
Genetic Algorithms Digest    Friday, 29 Decemeber 1989    Volume 3 : Issue 22

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

Today's Topics:
	-  GAs for NNets
	-  GA Workshop?
	-  Fitness distribution
	-  Variable string length
	-  Pattern learning plan

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

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

IJCNN Session on Evolutionary Processes (v3n10)               Jan 15-19, 1990
Double Auction Tournament - Sante Fe Institute  (v3n12)       Mar 1990
Workshop on GAs, Sim. Anneal., Neural Nets - Glasgow (v3n15)  May 9, 1990
7th Intl. Conference on Machine Learning (submissions 2/1/90) Jun 21-23, 1990
Workshop Foundations of GAs (v3n19)                           Jul 15-18, 1990
Conference on Simulation of Adaptive Behavior, Paris (v3n21)  Sep 24-28, 1990

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

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

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

From: George Robbins <logcam!george@relay.EU.net>
Date: Thu, 14 Dec 89 09:09:16 GMT
Subject: GAs for NNets

	Hi, I am currently working on a system which designs neural net st
structure using GAs, and have a question for anyone more neurally aware than
myself. It has been suggested to me that a useful way to cut down the search
space of possible nets is to only allow nets which have some regular pattern
of connectivity between their layers, (oh, forgot to say we're principally
interested in nets with hidden layers). It has further been argued that doing
this will help to prevent the nets performance from fitting too closely to the
training set, and hence losing the ability to generalise usefully. 
   While I can see/imagine that regular patterns of connectivity may well be 
useful in image-processing type tasks, (where you're repeating some task
over a large area of input nodes) I can see no a priori reason why regular
connectivity patterns should be at all useful elsewhere. In fact i've been
noted the work of Todd,Miller, and Hegde with interest: they found that 
irregular connectivities peformed well on a number of tasks - e.g. the four
quadrant problem. 
   Work using nets whose connectivity contains regular patterns abounds -
perhaps through us imparting our own design biases to the nets we build.
My questions are these:
	1. does anyone out there have any experience of trying to evolve
	   regular/irregular connectivity patterns in NNs.

	2. does anyone know of any work done in NNs, (not necessarily evolved
	   ones) which used irregular, (randomised or otherwise) connectivities

	3. are there any informed opinions out there on whether regular
	   connectivity pattens are likely to be useful.

thanks for your time
	george@uk.co.logcam

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

Date: Mon, 18 Dec 89 16:34:52 EST
From: S Khuri <@JHMAIL.HCF.JHU.EDU:kttur_s@JHUNIX.BITNET>
Subject: GA Workshop?

--To whom it may concern,
        Could you please send me some information on the workshop
that will take place at Indiana University in July?
        Could you please include me on your mailing list too?
        Thanks.
--Sam

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

Date: Thu, 14 Dec 89 17:09:38 -0500
From: androula@lips1.ecn.purdue.edu (Ioannis Androulakis)
Subject: Fitness distribution

 Is there any way to identify any particular distribution
 that the performance measure, in GA search, follow, or
 that distribution should be viewed as a random function?
 In other words, do you think that, after the "initial 
 transient" has been overcome, we could argue that the
 performance measure will follow some sort of a "normal"
 distribution which becomes sharper as the overall
 convergence is approached?
 
 Thank you,
 Ioannis P. Androulakis

 e-mail : androula@lips1.ecn.purdue.edu

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

Date: Tue, 19 Dec 89 15:29:10 -0500
From: androula@lips.ecn.purdue.edu (Ioannis Androulakis)
Subject: Variable string length

   I am interested in any work that deals with
   variable string lengths. 
   Thank you in advance,
   ioannis

   androula@lips1.ecn.purdue.edu

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

Date: Wed, 20 Dec 89 16:26:02 -0500
From: androula@lips1.ecn.purdue.edu (Ioannis Androulakis)
Subject: Pattern learning plan

    I am interested in any information concerning the
    pattern learning plan. In other words how can I
    exploit the pattern of changes causing improvements
    from one generation to another.
    Thank you,
    Ioannis P. Androulakis
    androula@lips1.ecn.purude.edu

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

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

