COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	vijay@ernie.berkeley.edu
C00019 00003		In one sense this gang of four are among the last holdouts
C00020 ENDMK
C⊗;
vijay@ernie.berkeley.edu
The rationalist approach to AI

Copyright 1986 John McCarthy

	In their various ways, Winograd, Searle and
the Dreyfus's are attacking rationalism.  I have commented
separately on their essays, but I would like to enunciate a
doctrine about AI that bears some resemblance to what they are
attacking.  In its present form it recognize that some of the
phenomena the Dreyfus's and Winograd have mentioned present
real problems that AI systems must solve.

	I shall begin with the programme of my 1958 "Programs
with Common Sense" (McCarthy 1960) and then modify it.

	It was proposed to represent a robot's information about
the world as sentences of mathematical logic.  This information
would include both general information about the world, including
especially general information about the effects of actions and
other events.  It also includes information obtained through the
robots sense organs about the particular situation.

	Qualification 1.  It was explicitly mentioned in that paper
that certain information, especially pictures, would be too longwinded
to represent pixel by pixel as sentences and would be represented
by any convenient data structure.  It was also stated that certain
computations would most efficiently be made by programs that didn't
do logical deduction per se.

	Goals and a general principle that the robot should do
what was likely to achieve its goals were also to be represented
by sentences.  A subset of the sentences in the memory of the
machine were in a small database corresponding to consciousness.

	The program would attempt to deduce from the sentences
in consciousness a sentence of the form  should(<action>).  When
it did, it would do the action.  The actions included physical
and mental actions.  The latter included modifying consciousness
by getting more information from memory, forgetting some
information, making observations both of the outside world
and the machine's own memory.  Computations with the data
not represented by sentences would be possible actions and
could result in sentences.

	There was an example of deducing from data in consciousness that
a certain two step plan would succeed in getting to the airport.
Beyond this details were not given, because they weren't available.
Nothing was said about how long all this would take.  My opinion was
that conceptual problems remained to be solved, but I certainly expected
faster progress than has occurred.

	I will refer to this approach and sufficiently similar
proposals by others as the "reasoning program" approach.  I don't
claim that it is the only workable approach to AI.  However, its
variants have had the most success so far.

	While many attempts have been made to realize the plan
of "Programs with Common Sense" or variants of it, I have never
felt that the conceptual problems had been solved well enough
for my own next step to be an implementation.  Nevertheless,
progress has been made both in implementing systems meeting
part of the goals and in developing improved concepts.  Some
of the progress, e.g. STRIPS and Microplanner, involved putting
more information in the program or in production rules, thus
avoiding some of the combinatorial difficulties of a pure
logic approach with the almost uncontrollable predicate
calculus theorem provers that have been available.

	Putting information that humans represent as facts in the
form of program has led to excessively specialized systems.
Thus it has always seemed to me that the solution would
eventually involve theorem proving problem solvers that
were controlled by referring to declarative meta-information.
Others have independently come to this conclusion but its
realization has been difficult.
One of the most serious current attempt along these lines is Michael
Genesereth's MRS.

	While I have described this rationalist approach in
terms of my own work, many of the concepts in similar
or variant form have been developed independently by other
people.

	Akin to this viewpoint is Allen Newell's (1980) notion
of the "logic level" and the ideas of my "Ascribing Mental
Qualities to Machines" (1980) and Daniel Dennett's "intensional stance".
These notions all take the view that beliefs and goals may
legitimately be ascribed to physical systems independently
of whether sentences in some language are explicitly represented.
The ascription is legitimate when certain minimal properties
of the concepts are realized and useful when it helps understand
interesting aspects of the structure, state or behavior of the system.
These ascriptions are piecemeal, and do not require anything
like the full set of properties of the human mind.  While
all of us believe that eventually the full set of human
mental properties will be understood and realized in
computer programs, no-one currently claims to understand
what they all are.

	Now I shall return to the reasoning program approach.
It has been modified in various ways.  The most important
modification is the addition of non-monotonic reasoning since
the late 1970s.  In some sense non-monotonic reasoning was
already anticipated in the 1958 proposal, because the ability
to observe its own consciousness could generate sentences
that sentences of a certain kind did not exist in consciousness,
and such sentences could be realized deductively.  However,
my attempts to work this out at that time merely confused me.

	Non-monotonic reasoning has been proposed both at the
program level and at the logical level.  At the program level
we have Jon Doyle's TMS and the more recent ATMS of Johann de Kleer.
Less systematic approaches to default reasoning have been included
in many programs.
At the logic level we have my circumscription, the McDermott
and Doyle non-monotonic logic and Reiter's logic of defaults.

	Non-monotonic reasoning is important in solving some of the
problems cited by the gang of three as reasons why logic can't be
successful.  It is also relevant to Carl Hewitt's and Marvin Minsky's
objections to logic as the basis of AI.
	
	It seems to me that these objections are based on an observation
and an intuition.  The observation is that the collections of facts
used as a basis of reasoning by all contemporary AI programs are too
specialized to the problem the particular program is solving.  Thus
they could not be incorporated as is in a general common sense database.
This is also true of the rules in the various semi-logical approaches
used in expert system shells.  More specifically there is the qualification
problem, e.g. the problem that specification of the conditions under
which an action (rowing a boat across a river) can be performed cannot
be made complete, because one can always invent some exotic condition
that would prevent the action from being successful.

	My 1980 and 1986 papers in Artificial Intelligence on circumscription
make proposals for relieving this problem by using the circumscription
method of non-monotonic reasoning.  Proposals for this and other
ways of incorporating non-monotonic reasoning into logic have been
in the literature since 1980.  However, none of the critics of AI
have bothered to comment on them.  Perhaps the logic, though not deep,
is too technical for them.

	My opinion is that more is required than the uses of non-monotonic
reasoning that have been developed so far.  To this end I am working
on a formalized notion of context.  The idea is to use wffs  holds(p,c)
where  p  is a proposition and  c  is a context, e.g. the context of
the Sherlock Holmes stories or the context of a particular real
conversation or piece of writing.  To some extent contexts correspond
to sets of assumptions, but a key idea is that the context is never
assumed to be fully known - like situations in situation calculus.
There are axioms relating different contexts, especially general
contexts and their specializations.  These axioms involve non-monotonic
assumptions, so that properties of contexts are assumed to be inherited
by generalizations and other related contexts unless there is information
to the contrary.

	More details have been developed, but I'm not ready to publish.
Moreover, I have found the vaguely stated conundrums proposed by Hubert Dreyfus
and by Winograd and Flores helpful in constructing examples of the
non-monotonic reasoning required.  In particular, Dreyfus's ancient
notion of "ambiguity tolerance" admits an interesting non-monotonic
formalization.

	Suppose there a law is passed against attempting to bribe a public
official, and a knowledge engineer has the responsibility of putting
it in a database to be used by an expert system that advises prosecutors
on when they should seek an indictment.  Must this expert anticipate the
following defense that may be offered twenty years hence.  "It is true
that my client offered this man $5000 to fix his drunk driving conviction
under the impression that he was the Commissioner of Motor Vehicles.
However, the Governor had never signed his commission and therefore
he wasn't a public official."  The issue is whether the individual
whom the defendant attempted to bribe was actually a public official
or is believing that he was a public official sufficient to make him
guilty of the attempt.  Our contention is that the knowledge engineer
can put the law in the database without anticipating this ambiguity
provided he uses an "ambiguity tolerant" formalization.  Ambiguity
tolerance is achieved in this case by the non-monotonic assumption
that the law is unambiguous in a particular case unless an ambiguity
is exhibited.

	A more ambitious treatment of a law that may contain unknown
ambiguities is Robert Kowalski's logical formalization of the British
Nationality Act.

	The Gang of Three may grumble that it is unfair to bring
unpublished work to bear on their argument against rational and
logical approaches to AI.  However, they are claiming to have proved
that logic won't work, but their proofs are merely appeals to intuition.
Therefore, it seems legitimate to bring up undeveloped approaches
as counterintuitions.

	Many of the issues are treated at greater length in my "Ascribing
Mental Qualities to Machines" published in Martin Ringle's Philosophical
Perspectives in Artificial Intelligence, Harvester Press, July 1979.
This publisher is so obscure that I have never met anyone who has
seen the book.  Therefore, I will be glad to supply a copy of the
paper to anyone who wants it.
	In one sense this gang of four are among the last holdouts
for a geocentric point of view.  Inteligence to them is a function
of humanity.