COMMENT ⊗   VALID 00006 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	\input memo.tex[let,jmc]
C00005 00003	\section{Introduction}
C00011 00004
C00012 00005	\smallskip\centerline{Copyright \copyright\ \number\year\ by John McCarthy}
C00013 00006	thomas[e87,jmc]		Position paper for Journal of Philosophical Logic
C00017 ENDMK
C⊗;
\input memo.tex[let,jmc]
\title{Logical Problems Posed by Artificial Intelligence}

\noindent Abstract: The most straightforward approach to AI is to express
what the system needs to know in a first order language and have it
programmed to deduce what it should do to achieve its goals.  This
requires formalizations of common sense knowledge and reasoning.
Specifically, facts about the effects of actions must be formalized.  The
approach has had to be modified to include non-monotonic reasoning, and
further modifications may be required involving the formalization of
context.  However, using logic as the basis for AI still looks promising
--- even after 30 years.  Our object is to explain the logical problems
and what has been accomplished in the hopes that logicians will find them
interesting.

\vfill\eject
\section{Introduction}

	I am grateful for the editor's invitation to contribute a
position paper on the relation between mathematical logic and
artificial intelligence.  This paper will be light on technical
detail, but I hope the references will permit the reader to
pursue the applications.

	My position is at one end of the spectrum of opinion.  Namely,
I believe that logic is of decisive importance for AI.  Mathematical
logical languages and mathematical logical reasoning will permit
intelligent programs that express what they know in logical formulas
and decide what to do by logically inferring that certain actions
are appropriate to achieving their goals.  This approach, first
enunciated in 1958 in (McCarthy 1960) encounters many difficulties
some of which require new logical formalisms, but it is still ahead
of any other approach in generality and promise.


Intelligence may be studied as a branch of biology or as a branch
of computer science.  The former approach regards it as a phenomenon
exhibited by humans and other animals and uses methods of psychological
experiment and theory or neurophysiological experiment and theory.
The computer science approach studies the relations among various
aspects of the world, goals structures that a system may have,
what a system with given opportunities to observe and act can learn
about the world and its particular situation and what strategies
are effective in achieving goals.  Its theories involve the relations
among these entities and its experiments involve testing what strategies
work.  This paper concerns primarily the computer science approach.
Note that the two approaches interact.

Within computer science, opinions and practices differ about the proper
role of mathematical logic as compared to more procedural formalisms or
declarative formalisms purportedly not based on logic.  Some use logical
languages to express information but rely on purportedly non-logical
programming to make inferences and decide what to do.  This paper concerns
using logical inference to decide what to do, first suggested
in (McCarthy 1960).  The emphasis is on {\it epistemologically adequate}
formalisms for expressing facts about the common sense world and the system's
goals and formalizing the reasoning involved in
deciding what to do.  The term {\it epistemologically adequate} refers
to suitability for expressing the information that is actually available
with given opportunities to observe and act.  Success in this endeavor
has been moderate but non-trivial.  The approach presents a very difficult
collection of problems in logic as well as in computer science.

It is evident that Leibniz, Boole and Frege all hoped to include common
sense knowledge in the domain formalizable by mathematical logic.
However, this has proved very difficult, and the difficulties have not
been easy to diagnose, let alone solve.  Since the successful applications
of mathematical logic have been to the foundations of mathematics, some
people have prematurely concluded that this is all it's good for.

Many AI systems use mathematical logic to formalize common sense knowledge
and reasoning.  Systems differ in how much logic is used, and some use far
less than the full first order logic.  For example, many expert systems
never infer new general sentences, but only go from the general to the
particular, and this suffices for many applications.



1. Express the facts about the world as sentences of mathematical
logic.  These include especially the facts about the consequences
of actions.
\smallskip\centerline{Copyright \copyright\ \number\year\ by John McCarthy}
\smallskip\noindent{This draft of THOMAS[E87,JMC]
 TEXed on \jmcdate\ at \theTime}
\vfill\eject\end
thomas[e87,jmc]		Position paper for Journal of Philosophical Logic
special issue on AI and Philosophical Logic, due dec. 1
msg.msg[1,jmc]/165p
e87.in[let,jmc]/165p
THOMASON@C.CS.CMU.EDU


Wittgenstein


	The use of logic in artificial intelligence is controversial
in a variety of ways.  Sometimes it's simple ignorance.

	Many AI researchers do not understand the inevitability of logic.
For example, one moderately prominent AI researcher said, ``Why do you
have this prejudice in favor of this language invented by Russell?''

	Sometimes it's some kind of existential despair.

Problems in formalizing common sense.
	1. ccntext
	2. non-monotonic reasoning
	3. arising from insufficient reification of ordinary kinds