
Genetic Algorithms Digest   Thursday, October 31 1991   Volume 5 : Issue 33

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

Today's Topics:
	- dialogue on uniform crossover
	- AISB - membership information

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

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

 First European Conference on Artificial Life (v5n10)         Dec 11-13, 1991
 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: pl160988@mtecv2.mty.itesm.mx (Ivan Ordonez-Reinoso)
Date: Mon, 7 Oct 91 13:57:05 CST
Subject: dialogue on uniform crossover

  Christer Ericson and I were having a discussion concerning Uniform
  Crossover.  We thought it would be a good idea to let other people
  participate on the discussion, so here it goes:

  Christer Ericson:
  ... my purely hypothetical suggestion was that a 0.04-uniform crossover
  should perform quite well.

  Ivan Ordonez-Reinoso:
  ... I have a copy of the ICGA-91 proceedings, and I read the Spears-DeJong
  Paper on Uniform Crossover. The conclusion seems to be that UC with a low
  swap probability is very little disruptive, and thus in certain cases may
  perform better than 2 point crossover, which is the least disruptive of
  all the traditional crossover operators.

  However, non-disruptive operators have the disadvantage that
  they perform very little exploration, and over-emphasize the importance
  of exploitation.. this can result in premature convergence in many
  cases. My personal opinion is that claims like "8 point crossover is
  better than uniform crossover" or "UC is better than 2 point crossover"
  cannot be made in general, because the performance of each particular
  operator depends on the characteristics of the objective function. I
  think that even the optimal allocation of parameters like population
  size and mutation probability are problem dependent..

  CE:
  The reason I thought about this, which I mentioned in my post, was that I
  thought it was pretty silly to distinguish between different crossover
  operators, and thereby handling crossover in a special way, compared
  to mutation which has only one parameter, namely its probability. I mean,
  the logical thing would be to handle these two operators in the same way
  and therefore either a) always use uniform crossover with some probability
  or b) have many different mutation operators (like say, n-bit mutation with
  some probability). I myself think that a) must be the better option simply
  because it keeps the parameters to a minimum (one per operator) thus making
  it simpler to draw conclusions about their performance, and it also
  couples the number of crossovers to the chromosome length, which seems like
  a reasonable thing to do. After posting that, I made a few minor experiments
  and found that the 0.04-uniform crossover I suggested, performed as well as,
  if not better than both 0.5-UC, and 1-point crossover (2-point crossover as
  well I think, but I can't remember for sure). Also, I used a steady-state
  GA so it might be the case that its use leads to premature convergence for
  a generational GA, but I had so such problems. But, you're obviously right in
  that the allocation of parameters is (probably highly) dependent upon
  the problem and the encoding used.

  IO:
  > I thought it was pretty silly to distinguish between different crossover
  > operators, and thereby handling crossover in a special way, compared
  > to mutation which has only one parameter, namely its probability...

  Here I have to disagree. Crossover should not be treated like mutation,
  because crossover is *the* genetic operator, while mutation is merely
  incidental, and many people think it's not even important at all!

  CE:
  Well, I'm of the opposite opinion, I believe that mutation plays a much
  greater role than people are willing to admit (or believe). And as far
  as I can tell I'm not the only one to believe that, in one of his ICGA89
  papers Schaffer made a few remarks on what the called naive evolution (NE),
  using only mutation as an operator about NE performing actually quite well.
  A year or so ago, I did an experiment with growing finite automatas, and
  my NE approach performed better than a traditional generational GA!
  Unfortunately, neither performed well enough though. :-(

  IO:
  Anyway, Im not quite sure that UC will always perform as well, even with
  low swap probabilities. I discussed about this with my thesis advisor,
  and he said he thinks that UC is not good precisely because
  it's uniform: no particular defining length schemas are favored, while
  in most problems low order schemas are the really important ones
  (remember the building block hypothesis). UC might behave fine when you
  try it on easy functions, like De Jong's test bed, but I think no
  experimental results exist for UC applied on harder functions.. I guess
  somebody should do that, maybe UC will work fine in any case!

  CE:
  Yes, that's certainly true that no one crossover operator will perform
  well on all problems, or rather there will always be a specialized crossover
  operator that performs better than a general operator on a specific problem.
  I believe however that UC will generally perform better than x-point
  crossover (for a given x). At least, it will be easier to find an UC that
  performs well than to find an x-point crossover simply because we only have
  to tweak one parameter (probability) instead of two (probability and what x
  to use) which will be a gain for some people who actually use GAs for
  real world problems (and as such need to get the GA working in a reasonable
  period of time, with a reasonable success rate).

  Well, time will tell. To conclude my opinion; I believe in steady-state GAs,
  uniform crossover, increased mutation probability and higher (ie non-binary)
  cardinality chromosomes (even non-linear structures as well). As you can see,
  I'm certainly no firm believer in the (strong) building block hypothesis! :-)

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

From: Judith Dennison <judithd@cogs.sussex.ac.uk>
Date: Tue, 8 Oct 91 14:30:13 +0100
Subject: AISB - membership information

        THE SOCIETY FOR THE STUDY OF ARTIFICIAL INTELLIGENCE
                AND SIMULATION OF BEHAVIOUR (AISB)

   AISB was established 26 years ago and is the U.K.'s original and foremost
   A.I. society with an international membership. It has over 900 members drawn
   from both industry and academia.

   For an annual membership fee of only 18.00 (pounds UK), 12.00 for students
   you will receive:

       * FREE ACCESS TO UP-TO-DATE INFORMATION IN THE FIELD THROUGH:

	 -  The AISB Quarterly Newsletter
	 -  The European Journal on Artificial Intelligence (AI Communications)
	 -  The AISB Electronic Mail Directory

       * SUBSTANTIAL REDUCTIONS AT AISB EVENTS INCLUDING:

	   -  A Biennial Conference - the next will be AISB93 in Birmingham.
	   -  Sponsored Conferences and Workshops
	       (e.g. AI and Education 1993, Edinburgh)
	   -  Schools and Tutorials
	       (e.g. AISB Post-Graduate Workshop to be held in January 1992)

   Other activities of the society of benefit to members include:

       * TRAVEL AWARD SCHEME for young researchers in the field
       * Access to the ECCAI TRAVEL AWARD SCHEME
       * The opportunity to receive new books in return for an 800 word
	   review to be published in the AISB Quarterly

       * Reduced rates to journals
	   'ARTIFICIAL INTELLIGENCE' (Elsevier Science) and
	   'ARTIFICIAL INTELLIGENCE REVIEW' (Kluwer Academic Publ).

   THESE POINTS ARE COVERED IN MORE DETAIL BELOW.

   Further information and sample copies of the 'AISB Quarterly' and
   'AI Communications' are available from Judith Dennison.

   YOU CAN JOIN NOW BY PRINTING OUT AND POSTING THE APPLICATION FORM BELOW

   ==================================================================

			   APPLICATION FORM
   Return to:

       Judith Dennison                     Email: judithd@cogs.sussex.ac.uk
       AISB Administration                 Tel:   +44 (0) 273 678379
       Cognitive and Computing Sciences    Fax:   +44 (0) 273 678188
       University of Sussex
       BRIGHTON BN1 9QH, UK

   NAME_____________________________________________________________

   ADDRESS_________________________________________________________

   _________________________________________________________________

   EMAIL ADDRESS___________________________________________________
   (please note this will be added to the AISB Email Directory unless
   otherwise requested)

   Annual Membership fees are as follows:

		   Regular    Student          (all figures are in pounds UK)

   UK              18.00      12.00           Students should sent proof of
   EUROPE          20.00      14.00           their full-time student status
   ELSEWHERE       25.00      18.00           e.g. copy of registration card

   Cheques should be in sterling made payable to AISB.

   Membership starts from the month of joining unless otherwise requested.

   Members wishing to pay by standing order can obtain details from
   Judith Dennison.

   ====================================================================

   FURTHER INFORMATION

   THE AISB QUARTERLY NEWSLETTER

   This is sent to all members and contains up to date news, conference
   announcements and reports, technical features and articles, book reviews,
   advertising material including information on courses, jobs and
   new publications.

   ECCAI - The European Co-ordinating Committee on Artificial Intelligence

   AISB has strong links with its European counterparts in particular
   through being a member group of ECCAI. AISB distributes
   'AI Communications' the official journal of ECCAI. 'AI Communications'
   covers the whole European AI Community and publishes refereed articles
   concerning scientific and technical AI procedures, as well as as opinions,
   policies and news.  All European AISB members receive this free of charge.
   ECCAI has also introduced a travel award scheme which AISB members are
   eligible for.

   THE AISB ELECTRONIC MAIL DIRECTORY

   AISB has published an EMAIL Directory updated at intervals.  This is
   available in booklet/on-line form to members and contains the electronic
   mail addresses of those members wishing to take part in it. Members
   are sent updated material regarding conferences, jobs, and general
   information regarding the society.

   AISB CONFERENCES

   AISB holds a bi-ennial conference which offers substantially reduced
   rates to members - reductions are significantly bigger than the cost
   of joining AISB! The next conference - the 9th in the series - will
   be held at the University of Birmingham in 1993. The conference is
   usually preceded by tutorials.

   AISB also sponsors other conferences and workshops in the field.
   Recent meetings include Turing 90, Interact 90, Computers and Writing 91.
   Future meetings include Psychology of Programming Workshop 92.

   POSTGRADUATE WORKSHOPS

   The AISB Postgraduate Workshops have proved very rewarding to students
   in the past - the next one is in January 1992 at the University of
   Nottingham.  They provide a forum for current PG students to meet and
   describe their work as well as hosting sessions on AI methodology and
   surviving the Ph.D. process.

   TRAVEL AWARD SCHEME

   AISB awards a limited number of travel grants to British, European or
   International conferences or workshops on AI.  The grants are aimed at
   young or new researchers in the field. Apply early - these are very
   popular!

   AISB also assists in the formation of local and regional AI Groups
   by providing small grants to enable groups to organise and to pay for
   external speakers.

	      JOIN NOW BY USING THE MEMBERSHIP FORM ABOVE

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