COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	siglun[f85,jmc]		Notes for Siglunch on 10,000 vs. 10,000,000
C00008 ENDMK
C⊗;
siglun[f85,jmc]		Notes for Siglunch on 10,000 vs. 10,000,000

1985 Nov 15

	The pattern-action rule paradigm is characteristic of present
expert system technology.  It has a rather complex relation to
intelligence characterized by the following.

	1. Most present applied AI systems are essentially production
rule systems.  I regard fancy displays and certainty factors as decorations
in their relation to actual intelligence, however important they may
be in the practical utility of the systems.

	2. Much intelligent human activity can indeed be modeled by
pattern-action rules.

	3. The principle of modesty in AI tells us to be enthusiastic
about a technique for modeling any important aspect of intelligence even
if the technique has important limitations.  So it is with pattern-action
rules and present expert system technology.

	4. However, pattern-action rules are limited in their ability
to directly express human knowledge.  There is a catch to this statement
of limitation expressed in the use of "directly".

	5. In the first place, pattern-action rules constitute a
universal programming language, and therefore any reasoning system
that uses any symbolic expression of facts can be coded as a pattern
action rule system.  However, this will involve many objects and
predicates in the pattern action system that purely concern the
symbol manipulation and maybe even none that are directly interpretable
in the world.  However, the sense of "directly" suggested by these
remarks has never been made precise.  If we prescribe a language we
may then be able to prove something about what can and cannot be
done by a pattern-action system using this language.

	6. Secondly, however, there is some hope of using pattern-action
systems to encode a more sophisticated fact representation system
without going all the way to using pattern-action rules as a mere
programming language.  If we suppose that pattern-action rules
model human or animal behavior partially, then this is a direction
we should consider.  Namely, the more sophisticated fact representation
system should be taken as an elaboration of a simple pattern-action
system.  The approach that seems plausible to me is to add certain
abstract objects to the concrete objects of the simple systems
and have pattern-action rules involving them.

	7. Suppose one tries to build a pattern-action system covering
a domain where more sophisticated knowledge representation and reasoning
is required and indeed used by humans.  The symptom is likely to be
a need for a vast number of rules.  Moreover, it will turn out that
many of the rules so encoded are made up by the expert or the knowledge
engineer and cannot plausibly be regarded as already present in the
expert.


	The rest of this talk consists of examples of phenomena where
"direct" pattern action rules don't seem to cover human knowledge and
their direct use doesn't cover human reasoning.  There isn't time in
preparing this written version to do more than list some areas in
which I have found examples.

	Here they are.

	1. MYCIN doesn't know bacteriology, and maybe it needn't.

	2. The rationalization of canning food to preserve it.

	3. What information is in the chess book?  Barbara Liskov's
thesis.

	4. Dogs and trash cans.

	5. Theoretical use of knowledge in general.

	6. Birthday announcement.

John McCarthy