Combining Analytical and Similarity-Based CBR

A technique is presented for combining analytical and similarity-based CBR. Analytical CBR permits transfer from a source problem to a target only if the \analytical criterion" is satissed, namely, that the justiication for the source solution also hold in the target. This strategy has the merit of preventing \unsound" analogies, but is infeasible when no complete, correct domain theory is known. When a partial theory is all that is available, the analytical criterion becomes merely a heuristic. Similarity-based CBR can be called upon to supplement the heuristic; a similarity metric chooses which source is best among those that pass the analytical criterion. The analytical criterion helps guide the metric toward relevant features to match on. Aside from improved analogical judgements, the combination of analytical and similarity-based CBR supports eecient case retrieval. The method of combination is described in the context of a partially implemented system for pronouncing surnames.