Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!linus!decvax!harpo!utah-cs!shebs
From: shebs@utah-cs.UUCP (Stanley Shebs)
Newsgroups: net.ai
Subject: Re: Rational Psychology
Message-ID: <1913@utah-cs.UUCP>
Date: Mon, 19-Sep-83 17:43:53 EDT
Article-I.D.: utah-cs.1913
Posted: Mon Sep 19 17:43:53 1983
Date-Received: Thu, 22-Sep-83 00:49:28 EDT
Lines: 72

I just read Jon Doyle's article about Rational Psychology
in the latest AI Magazine (Fall '83), and am also very interested in
the ideas therein.  The notion of trying to find out what is *possible*
for intelligences is very intriguing, not to mention the idea of developing
some really sound theories for a change.

Perhaps I could mention something I worked on a while back that appears
to be related.  Empirical work in machine learning suggests that there
are different levels of learning - learning by being programmed, learning
by being told, learning by example, and so forth, with the levels being
ordered by their "power" or "complexity", whatever that means.
My question: is there something fundamental about this classification?
Are there other levels?  Is there a "most powerful" form of learning,
and if so, what is it?

I took the approach of defining "learning" as "behavior modification",
even though that includes forgetting (!), since I wasn't really concerned
with whether the learning resulted in an "improvement" in behavior
or not.  The model of behavior was somewhat interesting.  It's kind
of a dualistic thing, consisting of two entities:  the organism and
the environment.  The environment is everything outside, including
the organsism's own physical body, while the organism is more or less
equivalent to a mind.  Each of these has a state, and behavior can be
defined as functions mapping the set of all states to itself.  Both the
environment and the organism have behaviors that can be treated in the
same way (that is, they are like mirror images of each other).   The
whole development is too elaborate for an ASCII terminal, but it boiled
down to this:  that since learning is a part of behavior, but it also
*modifies* behavior, then there is a part of the behavior function that
is self-modifying.  One can then define "1st order learning" as that
which modifies ordinary behavior.  2nd order learning would be "learning
how to learn", 3rd order would be "learning how to learn how to learn"
(whatever *that* means!).  The definition of these is more precise than
my Anglicization here, and seem to indicate a whole infinite heirarchy
of learning types, each supposedly more powerful than the last.  It doesn't
do much for my original questions, because the usual types of learning
are all 1st order - although they don't have to be.  Lenat's work on
learning heuristics might be considered 2nd order, and if you look at it
in the right way, it may actually be that EURISKO actually implements
all orders of learning at the same time, so the above discussion is
garbage (sigh).

Another question that has concerned me greatly (particularly since building
my parser) is the relation of the Halting Problem to AI.  My program
was basically a production system, and had an annoying tendency to get
caught in infinite loops of various sorts.  More misfeatures than bugs,
though, since the theory did not expressly forbid such loops!  To take
a more general example, why don't circular definitions cause humans to
go catatonic?  What is the mechanism that seems to cut off looping?
Do humans really beat the Halting Problem?  One possible mechanism is
that repetition is boring, and so all loops are cut off at some point
or else pushed so far down on the agenda of activities that they are
effectively terminated.  What kind of theory could explain this?

Yet another (last one folks!) question is one that I raised a while back,
about all representations reducing down to attribute-value pairs.
Yes, they used to be fashionable but are now out of style, but I'm
talking about a very deep underlying representation, in the same way
that the syntax of s-expressions underlies Lisp.  Counterexamples
to my conjecture about AV-pairs being universal were algebraic
expressions (which can be turned into s-expressions, which can be
turned into AV-pairs) and continuous values, but they must have *some*
closed form representation, which can then be reduced to AV-pairs.
So I remained unconvinced that the notion of objects with AV-pairs
attached is *not* universal (of course, for some things, the representation
is so primitive as to be as bad as Fortran, but then this is an
issue of possibility, not of goodness or efficiency).

Looking forward to comments on all of these questions...

					stan the l.h.
					utah-cs!shebs