COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	common[s82,jmc]		What is common sense?
C00008 00003	What is common sense?
C00010 ENDMK
C⊗;
common[s82,jmc]		What is common sense?

EXPERT SYSTEMS AND COMMON SENSE

or

CHALLENGING THE EXPERT SYSTEMS

	I'm no an expert on expert systems.  Rather I have been concerned
since 1955, in making computer programs with common sense.  This research
involves trying to express what everyone knows about the world as sentences
of mathematical logic and trying to express the methods of reasoning used
to draw common sense conclusions by (1) further sentences of logic, (2) by
the rules of inference of logic and (3) since the middle 70s by "non
monotonic" rules of conjecture.

	Progress has been slow.  Few have been attracted to the work
and almost no-one has stayed with it.  Maybe we're on the wrong track,
but many who were driven off by the difficulties of expressing common
sense facts in logical languages have returned to the opinion that there
is no alternative.  If we look at the history of science, we shouldn't
be surprised or dismayed that 25 years of artificial intelligence research
hasn't decisively cracked the problem of representing common sense
knowledge.  It was 100 years from Mendel to the elucidation of the
genetic code for proteins, and the problem of the genetic encoding
of physical form and of behavior are still almost untouched.  Another
historical example is the 1960s revival of Wegener's 1910 theory of
continental drift.  I wouldn't have bothered with these historical
analogies, except that my remarks about the present state of
expert systems will be critical, and I don't want anyone to say "Even
McCarthy is ready to give up on AI".  Since artificial intelligence
research seems to be subject to ideologically motivated attack, and
much of the funding comes from people demanding short range payoffs,
there is a tendency to avoid mutual criticism and for some scientists
to say little (often accompanied by thinking little) about the limitations
of the methods currently being employed.  This reluctance to recognize
difficulties slows progress.

	The expert systems approach to artificial intelligence attempts
to encapsulate in computer programs the knowledge that an expert has of his field.
The approach was pioneered by Edward Feigenbaum's Heuristic Programming
Project at Stanford and began with the Dendral program for inferring
chemical structures and continuing with the Mycin program for advising
physicians on the diagnosis and treatment of bacterial infections of
the blood.  Mycin carries out an intereactive English dialog with the
user about symptoms, tests that may be performed, the results of these
tests, anti-biotics that may be used and the results of treatment.
Experiments have led to rating its performance equal to experts in bacterial
infections of the blood and superior to that of graduating medical
students or non-specialist practicing physicians.

	[The folowing is based on memory of papers and talks about Mycin
and must be checked before any version of this is distributed].

	Mycin does all its work without any common sense knowledge.
It doesn't know about doctors, hospitals, life, death, health or
sickness.  Its ontology consists of aspects of the particular case
it is advising abbut.  Thus there ii no way to tell it that two patients
are members of the same family and have identical symptoms.  There
would be no way to tell it that one of the patients died after the
recommended treatment and ask it what should be done with the other.
It has extremely limited ability to be tolod about events taking
place in time.

What is common sense?

	Our object in asking this question is to arrive at a core of
capabilities rather than make a list of what common sense is used to
do.  Thus, since blind people have common sense, we suppose that vision
is not part of the core.  We also believe that natural language is
not part of the core, although this may not be easy to show and
contradicts many people's beliefs.

Events occur in time.

Facts are true in situations.

There are temporal and non-temporal facts.

Actions affect the future.

Concurrent events and actions.

Objects have extent in space and distances from each other.

People have purposes and perform actions.

People have beliefs which are not always correct or complete.

Assertions admit exceptions.

Handling "here" "now" and "me".