COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	@define(fl=flushleft,break before,below 0)
C00007 ENDMK
C⊗;
@define(fl=flushleft,break before,below 0)
@heading(Ambiguity Tolerance and Non-monotonic Reasoning)

@center(by John McCarthy, Stanford University)

@Fl(@b(Abstract:)) This paper
 discusses how the circumscription method of
non-monotonic reasoning described in (McCarthy 1980)
can be extended to allow reasoning with
unavoidably ambiguous concepts.  The idea is that both humans
and artificial intelligences must invent and use concepts with
hidden ambiguities but can use these concepts without
noticing any problem until circumstances arise that excite
the ambiguities.  Once noticed, the ambiguities may or may not
be resolved, but in either case the old ability to use the concept
isn't lost.
  It will be argued that this capability
is essential for intelligent behavior, but Dreyfus (1973) was mistaken
in asserting that it can't be formalized.

	Circumscription starts with a sentence of first order logic
and a predicate occurring in that sentences and produces a first
order schema whose intuitive content is that the predicate in question
has the minimal extension compatible with the sentence.  It can be
used to assert that the only entities of a given kind that exist
are those required by the assumptions made or that the only entities
that change when an event occurs are those that the assumptions
require to change.  Applying the method to ambiguous concepts seems
to require modification of the method and extensive reification of
concepts.
	
	@i(De re - de dicto) puzzles are a case in point, though they
probably won't turn out to be the basic case.  Suppose, for example,
that it has been declared a crime to "attempt to bribe a public official".
The concept may be used for years before the question arises of whether
any of the following defenses are valid.

	@i("I didn't know he was a
public official".

	"I mistakenly thought he was a public official,
but he wasn't".

	"When I let it be known that I would pay $5,000 to
any public official who would fix my drunk driving conviction, there
was no-one I was attempting to bribe, since there was no public official
who could fix the conviction").

	Our present interest isn't in resolving the philosophical
problems presented by these puzzles.
Rather we shall propose formalisms usable by
an artificial intelligence to reason with such ambiguous concepts as
@i(attempting to bribe).
Using the formalism should not require that the programmer know about
the possibilities for ambiguity, and the program should be able to use
the formalism without difficulty until ambiguous cases arise.  Even
when it has noticed an ambiguity, it should still be able to use the
formalism in unambiguous cases.

@FlushLeft(@b(References:))

@b(Dreyfus, H., (1973)): @i(What Computers Can't Do: A Critique of
Artificial Reason), Harper and Row.

@b(McCarthy, John (1980)): 
"Circumscription - A Form of Non-Monotonic Reasoning", @i(Artificial
Intelligence), Volume 13, Numbers 1,2, April.