ylfink@water.waterloo.edu (ylfink) (03/05/88)
DEPARTMENT OF COMPUTER SCIENCE
UNIVERSITY OF WATERLOO
SEMINAR ACTIVITIES
ARTIFICIAL INTELLIGENCE SEMINAR
- Friday, March 11, 1988
Dr. Henry Kautz, from AT&T Bell Labs, Murray Hill, New
Jersey, will speak on ``A Logical Reconstruction of
Plan Recognition''.
TIME: 1:30 PM
ROOM: MC 5158A
ABSTRACT
Algorithms for plan recognition have usually been based
on matching a sequence of observed events into a
hierarchically structured ``plan library'', which
specifies the abstraction and decomposition (substep)
relations between event types. This talk describes a
logical reconstruction of this kind of taxonomic
recognition. The result is a general competence theory
of plan recognition, whose semantic basis can be used
to justify particular recognition algorithms.
We show that the logical representation of the plan
library must be strengthened by a number of assumptions
for use in recognition. The assumptions are that event
types are disjoint, unless known otherwise; that the
decomposition hierarchy is complete; and that the
agent's actions are, if possible, all part of the same
plan. These assumptions are developed through the
construction of a certain class of minimal models of
the plan library. Circumscription provides a general
non-constructive method for specifying a class of
minimal models. For the specific case at hand,
however, we can mechanically generate a set of first-
order axioms which precisely capture the assumptions.
The resulting theory correctly handles such difficult
matters as disjunctive information, shared steps, and
concurrent unrelated plans.