Case-Based Reasoning

Origin. Roger Schank and Janet Kolodner at Yale; extensively developed and applied in Japanese AI research for expert systems, design support, and diagnosis. Japanese work emphasized case libraries and similarity metrics for engineering domains.

Mechanism. Solve new problems by retrieving and adapting solutions from similar past cases. The case base is the knowledge base: each case is a problem-solution pair with features. When a new problem arrives, retrieve the most similar cases, adapt their solutions to fit the new problem, evaluate and revise as needed, and retain the new case for future use. Learning is case accumulation; expertise is a large library of relevant cases.

Procedure. Define the case representation: what features characterize problems and solutions? Build the initial case base from historical examples or expert elicitation. For each new problem: (1) Retrieve — find cases with similar features using similarity metrics; (2) Reuse — apply the retrieved solution, adapting for differences; (3) Revise — evaluate the outcome and correct if needed; (4) Retain — store the new problem-solution pair as a case. The system improves as the case base grows.

Applies to. Design support, diagnosis, help desks, legal reasoning, any domain where past examples are a valid basis for current decisions.

Limitations. The similarity metric determines everything; a poor metric retrieves irrelevant cases. Adaptation is often the weak link — knowing a similar case does not provide the transformation rules for the new case. Case bases grow indefinitely and become slow without case management. Also: case-based reasoning cannot extrapolate beyond its cases; novel situations with no precedent receive poor or no solutions.