Derivational Replay in an Universal Classical Planning Framework

The status of plan adaptation has been somewhat controversial in the AI-planning literature, due to a conflict between worst-case complexity analyses (which have led to pessimistic conclusions about the utility of plan adaptation) and empirical results (in which plan adaptation has performed significantly better than planning from scratch). This paper provides a step toward the resolution of this conflict, for the case of a prominent planadaptation technique called derivational replay. First, we provide a general model of derivational replay. Next, we use this model to show that derivational replay does not satisfy a key assumption in the previous complexity analyses, and thus those analyses do not apply to it. Finally, we show that the search space for a planner using derivational replay is no bigger than it would be without derivational replay, and that derivational replay can reduce the size of the search space by n exponential amount.