COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002		This is to defend nonmonotonic reasoning from Joseph's
C00006 ENDMK
C⊗;
	This is to defend nonmonotonic reasoning from Joseph's
Agassi's criticism in his July 1988 review of the Daedalus special
issue on AI.

d69%taunos.bitnet@forsythe.stanford.edu
garble?
	I'm about to respond in a letter to the editor to the
criticism in your review of my article in the Daedalus special issue
on AI.  However, the second sentence quoted from your review seems to
be garbled.  Can you confirm it or give me a corrected version?  On
page 21, you wrote,

     ``The claim that ordinary reasonable thinking is
     nonmonotonic is false.  The examples McCarthy takes are
     not of violation of logic but of statements taken in
     the context in which they are stated, the context
     statements taken as stated and agreed upon, and of
     statements of context easily and naturally altered upon
     correction.''

By the way, would you accept the correction that I didn't say all
ordinary thinking is nonmonotonic - only some important parts of it?

Sincerely,
John McCarthy

Consider the sentence, ``Fred owns a bird named Tweety'', representable
in logic as 
%
\eqno $$owns(Fred,Tweety) ∧ bird(Tweety).$$
%
	Any of the well known
systems of nonmonotonic reasoning (circumscription, the logic
of defaults, or autoepistemic logic) admits the inference from (1)
and suitably encoded common sense knowledge (A) about birds of
%
\eqno $$flies(Tweety).$$
%
Each of these methods works with (1) as expressed above in first
order logic.  If also given the sentence, ``Tweety is a penguin'',
expressed as
%
\eqno $$penguin(Tweety),$$
%
none of them will infer (2).  To put it more
precisely, a program taking into account (1), (3) and (A) will not
infer (2) if it uses any of the above-mentioned systems.

	I have been told that probability theory will do even better,
taking into account some such fact as ``The probability that a randomly
mentioned bird can fly is 0.95, but no penguin can fly.''  Indeed
a human user of probability theory will do better, but suppose the
human wants express his knowledge of the particular facts in logical
sentences, so that the conclusions with and without the fact
that Tweety is a penguin follow logically.  When the program works without
the fact that Tweety is a penguin, it must use some encoding of the
fact that this is all that was said.