Most industrial-strength planners employ hierarchical planning by
recursively decomposing tasks into subtasks. In spite of this success,
there have hardly been any learning systems applicable to hierarchical
planning. There are two opportunities for learning useful knowledge
for hierarchical planning: (1) learning task decompositions
themselves; and (2) learning control knowledge for selecting/rejecting
among various task decompositions. My work focuses on the former---in
particular, on learning goal-decomposition rules (d-rules) that
recursively decompose goals into subgoals under particular conditions.
I map the problem of learning d-rules to the problem of learning Horn
programs. Then, I describe theoretical results on learning special
cases of Horn programs. Next, I present two implementations for
learning d-rules loosely based on the theoretical results---one by
generalizing the example plans executed by the user, and the other by
solving user-given exercises with search. Finally, I demonstrate
the performance of these implementations on two planning domains.
Date: Wed., May 6; Time: 4:15-5:30PM; Place: Cordura 100
Return to seminar schedule.