COMMENT ⊗   VALID 00005 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	nonmon[s85,jmc]		Non-monotonic reasoning for AAAS
C00003 00003	Human non-monotonic reasoning
C00007 00004	slides
C00008 00005	outline of talk
C00012 ENDMK
C⊗;
nonmon[s85,jmc]		Non-monotonic reasoning for AAAS

glue together
	1. uses of non-monotonic reasoning
	2. from intro to 1980 paper
	3. bird axioms
	4. blocks axioms
	5. references to Doyle, Reiter, Lifschitz

	6. relation to probabilistic and competition with fuzzy
at least we think it covers the more significant phenomena

	7. Shedding light on ordinary human argumentation

In this paper masculine pronouns embrace not only the feminine but
also the robotic.
Human non-monotonic reasoning

	The study of formalized non-monotonic reasoning has been motivated
primarily by artificial intelligence research.  However, it also suggests
new ways of thinking about human reasoning.  The following considerations
are tentative.

	We may distinguish between reasoning aimed at solving a problem,
e.g. deciding what to do, and reasoning aimed at convincing other people
of the truth of certain propositions or to take a certain action.  Our
considerations apply to both but somewhat differently.

	In either case the reasoner {\it takes certain facts into account}
and draws conclusions correct in the simplest or perhaps
in a ``standard'' model of this collection of facts.  The reasoner keeps
in mind that some of his beliefs may be untrue and also that there may
be additional relevant facts that he has failed to take into account.

	The old-fashioned way of trying to look at this is probabilistic.
It is supposed that the reasoner has assigned some probabilities to the
possibilities that he wants to ignore and that these probabilities are low
enough for the time being.  However, it is difficult to use this model
unless we can say what these other possibilities are.  The best we can
the probabilist can do is to assign a probability to the proposition
that the simplest model of these facts is inapplicable.  He still has
to identify what this model is.

	The inability of humans to agree may be illuminated by
considerations of non-monotonic reasoning.  In considering any question,
a human is limited in the number of facts he can take into account.
What facts a person uses is affected by non-logical and even
non-logical factors.  These include the following:

	. What one's friends have been saying.

	. What one has heard recently.

	. Wishful thinking.  

When two people argue, replies often begin ``Yes, but ...''.  This
is often the signal of a complaint that one's interloctor hasn't
taken some of one's favorite facts into account and is usually
a prefix to ignoring some of his.  It is often true that each party
to the dispute has reasoned correctly though non-monotonically, i.e.
his conclusions are true in the simplest models of his collection
of facts.
slides

definition of circ

birds

moving and painting blocks

truth maintenance

logic of defaults

references, especially Lifschitz

advice taker slogans
outline of talk

	Intellligence exists --- as exemplified by humans.  However, we
humans don't yet understand how intelligence can work.  It isn't merely
that we don't know the mechanisms actually used; we don't know any
mechanism that will give rise to intelligent behavior in as wide
a variety of situations as humans exhibit.  Indeed there are certain
rather narrow classes of situations in which we don't yet know how
to make intelligence work.

	This fact dominates the relation between research in artificial
intelligence and the research in natural intelligence via physiology and
psychology.  Physiology and psychology can provide clues as to the
way natural intelligence functions, and this can help design intelligent
computers.  AI can isolate difficult classes of problems for psychologists
to ponder.  It has also given rise to concepts of intellectual process
that have influenced psychologists to study information processing models
of intelligence.  Indeed one may say that AI research was the largest
single influence in the demise of behaviorism.

	However, there are barriers limiting the interactions among
physiology, psychology and AI.  While physiologists have determined
much about the structure and function of neurons and synapses, it
turns out that this information is of no direct use for psychology
and AI.  This is because of universality theorems that tell us that
almost any structure of neuron can serve as a component of any
information processing system.  (slide?)

	AI work has been less helpful to psychologist than it might,
because AI programs usually take advantage of the high speed and
large general purpose temporary storage of computers which far exceeds
that available to humans and animals.  In some applications this more
than makes up for our inability to program the biologically achieved
level of pattern matching and other aspects of intelligence.  Mostly
it doesn't.

	One area that has concerned AI research for thirty years
is the use of facts in ways that were not pre-planned.