Newsgroups: rec.arts.int-fiction
Path: gmd.de!nntp.gmd.de!xlink.net!howland.reston.ans.net!europa.eng.gtefsd.com!news.umbc.edu!eff!news.kei.com!ub!acsu.buffalo.edu!goetz
From: goetz@cs.buffalo.edu (Phil Goetz)
Subject: Controlling plots with chaos
Message-ID: <CqqnHr.EGE@acsu.buffalo.edu>
Summary: An agent architecture
Keywords: Maes, Hopfield, chaos theory, agents
Sender: nntp@acsu.buffalo.edu
Nntp-Posting-Host: hydra.cs.buffalo.edu
Organization: State University of New York at Buffalo/Comp Sci
Date: Wed, 1 Jun 1994 21:53:50 GMT
Lines: 80

Here's an off-the-wall idea, not really thought out:

We want characters that act in a reasonable way, with some independence.
But we also want them to do things that the author scripts into the story.

Envision the character as controlled by a network of nodes which stand
for environmental conditions and behaviors.  Activation spreads from
the current behavior and the current environmental condition.  This
is what Maes' architecture (refs below) does.  She provides a way
to automatically construct a reasonable network, so that conditions
spread activation forward to appropriate behaviors, and desired behaviors
spread activation backward to behaviors that enable them.

Expand the network to include facts that describe the state of the world.
You may (this is what my dissertation proposal is about) include
mental actions of inference as "behaviors" as well.  Now your agent
is ready to face the world.

But we want it to 1. do the things we want it to in the situations we
want it to, and 2. do reasonable things the rest of the time.

Call to mind the Hopfield network [Hopfield 1982].  You have a net of
interconnected nodes.  You have a set of patterns which you wish to
store.  Each pattern tells whether each node is on or off.  You
assign weights between the nodes such that the patterns you want to
recall are stable basins of attraction, meaning that given a pattern
nearby, the network will fall into the pattern to recall.

Look at your agent network again.  Enumerate the situations and responses
you desire.  For each one, set a pattern for the network where the
situation and response are "on" and everything else is "off".
Train the network so those patterns are stable attractors.
Then the given situation will provoke the desired behavior, and
similar situations probably will.

The problem is that the Hopfield net does not train, and so if we
use the straight Hopfield weight-assigning rule, we won't get reasonable
behavior in other situations.  We need a weight-assignment rule which
incorporates aspects of both Hopfield's method and Maes' method,
and we might want some learning also.  For instance,  when a goal is
attained after a behavior, or one behavior is followed by another,
strengthen the connections between them.  We might also use a
bucket-brigade algorithm [Holland 1981, 1992] to allocate credit for
accomplishing goals.  This would make successful behaviors "stronger"
(have a higher base level of activation).

Phil goetz@cs.buffalo.edu

(The thread title refers to the chaos-theoretic underpinnings of
the Hopfield net.)


References and further reading:

John Holland (1985).  "Properties of the bucket brigade."  In
J.J. Grefenstette (ed.), _Proceedings of an International Conference
on Genetic Algorithms and their Applications_: 1-7.

John Holland (1992).  _Adaptation in Natural and Artificial Systems_,
2nd edition: 176-179.  MIT Press.  Note: This is not in the 1975 edition.

J. Hopfield (1982).  "Neural networks and physical systems with
emergent collective computational abilities."  In _Neurocomputing_,
ed. Anderson and Rosenfeld, MIT Press 1988.  Originally in
Proc of the National Academy of Sciences 79: 2554-2558.

Pattie Maes (1989).  The dynamics of action selection.
_Proceedings of the Eleventh International Joint Conference on
Artificial Intelligence_ (IJCAI-89): 991-997.  San Mateo, CA: Morgan Kaufmann.

Pattie Maes (1991a).  A bottom-up mechanism for behavior selection in
an artificial creature.  In [Meyer and Wilson 91 (From Animals
to Animats)]: 238-246.

Pattie Maes (1991b).  The agent network architecture (ANA).
_SIGART Bulletin_ 2(4): 115-120.

Christine Skarda and Walter Freeman (1987).  "How brains make chaos
in order to make sense of the world."  Behavioral and Brain Sciences
10: 161-173 (commentary thru p. 195).
