An Explanation-Based Approach to Improve Retrieval in Case-Based Planning

When a case-based planner is retrieving a previous case in preparation for solving a new similar problem, it is often not aware of the implicit features of the new problem situation which determine if a particular case may be successfully applied. This means that some cases may be retrieved in error in that the case may fail to improve the planner’s performance. Retrieval may be incrementally improved by detecting and explaining these failures as they occur. In this paper we provide a definition of case failure for the planner, dersnlp (derivation replay in snlp), which solves new problems by replaying its previous plan derivations. We provide EBL (explanation-based learning) techniques for detecting and constructing the reasons for the failure. We also describe how to organize a case library so as to incorporate this failure information as it is produced. Finally we present an empirical study which demonstrates the effectiveness of this approach in improving the performance of dersnlp.