Craig Boutilier
Department of Computer Science
University of British Columbia
Vancouver, BC, CANADA, V6T 1Z4
email: cebly@cs.ubc.ca
Moises Goldszmidt
Rockwell Science Center
444 High Street
Palo Alto, CA 94301, U.S.A.
email: moises@rpal.rockwell.com
Abstract
We examine a number of techniques for representing actions with
stochastic effects using Bayesian networks and influence diagrams.
We compare these techniques according to ease of specification and
size of the representation required for the complete specification
of the dynamics of a particular system, paying particular attention
the role of persistence relationships.
We precisely characterize two components of the
frame problem for Bayes nets and
stochastic actions, propose several
ways to deal with these problems, and compare
our solutions with Reiter's solution to the frame problem for the situation
calculus. The result is a set of techniques that permit
both ease of specification and compact representation of probabilistic
system dynamics that is of comparable size (and timbre) to
Reiter's representation (i.e., with no explicit frame axioms).
To appear, CSCSI-96, Toronto, May 1996
Return to List of Papers