Genetic Algorithms Digest    Friday, 4 March 1988    Volume 2 : Issue 5

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

Today's Topics:
	- Information for newcomers
	- Holland's work in cellular automata
	- Neural nets and GA's

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

Date:         Fri, 26 Feb 88 13:25:12 CST
From: Dave Goldberg <DGOLDBER@UA1VM.BITNET>
Subject: Information for Newcomers and a Concern

The last distribution contained several pleas from GA newcomers who were
seeking basic information on GAs.  I still have a number of copies of my 1986
GA Bibliography available as well as some reprints that may useful.  I might
also mention that I have a new book in press, GENETIC ALGORITHMS IN SEARCH,
OPTIMIZATION, AND MACHINE LEARNING.  Addison-Wesley hopes to have it bound by
AAAI.  Last, I was somewhat disturbed by Goldthwaite's comments decrying the
engineering orientation of Holland's students.  As Lashon Booker pointed out,
there are many examples where that has not been the case, but pointing this
out sidesteps the underlying issue.  One beauty of the genetic algorithm
movement has been that so many different types of individuals have been able
to make important contributions.  Mathematicians, computer scientists, social
scientists (even psychologists), economists, engineers, management scientists,
and so on have all contributed in important ways.  This is largely the result
of John Holland's insistence upon interdisciplinary effort, and I hope we are
not about to throw away this important characteristic of our heritage to engage
in unproductive, internecine bickering and rivalry.  Interdisciplinary research
is no easy task.  We genetic algorithmists, of all people, should appreciate
the inherent good in maintaining useful diversity amongst our number.

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

Date: Sun, 21 Feb 1988 17:03:30 LCL
From: Kislaya Prasad <PRASAD@SUVM.ACS.SYR.EDU>
Subject: AIList V6 (Genetic Algorithms)

	[ from AI-List Digest -- JJG]

While only superficially familiar with the current Genetic Algorithms
literature I remember reading a paper by Holland a few years back which
I found very interesting:

Holland, J. (1970) "Logical Theory of Adaptive Systems", in A.W. Burks (Ed.)
Essays in Cellular Automata, University of Illinois Press.

(Remember Cellular Automata?)
My question to those familiar with the literature is:
How does this work of Holland relate to the recent literature? I don't see it
referred to at all. Is there a strong reason for this, or just that it is
now old stuff? It seems to me that this paper had many of the fundamental
ideas of today (especially with respect to parallel implementations).

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

Date: Thu, 3 Mar 88 16:51:40 est
From: ackley@flash.bellcore.com (David Ackley)
Subject: A neural network genetic algorithm: GA(1) or GA(2)?

Rik Belew offers the following distinction (GA-List V2N4):
 	GA(1) is 			   GA(2) is
 "a particular (class of) 	"a broader class ... of genetic 
  algorithms developed by	 algorithms (sometimes also called
  John Holland and his 		 'simulated evolution') that bear
  students.  This GA(1) has	 some loose resemblance to population
  as its most distinctive	 genetics... Generally, these
  feature the 'cross-over'	 algorithms make use of only a
  operator."		   	 'mutation' operator."

and then makes the following point:
> The complication comes with work like Ackley's thesis (CMU, 1987)
> which refers to Holland's GA(1), but which is most accurately
> described as a GA(2).

I'm a bit disappointed --- if not entirely surprised --- that Rik
takes the trouble explicitly to exclude my dissertation work ("A
connectionist machine for genetic hillclimbing", Kluwer Academic
Publishers, 1987, 0-89838-236-X) from his class GA(1).  I think
Holland's work is great, and I believe in the fundamental idea of
implicitly-parallel schema-processing via a population, emphasis, and
recombination.

Looking more closely at Rik's distinction, I think it turns out either
that membership in GA(1) is restricted to a small and somewhat quirky
"DNA-ish" subset of all possible combination rules, or that a model
such as mine must be allowed to sit around the "crossover campfire",
odd though its crossovers may appear to be.

[If we disregard the parenthetical "(class of)" in the description of
 GA(1), then Rik is obviously right, since I was not a student of John
 Holland, and the neural network algorithm (called SIGH, for
 "Stochastic Iterated Genetic Hillclimbing") that I described in my
 dissertation is not --- in its particulars --- one that was explored
 by John Holland or his students.  I made it up myself; its
 particulars are densely-connected non-linear processing elements, not
 boolean operations on bit strings.] 

GA(1) only gets interesting when we include the parenthetical remark,
and admit that while the specific algorithms produced by Holland&Co.
are important as examples of implicitly parallel algorithms, there is
no reason to expect that their form should be prescriptive of the
class as a whole.  That is a good thing, too, if we are interested in
understanding the mind and the brain.  Although cut-and-swap crossover
is a simple, elegant combination rule for DNA, it is not ideally
suited to neural networks.

As one example, the location of the cut-point(s) is global information
which would have to be broadcast to all of the neural units
representing the offspring, so that each would know which parent's
value to copy.  By contrast, using SIGH's related-subpopulation-voting
combination rule, only local information is needed.  In SIGH,
recombination is a fully parallel operation.

Yes, SIGH's combination rule looks weird.  Yes, there may be more than
two parents, perhaps many more --- but selective apathy helps avoid
the "homogenization" problem that afflicts mandatory voting schemes.
Yes, SIGH is hard to analyze, because recombination emerges from the
voting scheme rather than being a hardwired, primitive operation.  You
may say that recombination in SIGH is not "cross-over", but it is
definitely something more than "only a mutation operator".

 David Ackley
 Bell Communications Research 2B-324
 435 South St
 Morristown, NJ  07960-1961

 ackley@bellcore.com

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

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


