Case-Based Reasoning

Case-Based Reasoning (CBR) solves a new problem by reusing the solution to a similar problem from the past. Instead of encoding knowledge as general rules, it stores it as cases: concrete records holding a problem description, the solution applied to it and, where known, the outcome. Faced with a new query, the system searches its case base for the closest matches and builds an answer from them.

The reference model was set out by Agnar Aamodt and Enric Plaza in 1994. Their definition organises the process into four linked phases and, along the way, turns CBR into a form of learning: every problem solved can return to the base as one more case.

The four-R cycle

The core of CBR is a four-step cycle, the four Rs. Retrieve: starting from the new problem description, find the most similar case or cases in the base. Reuse: take the solution of the retrieved case and carry it over to the current problem. Revise: test that proposed solution, in the real world, with an expert or in simulation, and repair it if it fails. Retain: fold the useful experience back into the base as a new case, with its indexes, so it serves future problems. All four phases are needed: without retrieval there is no starting point, without retention there is no learning.

Similarity and adaptation

Two decisions hold the whole cycle together. The first is how to measure resemblance. Retrieval usually relies on nearest-neighbour (k-NN) methods: each case is described by attributes and compared with the new problem through a similarity function. Not every attribute counts equally, so each is given a weight; tuning those weights and refining the metric is much of the work, and it is domain-specific. The second decision is adaptation: the retrieved solution rarely fits as is. Aamodt and Plaza distinguish two routes, transformational reuse, which alters the old solution with operators until it fits, and derivational reuse, which replays on the new case the method that produced it. Automatic adaptation remains the hardest and least generalisable part of CBR.

Limits: case-base quality and maintenance

A CBR system is only as good as its case base. Noisy or poorly described cases degrade retrieval; gaps in coverage leave the base unable to answer problems it has no precedent for; and a base that only grows becomes slow and redundant. Hence the maintenance problem: which cases to keep, merge or delete without losing competence. Smyth and Keane (1995) formalised this with the notions of coverage and reachability, deleting cases while preserving problem-solving power. There is no universal recipe: similarity, adaptation and deletion criteria are designed for each domain. Even so, the approach shines where prior experience is the key knowledge: technical support and help desks (systems such as Compaq's SMART or HOMER), medical diagnosis, and legal reasoning from precedent, like the HYPO system.