COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002
C00011 ENDMK
C⊗;

John McCarthy will direct the work on formal reasoning and basic
artificial intelligence and will continue his own research in the
epistemological problems of artificial intelligence and the theory of
program verification.  Besides this he will take part in the specification
and implementation of ANALYST.

McCarthy's recent research efforts have mainly involved extending the
circumscription method of non-monotonic reasoning described in [McCarthy
1980].  Formal methods of non-monotonic reasoning are a development of the
last five years.

There is a good prospect for breakthroughs on the long-standing
"qualification problem" in the next few years using circumscription and
various generalizations.  Two ideas are involved.

First, frame axioms are derivable from a slight generalization of the
circumscription method of [McCarthy 1980].  The generalized
circumscription gives a sentence schema stating that certain functions and
predicates minimize the truth value of a certain sentence while holding
other sentences true.  It is a logical analog of the mathematical method
of minimization subject to constraints of the calculus of variations
although it is not yet apparent that the analogy can be exploited.  It
allows us to specify that the only entities that change when an event
takes place are those that the description of the event and associated
facts force to change. Circumscription permits this to be done without an
initial commitment as to what these facts will be.

For example, we can tell ANALYST that the event of a person moving from
one city to another leaves fixed everyone whom it doesn't say moves.  The
fact that a person's family moves with him is then an add-on rather than a
modification.  This meet the requirement that the person or computer
program adding the fact that when a person moves his family also moves
need not understand the initial axiom giving the effects of moving.  It is
a key property of intelligent systems that they can be advised and their
behavior modified by someone who does not understand the details of how
they work.  It will be especially important when knowledge representation
and problem solving systems like ANALYST contain many tens of thousands of
facts.

The second new application of circumscription is more difficult but
probably more far-reaching.  Suppose we need to tell ANALYST that the
Russians sometimes attempt to bribe foreign officials to reveal secrets
and that they can be embarassed when such attempts are revealed.  This
seems straightforward and a human would have no difficulty absorbing the
information.  However, for any present knowledge representation system to
accept the information, a number of very pedantic issues would have to be
resolved: (1) %2Does attempting to bribe an official include offering a
bribe to someone who is mistakenly believed to be an official?  (2) Does
it include offering a bribe to someone who is not believed to be an
official but turns out to be?  Must there be an individual who is offered
the bribe or can letting it be known that a bribe is available count as an
attempt to bribe an official%1?  Intuitively, all these issues seem to be
irrelevant.  They proably won't come up, and if they did, they could be
resolved in an ad hoc way.  However, present knowledge representation
systems require require definitions applicable to all cases.

Circumscription and other forms of non-monotonic reasoning apparently
allow us to make systems that may be characterized as "ambiguity tolerant"
- something that Hubert Dreyfus said was impossible for computers.  The
idea, not yet well worked out, is to be able to say that a sentence is
unambiguous unless two interpretations are proposable.  This will allow an
intelligence specialist to tell ANALYST about propensities to attempt to
bribe without being aware that ambiguities are possible, and it will allow
ANALYST to use the concept without problem until and unless ambiguous
examples actually arise.

McCarthy will also continue his research on the ELEPHANT formalism for
representing sequential programs in first order logic.

Goals and milestones

⊗ December 1981 Completion of paper on ELEPHANT.

⊗ June 1982 Completion of paper on use of circumscription to
get "ambiguity tolerance".