COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.<<hannov.2[w86,jmc]	Abstract for Hannover Fair Lecture>>
C00007 ENDMK
C⊗;
.<<hannov.2[w86,jmc]	Abstract for Hannover Fair Lecture>>
.require "memo.pub[let,jmc]" source;
.cb Expert Systems and Common Sense

.cb "John McCarthy, Stanford University, Stanford, California"

	Most technological applications of artificial intelligence (AI)
involve rule-based expert systems.  These expert systems, prominently
developed in Edward Feigenbaum's Heuristic Programming Project at
Stanford University, represent the "rules of thumb" of experts in
a certain domain as pattern-action rules in a computer.  
Such programs compare the pattern parts of their rules against the
facts in their database about a problem.  A match gives them
values for the variables in the pattern part of the rule.  The variables
are then used in the action part of the rule in order to determine
what to do.

	An example from the domain of the Mycin expert system that
diagnoses bacterial infections might take the form (translated into
English): If the patient has a rash on a certain part of his body,
his white blood cell count is ⊗x, his temperature is ⊗y  and ⊗x
and ⊗y satisfy a certain condition, then diagnose him as having
a certain disease.

.<<example?>>
	Many useful tasks can be done by such collections of rules -
anywhere from tens of rules to several thousand.  However, much human
common sense knowledge and scientific knowledge cannot be readily
fitted into such a framework.  For example, although Mycin does
good job of diagnosis within its domain, it doesn't know that
bacteria are little organisms that grow and reproduce and can
emit toxins that are harmful.  It doesn't know that bacteria enter
the body in certain ways and that they can be killed.  All these
facts were used in the design of Mycin's rules, but they aren't
in the system itself.  In fact Mycin and many other expert systems
don't know about processes occurring in time like the growth and
reproduction of bacteria or the patient getting well or dying.  If
Mycin is told that the patient is dead, it replied "Unrecognized
response".  Nevertheless systems like Mycin are useful.

	There is another thread of AI research dating back to
a 1958 paper of mine called "Programs with Common Sense"
that involves the study of common sense knowledge and its expression
in the machine in logical form.  While some human knowledge seems
to correspond to pattern-action rules, other knowledge seems to
take a more general form.  A much studied class of common sense
knowledge involves the effects of events including actions.  For
example, we may ask what are the basic facts about the action
of using a boat to cross a body of water.  This is a process
taking time, it requires being in legal and physical possession
of an operative boat at the beginning point of the journey.  It
involves the elementary fact that the passengers must enter the
boat, and stay in it while someone competent to do so, propels
the boat across and lands it on the other side.

	The lecture will discuss what is general
common sense knowledge, what expert systems require its use in
addition to partern-action rules, and the problem of adding this
capabiity to expert systems.