Genetic Algorithms Digest    Thursday, 24 March 1988    Volume 2 : Issue 10

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

Today's Topics:
	- Classifier Systems
	- Parallel GAs
	- GAs for Image Registration
	- GA Research at Mitre Corp.

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

Date:       Mon, 21 Mar 88 19:20:03 GMT
From:       Gary Roberts <gary@EDAI.ED.AC.UK>
Subject:    Research on Classifier Systems

APPLICATION AREA: Discovery Learning in a Rich Simulated Environment

GENERAL APPROACH: I am experimenting with various modifications to the
standard classifier systems.

GA TOOL: Classifiers are produced via genetic algorithm.

RESULTS: I have attempted duplication of Wilson's "Animat" as reported in
"Knowledge Growth in an Artificial Animal", from the 1985 GA Conference
proceedings.  So far, I have not managed to match that performance level.
Note, however, that I have not yet implemented the distance metric
described.  It does not appear difficult.

PROBLEMS: None, at the moment.

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

Date: Tue, 22 Mar 88 09:31:36 EST
From: "Worthy N. Martin" <wnm@uvacs.cs.virginia.edu>
Subject: Parallel GAs

  I am actively working on a parallel version of GA
which Dana Richards, Jim Cohoon and I refer to as
Punctuated Equilibrium - GA. This model is being designed
with message passing multiprocessors in mind. Our results
on the ordered linear array (OLA) problem were presented
at the last GA conference. We are currently expanding the
model and applying it to several VLSI problems.

   Worthy Martin

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

Date:  Wed, 23 Mar 88 20:45:03 CST
From:  John Grefenstette <vuse!ip0!jjg@uunet.UU.NET>
Subject:  GAs for Image Registration

This note describes some work that Mike Fitzpatrick and I
have done at Vanderbilt University.

APPLICATION AREA:  Medical image registration.  The problem is
to compensate for patient/camera motion occurring between the
taking of two x-ray images, so that the earlier mask image can be
subtracted from the later one to produce a clear image of particular
arteries (into which dye has been injected in the meantime.)
Unavoidable motion by the patient/camera produces blurs,
or motion artifacts, that may require a repetition of the procedure.

GENERAL APPROACH:  We define a parameterized class of transformations
that may compensate for the motion, and use GAs to search for high
performance transformations.  The evaluation of a transformation
is based on the average gray-level pixel difference between the
target image and the transformed mask image.  Minimizing this
difference generally minimizes the motion artifacts, since the
only remaining difference comes from the (generally small) artery.

GA TOOL: Genesis, interfaced to a displaytool that allows
us to watch as the image improves.

RESULTS:  Preliminary results have been very encouraging, but we
are just beginning to experiment with real coronary artery images.

From the GA point of view, the most interesting results concern
the performance of GAs when the evaluation function is too
complex to compute exactly, but can be estimated by sampling.
In our problem, the computation of the exact difference between two
images requires involves thousands of pixel comparisons.  However,
we have found that GAs do extremely well with fairly crude estimates.
For example, we routinely estimate the "fitness" of a transformation
by sampling about 20 pixels, so that it becomes feasible to evaluate
several thousand candidate transformations in a few cpu minutes.
We have explored tradeoffs between the number of samples per eval and
the population size, as well as between the samples per trial and
the number of generations that can be run in a fixed amount of time.

For more details, see:

Fitzpatrick, J. M., D. R. Pickens, J. J. Grefenstette, R. R. Price
and A. E. James, (1987).  A technique for automatic motion correction
in DSA. Optical Engineering 26(11), (pp. 1085-1093).

Grefenstette J. J. and J. M. Fitzpatrick, (1985). Genetic search
with approximate function evaluation. Proceedings of the First
International Conference Genetic Algorithms and Their Applications
(pp. 112-120). Pittsburgh, PA: Lawrence Erlbaum Assoc.

Grefenstette J. J. and J. M. Fitzpatrick, (1987).
Genetic algorithms using Monte Carlo function evaluations.
Submitted to Machine Learning Journal.

PROBLEMS:  We are in the process of doing an extensive experimental
comparison of a number of search techniques in this domain, including
GAs, simulated annealing, iterated hill-climbing, and specialized
methods. 

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

Date: Wed, 23 Mar 88 12:14:58 EST
From: jima@mitre.arpa
Subject: GA Research at Mitre Corp.

John,

  Good to hear from you! Yes, I definitely want to be on the GA list.
Please also add Karl Keller, keller@mitre.arpa, if he's not on it already.

  OK. Our application area is tactical intelligence (for now, anyway),
although we've put a lot of effort into building a simulation environment
that should be appropriate to other domains. The problem as we've defined
it is one of unsupervised learning by predicting an adversary's actions.

  The idea is to implement the bucket-brigade and similar approaches in
a reasonably standard production system to demonstrate its appropriateness for
a large class of AI programs. This year we are developing a formal analysis
and empirical evidence of the behavior of these algorithms for credit
assignment, with particular interest in absolute measures of utility in
complex inference networks (the usual approach leading to utility measures
that depend in part on the ply of the rules in a rule chain).
  We hope to buffer all hypothesized new structural relations in the domain
throughg a hypothetical worlds facility (which is making the tracking of rule
wealths pretty interesting). The only new relations we are presently worried
about are new rules. Part of our work in rule generation has been to make
sense of crossover for structured representations. We will be reexpressing
operators in meta-level rules in hopes of apportioning credit for successful
adaptations among them.

Tool:
  Pattern-invoked (RETE) rule interpreter which inferences over terms of 
a domain. Terms and relations (including rules themselves) are declaratively
represented in a frame system, rules are the procedural updating mechanisms
for the domain, much like in KRYPTON. Right now LISP code generates new rules
by, e.g., a crossover-like operater, but, as mentioned, we are working to have
that taken over by a set of meta-rules (summer).

Results:
  There are some problems with just importing the bucket brigade into a
production system environment for rule evaluation (mainly the one mentioned
above and the confounding of rule utility and its "reliability"). For
unsupervised domains the algorithm actually seems more appropriate. Keller
has found a nice meta-level procedure for speeding up apportionment which is
a powerful generalization of the bridging idea. Crossover is an interesting
search operator which is hard to interpret for context-sensitive
representations, but something in the ballpark has been implemented, and the
right thing can probably be done with a powerful grammatical representation
of rule derivations. We haven't done it yet, though.

Problems:
  Other than catching the AI chill? We don't have a measure for partial match
that we really believe in/ is reasonable to compute for production systems,
although we get by with OPS-like syntactic solutions. As mentioned, the
tracking of credit in multiple worlds is pretty bad, but even worse is that
sequences of new hypotheses (new worlds) that unify in "normal" systems should
properly fail to unify in ours. This can happen because the same rule will
usually have different wealths in the worlds leading up to the unified world.
  Our (my) main problem, though, is the bit-string representation that is
erroneously (yes!) viewed as optimizing implicit parallism. Some chance that
we'll be able to turn our attention to that this summer, after meta-level
inference is in place. (We should all be so lucky to have such problems!)

Jim Antonisse
MITRE

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

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

