MVILAIN@G.BBN.COM (Marc Vilain) (09/30/87)
BBN Science Development Program
AI/Education Seminar Series
Dominance and Subsumption in Constraint-Posting Planning
Michael Wellman
MIT Artificial Intelligence Lab
(MPW@ZERMATT.LCS.MIT.EDU)
BBN Laboratories Inc.
10 Moulton Street
Large Conference Room, 2nd Floor
10:30 a.m., Friday, October 2nd 1987
Abstract: By integrating a dominance prover into the plan search
process, the traditional constraint-posting planning paradigm can be
generalized to permit partially satisfiable goals. In this approach,
the view of planning as theorem proving is retained, although the
emphasis is on deriving facts about the admissibility of classes of
candidate plans. Plan classes are generated by posting constraints at
various levels of abstraction, then classified within a plan
specialization graph that manages inheritance of properties and
dominance characteristics. Efficient computation of plan class
subsumption is essential for effective use of dominance results.
I illustrate this planning framework with examples from SUDO-Planner, an
application to medical therapy currently under implementation. Medical
therapy has been an unattractive domain for AI planning techniques
because of the omnipresence of uncertainty and partially satisfiable
objectives. SUDO-Planner's knowledge base contains descriptions of
therapy actions at multiple levels of abstraction, with effects
represented by qualitative probabilistic influences. The nature of the
dominance results derivable by SUDO-Planner suggest that many
"metaplanning" rules may be recast as dominance conditions at
sufficiently high levels of abstraction.
-------