Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!cs.utexas.edu!hellgate.utah.edu!helios.ee.lbl.gov!nosc!ucsd!ucsdhub!hp-sdd!ncr-sd!ncrcae!usceast!park
From: park@usceast.UUCP (Kihong Park)
Newsgroups: comp.ai
Subject: Re: Chaos and AI
Message-ID: <3142@usceast.UUCP>
Date: 15 Mar 90 01:46:40 GMT
References: <6925@cps3xx.UUCP>
Organization: University of South Carolina, Columbia
Lines: 33

In article <6925@cps3xx.UUCP> leekin@frith.egr.msu.edu () writes:
>How does Chaos related to AI disciplines such as Connectionism?

There was some recent discussion concerning this topic on comp.ai.neural-nets.
Some notes on this issue:
	It seems that from a physiological standpoint of view, chaotic
neural activity patterns might be of great interest. But, apart from that,
there is this more general aspect. If one has a system consisting of simple,
homogeneous elements arranged according to some connectivity pattern
(possibly local), then the time-evolution of the global configuration of
the system might very possibly be chaotic. In fact, the so-called class
3 cellular automata(Wolfram) are often of this kind.

	In many cases, the evolving pattern is so disorganized as to
escape polynomial-time prediction algorithms. This empirical observation
prompted Wolfram to suggest that they be used as pseudo-number generators.

	As to its impact in the design of intelligent systems, if one's
system is based upon a parallel, distributed substrate on which computations
are carried out, then it is to be expected that chaotic behavior will be
encountered. In some cases, this might be a useful feature, but in most
cases, it will be an unpleasant hinderence if not controlled well.

	Look at Hopfield nets for example. Fixed points and limit cycles
are the "desired" behavior patterns which can encode memories whereas
erratic behavior is not so. One final comment may be due: namely, it has
to be remembered that in any discrete, deterministic, finite system,
the system's behavior will eventually be periodic. But, before the period
is reached, the behavior pattern up to then might very well look
chaotic. And in case of neural nets where the effective connectivity is
variable, one is even less subject to the periodicity constraint. But,
again, I think it is the periodic aspect, and less the chaotic aspect 
which will be most useful in the design of intelligent systems.