Partial-Order Planning: Evaluating Possible Efficiency Gains

Although most people believe that planners that delay step-ordering decisions as long as possible are more e cient than those that manipulate totally ordered sequences of actions, this intuition has received little formal justi cation or empirical validation. In this paper we do both, characterizing the types of domains that o er performance di erentiation and the features that distinguish the relative overhead of three planning algorithms. As expected, the partial-order (nonlinear) planner often has an advantage when confronted with problems in which the speci c order of the plan steps is critical. We argue that the observed performance di erences are best understood with an extension of Korf's taxonomy of subgoal collections. Each planner quickly solved problems whose subgoals were independent or trivially serializable, but problems with laboriously serializable or nonserializable subgoals were intractable for all planners. Since di erent plan representations induce distinct search spaces, the subgoals for a given problem may be trivially serializable for one planner, laboriously serializable for another, and nonserializable for a third. We contend that the partial-order representation yields superior performance because it more frequently results in trivial serializability. This research was greatly improved by suggestions from Paul Cohen, Ernest Davis, Oren Etzioni, Steve Hanks, James Hendler, Craig Knoblock, Rao Kambhampati, David McAllester, Steve Minton, Edwin Pednault, Ying Sun, Josh Tenenberg, Brian Williams, Mike Williamson, and two anonymous referees. We thank Steven Soderland, with whom we started this project. This research was funded in part by National Science Foundation Grant IRI-8957302, O ce of Naval Research Grant 90-J-1904, and the Xerox corporation.