ARCHBOLD@SRI-AI.ARPA (01/24/84)
From: Armar Archbold <ARCHBOLD@SRI-AI.ARPA>
[The following is a review of a Stanford talk, "Reflections on AI", by
Dr. Ron Rivest of MIT. I have edited the original slightly after getting
Armar's permission to pass it along. -- KIL]
Dr. Rivest's talk emphasized the interest of small-scale studies of
learning through experience (a "critter" with a few sensing and
effecting operations building up a world model of a blocks environment).
He stressed such familiar themes as
- "the evolutionary function and value of world models is predicting
the future, and consequently knowledge is composed principally of
expectations, possibilities, hypotheses - testable action-sensation
sequences, at the lowest level of sophistication",
- "the field of AI has focussed more on 'backdoor AI', where you
directly program in data structures representing high-level
knowledge, than on 'front-door' AI, which studies how knowledge is
built up from non-verbal experience, or 'side door AI', which studies
how knowledge might be gained through teaching and instruction using
language;
- such a study of simple learning systems in a simple environment -- in
which an agent with a given vocabulary but little or no initial
knowledge ("tabula rasa") investigates the world (either through
active experiementation or through changes imposed by perturbations
in the surroundings) and attempts to construct a useful body of
knowledge through recognition of identities, equivalences,
symmetries, homomorphisms, etc., and eventually metapatterns, in
action-sensation chains (represented perhaps in dynamic logic) -- is
of considerable interest.
Such concepts are not new. There have been many mathematical studies,
psychological similations, and AI explorations along the lines since the
50s. At SRI, Stan Rosenschein was playing around with a simplified learning
critter about a year ago; Peter Cheeseman shares Rivest's interest in
Jaynes' use of entropy calculations to induce safe hypotheses in an
overwhelmingly profuse space of possibilities. Even so, these concerns
were worth having reactivated by a talk. The issues raised by some of the
questions from the audience were also intesting, albeit familiar:
- The critter which starts out with a tabula rasa will only make it
through the enormous space of possible patterns induceable from
experience if it initially "knows" an awful lot about how to learn,
at whatever level of procedural abstraction and/or "primitive"
feature selection (such as that done at the level of the eye itself).
- Do we call intelligence the procedures that permit one to gain useful
knowledge (rapidly), or the knowledge thus gained, or what mixture of
both?
- In addition, there is the question of what motivational structure
best furthers the critter's education. If the critter attaches value
to minimum surprise (various statistical/entropy measures thereof),
it can sit in a corner and do nothing, in which case it may one day
suddenly be very surprised and very dead. If it attaches tremendous
value to surprise, it could just flip a coin and always be somewhat
surprised. The mix between repetition (non-surprise/confirmatory
testing) and exploration which produces the best cognitive system is
a fundamental problem. And there is the notion of "best" - "best"
given the critter's values other than curiosity, or "best" in terms
of survivability, or "best" in a kind of Occam's razor sense
vis-a-vis truth (here it was commented you could rank Carnapian world
models based on the simple primitive predicates using Kolmogorov
complexity measures, if one could only calculate the latter...)
- The success or failure of the critter to acquire useful knowledge
depends very much on the particular world it is placed in. Certain
sequences of stimuli will produce learning and others won't, with a
reasonable, simple learning procedure. In simple artificial worlds,
it is possible to form some kind of measure of the complexity of the
environment by seeing what the minimum length action-sensation chains
are which are true regularities. Here there is another traditional
but fascinating question: what are the best worlds for learning with
respect to critters of a given type - if the world is very
stochastic, nothing can be learned in time; if the world is almost
unchanging, there is little motivation to learn and precious little
data about regular covariances to learn from.
Indeed, in psychological studies, there are certain sequences which
will bolster reliance on certain conclusions to such an extent that
those conclusions become (illegitimately) protected from
disconfirmation. Could one recreate this phenomenon with a simple
learning critter with a certain motivational structure in a certain
kind of world?
Although these issues seemed familiar, the talk certainly could stimulate
the general public.
Cheers - Armar