MPhil Thesis Defence


"Learning Domain-Specific Control Information in Planner-R to Improve its
Performance"

By

Mr. Wai-Kay Lau


Abstract

Planner-R is a planner written by Fangzhen Lin, which was competed in the
AIPS2000 planning competition in both the automatic and hand-tailored
tracks. It was selected as one of the Distinguished Planners in the
competition.

Planner-R works very well in some well-known benchmark domains, like the
blocks-world domain, in the sense that it can generate a plan of
reasonable length in relative short period of time. However, it does not
work well in some other domains, like the logistics domain.

To improve the performance of Planner-R, we explore in this thesis the
possibility of having Planner-R to automatically learn some
domain-specific control information. Specifically, we consider the problem
of learning an ordering on action types of a domain that is best suited
for Planner-R, and that of learning a set of rules to guide the planner in
selecting the best action to achieve a simple goal in the domain. We shall
present our learning algorithms for solving these two problems, and show
that with the information learned, Planner-R indeed will perform better,
both in terms of the running time and the quality of plans generated.


Date:			Thursday, 2 August 2001

Time:			10:00a.m.-12:00noon

Venue:			Room 4333
			Lift 3

Committee Members:	Dr. Fangzhen Lin (Supervisor)
			Dr. James Kwok (Chairman)
			Dr. Helen Shen


**** ALL are Welcome ****