COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	halper.abs[e85,jmc]	Draft abstract for Halpern conference on knowledge
C00006 ENDMK
C⊗;
halper.abs[e85,jmc]	Draft abstract for Halpern conference on knowledge

Non-monotonic reasoning may require reifying states of knowledge.

	The use of circumscription to solve the frame problem
is discussed in (McCarthy 1984) and? (Lifschitz?).  The formalism
given in that paper isn't quite right, and trying to fix its difficulties
has led us to some interesting phenomena, which we will discuss in
this paper.

	Here's the difficulty.  In the axiomatization of that paper,
it is considered abnormal for one object to be on another.  This results
in the desirable consequence that when abnormality is circumscribed
only those  on(x,y,s)  relations that can be inferred are considered
true.  Now suppose we consider the new situation  result(move(A, top(B)),s0).
If  A  and  B  inferred to be clear in the situtation  s0,  then we wish to
infer the action  move(A,top(B))  succeeds.  However, this makes
allows us to infer  on(A,B,result(move(A,top(B)),s0)),  which implies
an abnormality that can be avoided if the action is unsuccessful.
Therefore, circumscription of abnormality results in a disjunction,
whereas we prefer to conclude definitely that the action is successful.

	There are several approaches.

	We can minimize  on(x,y,S0)  with higher priority than
on(x,y,result(move(A,top(B)),S0)).  This is what we want to happen,
but it isn't immediately clear how to formulate the axioms generally
so that it will.

	It is tempting to give normality in an early situation
priority over normality in a later situation.  This will probably
work in planning problems, but it isn't generally correct.  Sometimes
we want to infer facts about the past from knowledge of the present.
In that case giving priority to early situations is wrong.  The
answer seems to be for the system to give priority to the situations it
knows about over those whose properties it must infer from them.
Our formalism must then take into account not merely what it
believes but also how the different assertions came to be believed, e.g.
which were inferred from which others.  We are then may be led
to reifying and formalizing some of the data taken into account
by ``reason maintenance systems''.

	For this purpose we are considering using a ``mental situation
calculus'' involving mental states and mental actions.  The full
paper will cover our results on this and other approaches.