COMMENT ⊗   VALID 00004 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	%ssp[f88,jmc]		Notes for talk to SSP forum
C00013 00003	\smallskip\centerline{Copyright \copyright\ \number\year\ by John McCarthy}
C00014 00004
C00025 ENDMK
C⊗;
%ssp[f88,jmc]		Notes for talk to SSP forum
\input memo.tex[let,jmc]
\title{Formalizing Common Sense Knowledge and Reasoning in Mathematical Logic}
\centerline{Symbolic Systems Forum, 1988 October 21}

	This is to help people remember the points made in the talk.  It
isn't a connected exposition.

1. We can approach AI through

	a. Neurophysiology.  Learning about and simulating the nervous
system.  Connectionism takes ideas from neurophysiology but then goes
off independently.

	b. Psychology.  Experiments with human subjects.  Try to make
programs that behave similarly.  Newell and Simon.

	c. Computer science.  Study the information and goal-seeking
situations presented by the common sense world or some special domains
within it.  Develop methods for intelligent behavior.

	To claim that any of the above approaches is bound to fail goes
against past scientific experience.  To claim that any of them is bound
to succeed quickly goes against recent experience.

2. Within the computer science approach, logical languages are used to
differing extents.

	a. Daniel Dennett, Allen Newell, John McCarthy and
Stanley Rosenschein have all published papers advocating
ascribing mental qualities to machines when this helps understand
their behavior.  John Searle strongly disagrees.

	b. Expert systems often use logical sentences, but derive them
by ad hoc programs.

	c. Other expert systems use quantified implications as rules
but treat them differently from ground sentences used as facts.  They
have to do this because controlling logical inference is still not
well enough understood.

	d. Since my 1958 paper ``Programs with Common Sense'', I have
worked on representing knowledge of the common sense world in
mathematical logical languages.  One goal is a database of general
common sense knowledge that could be used by programs written after
the database was stuffed.  Another goal is the control of reasoning
by facts included in the database itself or derived by reasoning
about reasoning methods.

	3. Epistemologically adequate languages are capable of
expressing the facts, both general and special, that can actually
be obtained with the available opportunities to observe and
communicate.  Epistemological adequacy has been hard to achieve.

	4. There is no ultimate language.  Elaboration tolerant
formalisms permit improvements without throwing away what is
already known.  For example, an elaboration tolerant physics language
would permit the program itself to change the predicates and
functions in which volume depends only on pressure to one in which
it depends on temperature as well.

	Formalization of common sense knowledge and reasoning
requires ontologically rich languages.  Some kinds of entities
dismissed as meaningless by postivist minded philosophers have
to be used, often in ``approximate theories''.

	5. Ambiguity tolerance is also required in an adequate
language.  Thus the law that ``It is illegal to conspire a
Federal official'' can be used in unambiguous cases even though
it has a {\it de re} - {\it de dicto} ambiguity that judges have
had to resolve.

	6. Logical deduction as a means of inference has to be
supplemented by nonmonotonic inference.  Formal systems of nonmonotonic
reasoning were first developed in the late 1970s.  That's a short
time ago.  The circumscription method of nonmonotonic reasoning
involves formulas that express choosing minimal models in useful
orderings of the models of a theory.  Mathematically it is
analogous to the calculus of variations which treats the minimization
of an integral subject to constraints on the values of other
integrals.  There is also an analogy to linear programming.

	7. Situation calculus using $s' = result(e,s)$ is useful
but inadequate.

	8. Formal treatment of context $holds(p,c)$ is required.
Linguists and philosophers treat context too informally for AI
purposes.

	9. Formalization of common sense requires giving up some
of the characteristics of most usual scientific theories.  For example,
common sense dynamics is incomplete.  It only predicts the consequences
of some events, whereas the theory of gravitation predicts the
future of any system of gravitating bodies.  This incompleteness
is common in engineering.  For example, the specifications of
a flipflop only predict its performance under certain inputs.

	9. AI has important methodological differences from both
linguistics and philosophy in treating problems phenomena common
to the three domains.  For example, they ask, ``Is there a language
of thought?''  We ask, ``What kind of language of thought should
we build into our robot, and how should it differ from languages
of communication?''  What kind of communication language is needed
for communication among computers belonging to different organizations?
I believe linguists and philosophers will find that the AI point
of view will help them resolve or dissolve certain questions.

	10. Here are some formulas discussed in the lecture.  They
differ in how much they depend on context, i.e. on whether they
are appropriate for a general common sense database.

$$at(jmc,airport)$$
%
$$at(jmc,SJO,S91)$$
%
$$holds(at(jmc,SJO),S91)$$
%
$$value(location jmc,S91) = SJO$$
%
$$holds1(at(jmc,SJO),C85)$$
%
$$(∀p e s)(¬ab aspect1(f,e,s) ⊃ value(f,s) = value(f,result(e,s))$$
This one is a common sense law of inertia.
%
$$(∀x l s)ab aspect1(location x,move(x,l),s)$$
%
$$(∀x l s)(¬ab aspect2(x,l,s) ⊃ location(x,result(move(x,l),s)) = l$$
%
$$holds1(at(jmc,SJO),restrict(time=t0,C17)) ≡ holds1(when(t0,at(jmc,SJO)),C17)$$

	11. This approach to AI presents a number of reasonably well-defined
problems ranging from specific axiomatizations to developing whole theories.
The names of some of them are: dogs and trash cans, towers of blocks, missionaries
and cannibals, Lifschitz's list of problems in nonmonotonic reasoning,
concurrent events, continuous events, Mr. S and Mr. P.

	12. These matters will be treated in Winter 89 in
Computer Science 323 cross-listed as Philosophy 326.  Yoav Shoham's
CS 324 is also relevant.

	13. The Formal Reasoning Group in the Computer Science
Department does research in these matters.  Other members of the
group involved in formalizing common sense knowledge include
Vladimir Lifschitz and Arkady Rabinov.  In the Computer Science
Department, Nils Nilsson, Michael Genesereth and Yoav Shoham are
interested in somewhat similar problems.
\smallskip\centerline{Copyright \copyright\ \number\year\ by John McCarthy}
\smallskip\noindent{This draft of SSP[F88,JMC] TEXed on \jmcdate\ at \theTime}
\vfill\eject\end

Psychologists ask, ``Is there a language of thought?''

AI people ask, ``Should we design a language of thought?''

How should it differ from a language of communication?

	a. His name is Ivan Sag.  The internal designators of the
person to whom I am referring are clear to me.  If I want to say
it to an audience (of one or more), I have to think about what
they know.  Question for psychologists: At what age does a child
begin to adapt his designators to his audience?  What is the order
of steps in this process?

	b. Designation of nouns is one problem.  Suppose I am
asked whether a certain person is suitable for a certain job.
My opinion is based on my experience.  It is often difficult
to formulate a characterization in terms comprehensible to another
person.

elaborations of at(jmc,airport)

legitimacy of all approaches to AI.  An argument that any of the three
couldn't possibly succeed is not supported by previous scientific
experience.

Differences from what I imagine the more usual SSP approaches to be.

nonmon

flipflop example

attack Hayes's histories??? No.

design stance

need for formality