Least common subsumer trees for plan retrieval

Abstract

This paper presents a new hierarchical case retrieval method called Least Common Subsumer Trees (LCS trees). LCS trees perform a hierarchical clustering of the cases in the case base by iteratively computing the least-common subsumer of pairs of cases. We show that LCS trees offer two main advantages: First, they can enhance the accuracy of the CBR system by capturing regularities in the case base that are not captured by the similarity measure. Second, they can reduce retrieval time by filtering the set of cases that need to be considered for retrieval. We present and evaluate LCS trees in the context of plan retrieval for plan recognition, and present procedures for both assessing similarity and computing the least common subsumer of plans using refinement operators.