Some early case-based reasoning systems included components for learning limited forms of adaptation knowledge. For example, CHEF [Hammond1989] bases its adaptations on both a static library of domain-independent plan repair strategies and a library of special-purpose ingredient critics that are learned; PERSUADER [Sycara1988] uses previously-stored adaptation episodes to suggest adaptations. In both examples, the learned adaptations can only be reused in highly similar situations. However, the adaptation cases learned by DIAL can be reused more flexibly. Rather than learning only by storing specific adaptation episodes, DIAL stores both the specific adaptation and its derivational trace. In very similar situations, the adaptation can be reapplied directly; in less similar situations, the steps used to determine the previous adaptation can be replayed, taking into account differing circumstances. In reasoning about the information needed to carry out the adaptation task, it also relates closely to Oehlmann's [1995] metacognitive adaptation.
Smyth & Keane have developed a CBR system that ties similarity judgments directly to adaptability, using heuristics coded to recognize the difficulty of performing particular types of adaptations. They demonstrate that adaptation-guided methods produce significant improvements in the cost of performing adaptations. Our approach to similarity judgments is strongly in the spirit of their approach. However, in their work, similarity and adaptation knowledge are static. As our method learns new adaptations, it derives similarity criteria directly from its own experience with adaptation problems, changing both as it acquires adaptation experience.