it is *not* clear what are the regularities of the
  landscape, but I prefer to change encodings, rather than hope the UC
  would 'discover' the meaningful parts of solutions by chance. Did Mike
  Hobbs try something less disruptive (say two-point circular) with
  inversion?


  3.   Does anybody disagree?


  Best regards,
  E.Falkenauer (E-mail: PIERRE_LECOCQ_CRIF@eurokom.ie)

  [1] Falkenauer E. "A Genetic Algorithm for Grouping", in Procs of 5th
  Int'l Symposium on Applied Stochastic Models and Data Analysis, April
  23-26, 1991, Granada, Spain, World Scientific Pub.

  [2] Falkenauer E., Delchambre A. "A Genetic Algorithm for Bin Packing and
  Line Balancing", in Procs of 1992 IEEE Int'l Conference on Robotics and
  Automation (RA92), May 10-15, 1992, Nice, France.

  [3] Falkenauer E. "A Genetic Algorithm for Conceptual Clustering", Report
  FMS39, CRIF Industrial Automation, Brussels, Belgium. December 1991
  (available on request).

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

From: erwin@trwacs.fp.trw.com (Harry Erwin)
Date: Sat, 30 May 92 15:28:01 EDT
Subject: Re: Benefits of crossover

   A question that frequently comes up is whether crossover accelerates
   evolution in a genetic algorithm. An excellent book that addresses this
   and related issues is Evolutionary Genetics, by John Maynard Smith, Oxford
   University Press, 1989. It is slightly out of date in its treatment of
   repetitive DNA (which now seems to be instructions for expression of the
   portions of the genome coding for protein), but otherwise covers the
   issues of interest to the GA community in some depth, and with many
   pointers into the technical literature. JMS has other books that provide
   even more detail for those readers who are interested in following up
   further, but Evolutionary Genetics is an excellent starting point.

   BTW an answer to the question is that in small populations, recombination
   may well slow evolution since it breaks up favorable gene combinations
   when they occur. In large populations, recombination avoids Muller's
   ratchet and allows the merger of advantageous mutations earlier (i.e.,
   before they have become fixed).  So if you use recombination, make sure
   your population is large enough that it can take advantage of the
   acceleration.

   Cheers,
    Harry Erwin    erwin@trwacs.fp.trw.com

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

From: cacou@leland.stanford.edu
Date: Wed, 13 May 92 23:34:13 PDT
Subject: info on genetic algo. request

  I would like to have some information on the problem of adapting a neural
  network weight vector through genetic algorithm.

  Thank you very much in advance,
  Christophe Hurbin.

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

From: "Sam R. Thangiah" <SRT%SRU.bitnet@CUNYVM.CUNY.EDU>
Date: Sat, 30 May 92 11:51:43 EDT
Subject: Papers on Routing and Scheduling Problems

   The following papers on routing and scheduling problems are available from
   the conference proceedings:

   Thangiah, Sam R. and Kendall E. Nygard (1992),"MICAH: A Genetic Algorithm
   System for Multi-Commodity Transshipment Problems," Proceedings of the
   Eight IEEE Conference on Artificial Intelligence for Applications,
   Monterrey, California, 240-246.

   Thangiah, Sam R. and Kendall E. Nygard (1992), "School Bus Routing using
   Genetic Algorithms," Proceedings of the SPIE Conference on the
   Applications of Artificial Intelligence X: Knowledge Bases Systems,
   Orlando, Florida, 387-398.

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

From: Aaron Konstam <AKONSTAM@VM1.TUCC.TRINITY.EDU>
Date: Sun, 31 May 92 21:00:13 CDT
Subject: Request for information on GA and discriminant functions.

   Is any one aware of anyone who has done work in using Genetic algorithms
   for finding the optimum coeficients for discriminant functions (especially
   linear discriminant functions)?

   I would also be interested if any one has used genertic algorithms in
   conjunction with logistic regression.

   Post responses to the ga-list or directly to me at the address below.

   AARON KONSTAM
   Trinity University
   715 Stadium Dr.
   SAN ANTONIO, TX 78212
   (512)-736-7484
   AKONSTAM@TRINITY.EDU

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

From: Stewart Wilson <wilson@smith.rowland.org>
Date: Mon, 08 Jun 92 14:06:49 EDT
Subject: SAB92 Reminder Notice


  * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * 
                                                                            
  *                          R E M I N D E R                              *

  *                      From Animals to Animats:                         * 
        2nd International Conference on Simulation of Adaptive Behavior    
  *                       ----------------------                          *
		   Honolulu, Hawaii, December 7-ll, 1992                   
  *                                                                       *
									  
  *       OBJECT: To bring together researchers in ethology,              *
          psychology, ecology, cybernetics, artificial intelligence,      
  *       robotics, and related fields to further understanding of        *
          the behaviors and underlying mechanisms that allow animals     
  *       and, potentially, robots to adapt and survive in uncertain      *
          environments.
  *                                                                       *
	  DEADLINE -- Submissions must be received by the organizers       
  *                   by JULY 15, 1992                                    *
									  
  *       To receive the full Conference Announcement and Call for        *
	  Papers, please contact one of the Organizers:                    
  *                                                                       *
	     Jean-Arcady Meyer  meyer@wotan.ens.fr                        
  *                             meyer@frulm63.bitnet                      *

  *          Herbert Roitblat   roitblat@uhunix.uhcc.hawaii.edu           *
				roitblat@uhunix.bitnet                    
  *                                                                       *
	     Stewart Wilson     wilson@smith.rowland.org                  
  *                                                                       *
									  
  * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

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

From: rsjuds@snll-arpagw.llnl.gov (judson richard s)
Date: Fri, 5 Jun 92 12:55:02 -0700
Subject: Biocomputation Workshop, new deadline for posters

   ************************************
   *  NOTE DEADLINE CHANGE !!!!       *
   *                                  *
   *  NEW POSTER DEADLINE: JUNE 12    *
   *                                  *
   *  SCHEDULE ATTACHED               *
   ************************************

   BIOCOMPUTATION WORKSHOP
   Evolution as a computational process
   June 22-24, 1992
   Doubletree Hotel
   2 Portola Plaza
   Monterey, Ca 93940

   Sponsored by the Institute for Scientific Computing Research at LLNL and
   the Center for Computational Engineering at SNL

   This workshop brings together biologists, physicists and computer
   scientists with interests in the study of evolution.  The premisis
   of the workshop is that natural evolution is computational process
   of adaptation to an ever changing environment.  Mathematical theory and 
   computer modeling are therefore ideally suited to study evolution
   and conversely, evolution may be used as a model system to study the 
   computational processes of optimization and emergent pattern formation.

   Fifteen invited speakers will provide general reviews and summaries of their
   recent research.  Although oral presentations will be limited to the invited
   speakers, original research contributions are solicited for poster sessions 
   in the following areas:

   natural evolution
   artificial life
   genetic algorithms and optimization

   List of speakers:
   ----------------
   Stuart Kauffman      --- University of Pennsylvania, Santa Fe Institute
   Alan Templeton       --- Washington University, St. Louis
   Daniel Hillis        --- Thinking Machines Inc.
   Richard Hudson       --- University of California, Irvine
   Steven Frank         --- University of California, Irvine
   Alan Hastings        --- University of California, Davis
   Warren Ewens         --- Melbourne University and University of Philadelphia
   Marcus Feldman       --- Stanford University
   Lee Altenberg        --- Duke University
   Aviv Bergman         --- SRI  and Stanford University
   Mark Bedau           --- Reed College
   Heinz Muehlenbein    --- University of Bonn
   Eric Mjolsness       --- Yale
   Wolfgang Banzhaf     --- Mitsubishi Electric Corp.

   Schedule
   ---------
   Sunday - June 21    
   6:00 pm - 8:00 pm     RECEPTION

   Monday - June 22

   10:00 am - 10:45 am    Kaufmann    Evolution and co-evolution 
				      at the edge of chaos
   10:45 am - 11:30 am    Hastings    Use of optimization techniques to
				      study multilocus population genetics
   11:45 am - 12:30 pm    Altenberg   Theory on the evolution and
				      complexity of the genotype-phenotype
				      map
   12:30 pm - 2:00 pm     LUNCH

   2:00 pm  - 2:45 pm     Templeton   Gene tree overlay algorithms: a 
				      powerful methodology for studying
				      evolution
   2:45 pm  - 3:30 pm     Bedau       Measuring evolutionary activity
				      in noisy artificial systems
   4:00 pm  - 4:45 pm     Muehlenbein Evolutionary algorithms as a research
				      tool for a new evolutionary theory
   7:00 pm                BANQUET - MONTEREY AQUARIUM
			  After Dinner Speaker - Warren Ewens

   Tuesday - June 23

   10:00 am - 10:45 am    Banzhaf     An introduction to RNA-like 
				      algorithms with applications to
				      sequences of binary numbers
   10:45 am - 11:30 am    Hillis      Simulated evolution and the Red Queen
				      Hypothesis
   11:45 am - 12:30 pm    Frank       Coevolutionary genetics of hosts
				      and parasites
   12:30 pm - 2:00 pm     LUNCH

   2:00 pm  - 2:45 pm     Bergman     Learning your environment

   2:45 pm  - 3:30 pm     Feldman     Recombination and evolution

   4:00 pm  - 7:30 pm     POSTER SESSION

   Wednesday - June 24

   10:00 am - 10:45 am    Mjolsness   Connectionist grammars in evolution 
				      and development
   10:45 am - 11:30 am    Hudson      Title to be announced

   12:00 pm - 2:00 pm     GROUP LUNCH DISCUSSIONS AT LOCAL RESTAURANTS 

   Instructions for Submissions and registration:

   Authors should submit a single-page abstract clearly stating their results
   by June 12, 1992, to the Meeting Coordinator at the address listed below.
   Please indicate which of the above categories best applies to your paper.
   There will be no parallel sessions, and the workshop will be structured to
   stimulate and facilitate the active involvement of all attendees.  There
   will be sessions on the first 2 days from 9:00 AM till 5:00 PM with 1-2
   hrs lunch breaks. On the third day there will be a morning session and a
   short afternoon session only (maybe one talk until 3:00 PM).  Registration
   fees are $100 for full-time Ph.D. students and $250 for all others.  Fees
   include admission to a banquet, at the Monterey aquarium, to be held on
   Monday night.  (There is a $50 discount for students presenting posters at
   the meeting.)

   To obtain registration materials and housing
   information, please contact the Meeting Coordinator.

   For information only please contact eeckman@mozart.llnl.gov
   Electronic abstract submissions only at jb@s1.gov

   Meeting coordinator:  
   -------------------
   Chris Ghinazzi
   P.O. Box 808, L-426
   Lawrence Livermore Laboratory
   Livermore, CA 94550

   phone: (510) 422-7132

   email: ghinazzi@verdi.llnl.gov

   Please complete this form and return it to:  Evolution as a Computational 
   Process, c/o Chris Ghinazzi, Lawrence Livermore National Laboratory,
   P.O. Box 808, L-426, Livermore, CA 94550-9900.  Phone (510)422-7132 or
   FAX: (510)422-7819

   REGISTRATION FORM

   Name:
   Title:
   Organization:
   Address:
   City:
   State:
   Zip:
   Country:
   Citizenship:
   Telephone:
   email address:

   Registration Fees:
	   Regular 	 $250
	   Student		 $100
	   Student w/poster  $50

   Are you submitting a poster?	yes	no

   Total Payment Enclosed $________
	   Check or Money Order (payable in US dollars to UC Regents)

   Requests for refunds must be received in writing no later than June 1, 1992.
   Attendance is on a first-pay, first-serve basis.

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