COMMENT ⊗   VALID 00005 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	logica[e85,jmc]		Notes for Barwise Logic and AI conference
C00007 00003	Common Sense --- An AI Approach
C00013 00004	References:
C00017 00005	Non-monotonic reasoning
C00018 ENDMK
C⊗;
logica[e85,jmc]		Notes for Barwise Logic and AI conference

	The topic of this lecture seems to be peripheral to the
approach to the relations between logic and AI taken by CSLI
generally and in the organization of this summer school and
conference.  Conversely, my opinion is that the other approaches
miss the main problems involved in the use of logic in AI.  I even
think they are missing a point central to the issues of philosophy
and language which constitute the origin of the CSLI approach.  My
purpose isn't to debate the issue in this lecture.  I merely want
to begin by pointing out what I consider central and noting that
the others avoid it.

	In order to behave intelligently in the common sense world,
a human, animal or robot must know some facts about the world.  It
must know facts about the specific situation, but it must also know
more general facts including facts about the effects of the actions
it may take and the events that may occur.  The central problem in
applying logic to AI is what these facts are and how to express them
in logical languages and what rules of inference the need to be used
to get the conclusions required in order to take successful action.

	The problem differs from that which has concerned linguists,
because linguists are concerned only with expressing in logical
or other form those assertions which commonly occur in speech or
writing.  However, we don't ordinarily express in language the basic
facts we know about the effects of events in the common sense world.
Here are some examples:

1. The effects of physical events.  If I spill a cup of coffee on
the table, some people should jump aside to avoid being scalded
while others correctly consider themselves too far away to be
affected.

2. The effect of mental events.  A will be offended if B does X
without consulting him.

	There is another important difference between the AI requirements
for formalism and those sought by philosophers and linguists.
While AI is ambitious in its ultimate goals --- to create artificial
systems more intelligent than humans, it is modest in its goals
for specific systems.  We will settle for systems of limited
capability.

	The problem of expressing common sense knowledge in logic
has been around since 1958.  There has only been limited progress.
Other formalisms have been proposed, but they have had only
limited generality.  When attempts are made to generalize them,
they usually amount to a re-introduction of logical formalism such as
quantifiers.

	A big step forward in using logic to express facts about the
common sense world was made with the formalization of non-monotonic
reasoning starting in the late 1970s.  These formal systems include
default reasoning by programs, reason maintenance systems (TMS),
the McDermott-Doyle non-monotonic logic, Reiter's logic of defaults
and my circumscription.  I shall concentrate on circumscription and
its application to AI.  I will also mention some of its mathematical
logical properties and problems.
Common Sense --- An AI Approach

	Artificial intelligence shares with philosophy and linguistics an
interest in identifying the facts of the common sense world and expressing
them in a formal way.  However, out that AI research often takes a
different point of view about what are the important problems.  CSLI and
this summer school seem to be dominated by philosophical and linguistic
considerations.  Here are the main differences as I see them.

	1. Because AI is difficult, AI research has to have modest goals
at present.  While ``Programs with Commmon Sense'' were proposed in 1958,
we still don't have programs with general common sense.  Therefore, we
must settle for formalisms with limited expressiveness.  For example, the
situation calculus (McCarthy and Hayes 1969) is still the basis of much AI
research even though it doesn't handle concurrent events or continuous
time.  The intensional formalisms used in AI are even more modest in their
capability.

	2. On the other hand, much of the information about the common
sense world that AI must represent and use is almost always ignored in
philosophy and linguistics, because humans don't ordinarily represent it
in natural language.  Philosophers and linguists give only lip service to
the fact that human understanding of natural language depends on knowing
general facts about the consequences of actions and other events.

	3. There are a number of questions that philosophers regard
as ``scientific'', i.e. they seek the truth of the matter.  AI, on
the other hand takes an ``engineering'' approach, i.e. seeks to define
formalisms that will work --- in limited domains.  An example is
knowledge about knowledge.  We would like to be able to put in a
common sense database, usable by any suitable program, facts about
what information is in other databases, e.g. what is in the {\it Official
Airline Guide} about air travel.  We also want to express the facts
about what travel agents know and what services they can provide.
For this purpose, AI needn't seek an eternal answer to ``what is
knowledge'', but it will have to go considerably beyond the formalisms
proposed so far, whether in AI or in philosophy, just in order
to make a general air travel advisory expert system.

	5. However, the main area in which AI must go beyond philosophy
and linguistics is in studying the reasoning processes whereby the
facts of a situation can be combined with general knowledge and facts
about goals in order to decide what should be done to try to achieve the
goals.  The heuristic aspects of this problem were noticed first and are
still prominent.  Thus programs often use enormous amounts of computation to
decide what to do in situations humans find rather simple.  However,
much of the heuristic difficulty is a symptom of epistemological
inadequacy of the formalisms used to express knowledge, (i.e. they can't
express what every human knows (and many dogs)) and also a symptom of the
inadequate methods of reasoning heretofore available in mathematical logic
and in present AI systems.

	6. The present lecture concerns a new kind of formal reasoning
developed since 1975 in the AI research community.  The several forms
of non-monotonic reasoning are best considered as extensions to mathematical
logic.
References:

Formalization of  Common Sense Knowledge

progs w c.s.
mcc and hayes
hobbs and moore collection
charniak and mcdermott

Non-monotonic reasoning
mcc 1980
mcd and doyle
tms
reiter
mcc 1985
1984 conference

{\bf American Association for Artificial Intelligence}: {\it Workshop
on Non-monotonic Reasoning held at Lake Mohonk, N.Y. October
1984}.  Some copies may still be available.

{\bf Lifschitz, Vladimir (1985)}: ``Computing Circumscription'' in {\it 
Proc. IJCAI-85}.

{\bf McCarthy, John (1960)}: ``Programs with Common Sense,''
 in {\it Proceedings of the
Teddington Conference on the Mechanization of Thought Processes}, Her Majesty's
Stationery Office, London.

{\bf McCarthy, John and P.J. Hayes (1969)}:  ``Some Philosophical Problems from
the Standpoint of Artificial Intelligence'', in D. Michie (ed), {\it Machine
Intelligence 4}, American Elsevier, New York, NY.

{\bf McCarthy, John (1977)}:
``Epistemological Problems of Artificial Intelligence'', {\it Proceedings
of the Fifth International Joint Conference on Artificial 
Intelligence}, M.I.T., Cambridge, Mass.

{\bf McCarthy, John (1979)}: 
``First Order Theories of Individual Concepts and Propositions'', 
in Michie, Donald (ed.) {\it Machine Intelligence 9}, (University of
Edinburgh Press, Edinburgh).
%.<<aim 325,concep[e76,jmc]>>

{\bf McCarthy, John (1980)}: 
``Circumscription --- A Form of Non-Monotonic Reasoning'', {\it Artificial
Intelligence}, Volume 13, Numbers 1,2, April.
%.<<aim 334, circum.new[s79,jmc]>>

{\bf McCarthy, John (1984)}:
``Applications of Circumscription to Formalizing Common Sense Knowledge''.
This was distributed at the 1984 AAAI
conference on non-monotonic reasoning, which will not have a proceedings
and is being submitted for publication to {\it Artificial Intelligence}.
%  circum.tex[f83,jmc]

{\bf Reiter, Raymond (1980)}: ``A Logic for Default Reasoning'', {\it Artificial
Intelligence}, Volume 13, Numbers 1,2, April.
Non-monotonic reasoning
Logic of defaults - see Reiter 1980
Non-Monotonic logic - see (McDermott and Doyle 1980).
Reason maintenance -

Circumscription

We wish to infer those sentences which are true in the simplest
(or the standard) models of the collection of facts we are taking
into account.