COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	%chen[e84,jmc]		Non-monotonic reasoning essay for NSF
C00009 ENDMK
C⊗;
%chen[e84,jmc]		Non-monotonic reasoning essay for NSF
%
%500 to 600 words
%for current NSF grant
%send to
%Su-Shing Chen
%Program Director
%Intelligent Systems
%202/357-7345
\magnification\magstep1
\centerline{\bf Non-monotonic Reasoning Provides New Tools for Artificial
 Intelligence}
\bigskip
\baselineskip = 20pt

	Research in artificial intelligence (AI) is concerned with making
computers behave intelligently.  Part of this work is concerned with
special domains like diagnosing diseases, playing chess or evaluating
ore deposits.  However, experience with "expert systems" programmed
for some narrow domain has shown that there is a large difficult
problem still unsolved.  This amounts to giving computers some equivalent
of human common sense.  Without it the expert systems are very limited
and very difficult to modify to meet new requirements.  They also don't
know their own limitations.

	In order to make computer programs behave with common sense, they
 must be provided with facts about the
common sense world and ways to reason with these facts.  This is how they
will decide what to do to achieve the goals we give them.  Therefore, AI
researchers want to create a database of general facts about the common
sense world and ways of reasoning with these facts.

	The most useful way of expressing facts about the world involves
using the languages of mathematical logic.  Mathematical logic has been
developing since the middle of the last century, but it has mainly concerned
with expressing the facts of pure mathematics, although philosophers
have also used it.  Artificial intelligence has used it heavily, but since
the middle 1970s, it has become clear that the rigorous modes of reasoning
of mathematical logic have to be supplemented by equally well defined
ways of making useful conjectures.  Proposals for doing this go under
the general name of {\it non-monotonic reasoning}.  That intelligent
programs must make conjectures has been obvious all along.  What's new
is that precise methods of doing so have been discovered that fit nicely
into mathematical logic and supplement its old methods and which are
suitable for use by computers.

	Professor John McCarthy of the Computer Science Department of
Stanford University is one of the pioneers of artificial intelligence
research.  In the middle 1970s, he invented one of the important
methods of non-monotonic reasoning and has been developing it and
applying it to artificial intelligence problems with NSF and DARPA support
since that time.  He calls his method {\it circumscription}.

	Circumscription provides a way of telling the computer
that we want it to form the simplest or most standard interpretation
of the facts it is taking into account.  It also provides ways
for the computer to decide what other facts can be inferred given
that we want the simplest interpretation of the initial facts.
It often happens that a conclusion that holds in the simplest interpretation
of given facts does not hold any more in the simplest interpretation
of an enlarged collection of facts.  The phenomenon that some previous
conclusions may disappear when the set of assumptions is enlarged
doesn't occur in mathematical logic and is the reason for the mathematical
term {\it non-monotonic}.

	For example, if I tell you that I have a bird and hire you
to build me a cage for it, you will infer that the cage needs a top so the
bird won't be able to fly away.  However, if I add the fact that
my bird is a penguin, you will no longer draw the conclusion that
a top is needed.

	If circumscription is available to the computer reasoning
system, the common sense database can be provided with the fact
that a bird can be assumed to be able to fly unless there is information
to the contrary.  Without some form of non-monotonic reasoning,
whoever created the database would have to list all possible
exceptions to the ability of birds to fly.  This is entirely
impractical.

	Extending mathematical logic to include non-monotonic reasoning
has applications to AI (as has been said) but also to general
database technology, to the theory of computer programs and
even to philosophy.  It is leading to renewed efforts to provide
a general database of common sense facts and reasoning methods
for computer programs to use.
\end