Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10 5/3/83; site rochester.UUCP
Path: utzoo!linus!decvax!harpo!eagle!mhuxi!houxm!ihnp4!zehntel!hplabs!hao!seismo!rochester!gary
From: gary@rochester.UUCP
Newsgroups: net.ai
Subject: Re: "Rational Psychology"
Message-ID: <3352@rochester.UUCP>
Date: Tue, 11-Oct-83 12:37:52 EDT
Article-I.D.: rocheste.3352
Posted: Tue Oct 11 12:37:52 1983
Date-Received: Fri, 14-Oct-83 02:24:34 EDT
References: <12279@sri-arpa.UUCP>
Organization: U. of Rochester, CS Dept.
Lines: 50

This is in response to John Black's comments, to wit:

>     Having a theoretical (or "rational" -- terrible name with all the wrong
> connotations) psychology is certainly desirable, but it does have to make some
> contact with the field it is a theory of.  One of the problems here is that
> the "calculus" of psychology has yet to be invented, so we don't have the tools
> we need for the "Newtonian mechanics" of psychology.  The latest mathematical
> candidate was catastrophe theory, but it turned out to be a catastrophe when
> applied to human behavior.  Perhaps Periera and Doyle have a "calculus"
> to offer.

This is an issue I (and I think many AI'ers) are particularly interested in,                         
that is, the correspondence between our programs and the actual workings of
the mind. I believe that an *explanatory* theory of behavior will not be at
the functional level of correspondence with human behavior. Theories which are
at the functional level are important for pinpointing *what* it is that people
do, but they don't get a handle on *how* they do it. And, I think there are 
side-effects of the architecture of the brain on behavior that do not show up
in functional level models. 

This is why I favor (my favorite model!) connectionist models as being a
possible "calculus of Psychology". Connectionist models, for those unfamiliar
with the term, are a version of neural network models developed here at 
Rochester (with related models at UCSD and CMU) that attempts to bring the
basic model unit into line with our current understanding of the information
processing capabilities of neurons. The units themselves are relatively stupid
and slow, but have state, and can compute simple functions (not restricted to
linear). The simplicity of the functions is limited only by "gentleman's
agreement", as we still really have no idea of the upper limit of neuronal
capabilities, and we are guided by what we seem to need in order to accomplish
whatever task we set them to. The payoff is that they are highly connected to
one another, and can compute in parallel. They are not allowed to pass symbol
structures around, and have their output restricted to values in the range
1..10. Thus we feel that they are most likely to match the brain in power.

The problem is how to compute with the things! We regard the outcome of a
computation to be a "stable coalition", a set of units which mutually 
reinforce one another. We use units themselves to represent values of
parameters of interest, so that mutually compatible values reinforce one
another, and mutually exclusive values inhibit one another. These could
be the senses of the words in a sentence, the color of a patch in the
visual field, or the direction of intended eye movement. The result is 
something that looks a lot like constraint relaxation.

Anyway, I don't want to go on forever. If this sparks discussion or interest
references are available from the U. of R. CS Dept. Rochester, NY 14627.
(the biblio. is a TR called "the Rochester Connectionist Papers").
   
gary cottrell   (allegra or seismo)!rochester!gary or gary@rochester