
Genetic Algorithms Digest    Monday, 21 May 1990    Volume 4 : Issue 8

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

Today's Topics:
	- New Moderator
	- Reports from recent workshops?
	- Boltzmann Tournament Selection and Adaptive Default Hierarchies
	- Description of some current work at NRL
	- Forthcoming paper

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

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

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
Workshop Parallel Prob Solving from Nature, W Germany (v4n5)  Oct 1-3,   1990
2nd Intl Conf on Tools for AI, Washington, DC (v4n6)          Nov 6-9,   1990

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

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

Date: Mon, 21 May 90 10:12:35
From: GA-List-Request@aic.nrl.navy.mil (Alan C. Schultz)
Subject: New Moderator

     My name is Alan C. Schultz, and I am the new moderator of ga-list.
John has done an excellent job, but wishes to spend more time on research
and less on administrivia.  Hopefully, he will now have more time to spend
in discussion on this list!

     A little background on myself:  I am a researcher at the Navy Center
for Applied Research in Artificial Intelligence (part of the Naval
Research Lab), and I am a member of the machine learning group headed by
John.  My main areas of interest include using existing heuristic 
knowledge to seed GAs for faster (better?) solutions, using high level
representation languages with GAs, and using GAs to learn under simulation.

- Alan

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

Date: Fri, 18 May 90 9:15:16
From: GA-List@aic.nrl.navy.mil (Alan C. Schultz)
Subject: reports from recent workshops?

Any volunteers wish to report on the Workshop on GAs, Sim. Anneal.,
and Neural Nets that was help in Glasgow, or the Artificial Life
Workshop at Los Alamos?

- Alan

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

Date: Fri, 04 May 90 09:36:15 CDT
From: Dave Goldberg <DGOLDBER@UA1VM.ua.edu>
Subject: Boltzmann Tournament Selection and Adaptive Default Hierarchies

A technical report entitled "A Note on Boltzmann Tournament Selection
for Genetic Algorithms and Population-oriented Simulated Annealing"
(TCGA Report No. 90003) describes a new form of tournament selection
that achieves a Boltzmann distribution across a population of structures
in a genetic algorithm or in a parallel simulated annealer.  The
mechanism requires logistic acceptance and anti-acceptance competitions
for its operation, and proof-of-principle tests and fixed-point theory
show that the mechanism can achieve stable Boltzmann-distributed
populations.  The report contains sample Pascal code and suggests the
use of the procedure in parallel and as another form of niching.

A paper presented at the Arizona conference on AI, Simulation and
Planning in High Autonomy Systems describes a mechanism for achieving
accurate default hierarchy formation without reliance on
specificity-based bidding structures.  The paper "Reinforcement Learning
with Classifier Systems" by R. E. Smith and myself describes the
calculation of a separate priority measure within the classifier that
permits formation (as opposed to discovery) of multi-level DHs that
separate adaptively, thereby permitting near-perfect performance in
the presence of noise, bad rules, or other disturbances.

Copies of these papers are available upon request.   Snail-mail addresses
please.

Dave Goldberg
dgoldber@ua1vm.ua.edu or dgoldber@ua1vm.bitnet

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

Date: Thu, 17 May 90 14:56:46 EDT
From: spears@AIC.NRL.Navy.Mil
Subject: Description of some current work at NRL

	In order to generate more discussion on the GA-list, I
	will describe some of the work that I've been doing at
	the NRL AI Center (with Ken DeJong).

	1) I have done some work in the fixed length world, using GAs
	for solving boolean satisfiability problems (see 1989 GA Conf
	proceedings for details). I also have a student who is looking
	at other NP-Complete problems, including the max-cut and
	independent set graph problems. We feel that there is a whole
	host of NP-Complete problems that have natural GA representations
	(unlike TSP). In the case of SAT, I have also compared this to
	a NN algorithm for solving SAT problems, with the outcome that
	GAs appear better for harder, more complex problems.

	2) We have recently started work in the variable length world,
	by creating a GA concept learner. Our approach is similar to
	LS-1 (ie., a Pittsburgh approach). The variable length world
	raises several issues, such as hyperplane analysis, the right
	way to do crossover, taxation schemes, etc. Although we don't
	have the answers to any of this yet, we are slowly progressing.
	Oh yes, for GA fans, the GA concept learner is capable of good
	performance (from a predictive accuracy point of view, not from
	a time point of view (YET!)). We have taken on ID5R and C4.5
	on a few problems and have had very good results.

	3) We plan an in-depth revisitation of crossover, in an attempt
	to understand it better in both the fixed and variable length worlds.

	There are unfortunately many other topics we wish to explore, with
	so little time. I hope a few of these comments will spark questions,
	debates, and interest in some of these issues.

	Bill spears@aic.nrl.navy.mil

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

Date: Fri, 18 May 90 13:10:34 EDT
From: gref@AIC.NRL.Navy.Mil
Subject: Forthcoming paper

To appear in the Proceedings of the Seventh International
Conference on Machine Learning, June 1990.

	      Simulation-Assisted Learning by Competition:
		      Effects of Noise Differences
	     Between Training Model and Target Environment

  by Alan C.  Schultz, Connie Loggia Ramsey, and John J. Grefenstette

				Abstract

The problem of learning decision rules for sequential tasks is
addressed, focusing on the problem of learning tactical plans from a
simple flight simulator where a plane must avoid a missile.  The
learning method relies on the notion of competition and employs genetic
algorithms to search the space of decision policies.  Experiments are
presented that address issues arising from differences between the
simulation model on which learning occurs and the target environment on
which the decision rules are ultimately tested.  Specifically, either
the model or the target environment may contain noise.  These
experiments examine the effect of learning tactical plans without noise
and then testing the plans in a noisy environment, and the effect of
learning plans in a noisy simulator and then testing the plans in a
noise-free environment.  Empirical results show that, while best result
are obtained when the training model closely matches the target
environment, using a training environment that is more noisy than the
target environment is better than using using a training environment
that has less noise than the target environment.

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

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


