In this paper we describe the design and implementation of the derivation replay framework, DERSNLP+EBL (Derivational SNLP+EBL), which is based within a partial order planner. DERSNLP+EBL replays previous plan derivations by first repeating its earlier decisions in the context of the new problem situation, then extending the replayed path to obtain a complete solution for the new problem. When the replayed path cannot be extended into a new solution, explanation-based learning (EBL) techniques are employed to identify the features of the new problem which prevent this extension. These features are then added as censors on the retrieval of the stored case. To keep retrieval costs low, DERSNLP+EBL normally stores plan derivations for individual goals, and replays one or more of these derivations in solving multi-goal problems. Cases covering multiple goals are stored only when subplans for individual goals cannot be successfully merged. The aim in constructing the case library is to predict these goal interactions and to store a multi-goal case for each set of negatively interacting goals. We provide empirical results demonstrating the effectiveness of DERSNLP+EBL in improving planning performance on randomly-generated problems drawn from a complex domain.
Registration: ISBN 978-0-262-51091-2