COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.S Basic Research in Artificial Intelligence and Formal Reasoning
C00007 00003	.ss Recent Results in Formalizing Common Sense Concepts
C00011 ENDMK
C⊗;
.S Basic Research in Artificial Intelligence and Formal Reasoning

//pers John#McCarthy, Richard#Weyhrauch, Lewis#Creary, Jon#Doyle,
Richard#Gabriel, Jussi#Ketonen, Carolyn#Talcott, 
sra Jitendra#Malik, Joe#Weening


Applied research requires basic research to replenish the stock of ideas
on which its progress depends.

The long range goals of our work in basic AI and formal reasoning
are to make computers carry out the
reasoning required to solve problems.
This has involved studying the facts of the common sense world
and appropriate methods of common sense reasoning - both rigorous and
conjectural.  These facts and modes of reasoning constitute the
%2epistemological%1 (knowledge theoretic) aspect of the artificial
intelligence problem.  We have studied it apart from %2heuristic%1 (search
and pattern matching) problems as much as possible.

The need to study epistemological problems of AI apart from heuristics
has recently been recognized by many AI researchers.  We can cite Allen
Newell's not yet published presidential address to the AAAI
on the "logical level" of programs as well as work
by McDermott at Yale and Robert Moore at SRI.

Research in formalizing the facts of the common sense world for AI purposes
began with McCarthy's 1959 "Programs with Common Sense".  The much
used situational calculus was proposed in [McCarthy and Hayes 1969].
In recent years attention has turned to formalizing facts about knowledge,
belief and wants.  The problem of formalizing partial information about
concurrent processes long been a major barrier to progress in AI.

The work on rigorous reasoning has led to interactive proof generating
and proving programs including Weyhrauch's FOL and the newer EKL of
Ketonen.  Continued advances make it easier and easier to prepare
machine-aided and
machine-checkable proofs of mathematical results in general and the
correctness of computer programs in particular.

The work on conjectural reasoning has led to %2circumscription%1, a form
of non-monotonic reasoning described in [McCarthy 1980a].
Other forms of non-monotonic reasoning have been developed.

Indeed the epistemological studies have reached a point where we can
propose to make an experimental Advice Taker to implement the "intelligence"
ANALYST mentioned in our 1979 proposal as a potential long range
application.

Until now the main work of the Formal Reasoning Group has been
theoretical.  We plan to continue the theoretical work, but it
has advanced to the point where a major experimental program aimed at
eventual applications is possible.  We have decide to implement
a version of the Advice Taker proposed by McCarthy in 1958 [McCarthy
1959] concentrating on the "intelligence" ANALYST mentioned in our
previous proposal.

The advances that make ANALYST a feasible project include McCarthy
and Creary results on formalizing mental qualities and results of
McCarthy, Doyle and others on non-monotonic reasoning.  Besides our
own results, we rely on work by others including ...


.ss Recent Results in Formalizing Common Sense Concepts


Although much of our work is technical and abstract,  the results will have
important practical applications.  Also, consideration of solutions to
practical problems helps to isolate and clarify some of the technical issues.
For example, consider the problem of designing and effectively using data bases.
For data bases to include many types of information that decision makers
really need will require major advances in representation theory.  Programs
that use the information effectively impose requirements on the
representation and also need new modes of reasoning.  Thus current data
base technology at best allows simple relations to be represented - e.g.
"Smith is the supervisor of Jones."  Additions from current AI techniques
would allow simple generalizations of relations ("Every employee has a
supervisor except the director."), but this leaves a tremendous range of
representation problems untreated:
.skip 1
.bs
1. Mental states - what a person or group believes, knows, wants, fears, etc.

2. Modalities - what may happen, what must happen, what ought to be done, what
      can be done, etc.

3. Conjectures - if something were true what else would be the case.

4. Causality - how does one event follow because of another.
The preconditions of events and the consequences of events.
concurrent events and their laws of interaction and non-interaction.

5. Actions and their modifiers, e.g. "slowly".  

6. Ability - conditions under which a person or group can do something.

7. Obligation or owing.
.end

Facts of these kinds cannot be adequately represented in data bases at
present, and there are undoubtedly other phenomena essential for
intelligence which have yet to be discovered.  Before such facts can be
incorporated in data bases and question-answering programs in a general
way, basic research must determine the logical structure of these
concepts.

Before discussing our planned work on these epistemological problems
technically, we present an experimental application that we are now in a
position to undertake.