Ralph Barletta, William Mark
Proper indexing of cases is critically important to the functioning of a case-based reasoner. In real domains such as fault recovery, a body of domain knowledge exists that can be captured and brought to bear on the indexing problem-even though the knowledge is incomplete. Modified explanation-based learning techniques allow the use of the incomplete domain theory to justify the actions of a case with respect to the facts known when the case was originally executed. Demonstrably relevant facts are generalized to form primary indices for the case. Inconsistencies between the domain theory and the actual case can also be used to determine facts that are demonstrably irrelevant to the case. The remaining facts are treated as secondary indices, subject to refinement via similarity based inductive techniques.