The Case for Graph-Structured Representations

Case-based reasoning involves reasoning from cases: speciic pieces of experience, the reasoner's or another's, that can be used to solve problems. We use the term \graph-structured" for representations that (1) are capable of expressing the relations between any two objects in a case, (2) allow the set of relations used to vary from case to case, and (3) allow the set of possible relations to be expanded as necessary to describe new cases. Such representations can be implemented as, for example, semantic networks or lists of concrete propositions in some logic. We believe that graph-structured representations ooer signiicant advantages , and thus we are investigating ways to implement such representations eeciently. We make a \case-based argument" using examples from two systems, chiron and caper, to show how a graph-structured representation supports two diierent kinds of case-based planning in two diierent domains. We discuss the costs associated with graph-structured representations and describe an approach to reducing those costs, implemented in caper.