Case-Based Reasoning (CBR)
- Introduced by
- Janet Kolodner; Agnar Aamodt & Enric Plaza (R4 cycle)
- Year
- 1994
- Type
- Experience-based (analogical) problem solving
- Family
- Case-based reasoning
What does it do?
Case-based reasoning solves a new problem by retrieving similar problems solved in the past and adapting their solutions, rather than reasoning from first principles or a trained statistical model.
Formalized as the Retrieve-Reuse-Revise-Retain cycle by Aamodt and Plaza in 1994 and popularized by Janet Kolodner, CBR mirrors how human experts in medicine, law, and engineering reason by analogy from remembered cases, and it learns simply by storing each newly solved case.
The intuition behind the method
A doctor faced with a difficult patient does not re-derive medicine from scratch; she recalls similar patients and what worked for them, then adjusts for this patient's differences. CBR encodes exactly this: it keeps a library of past cases (problem plus its solution), finds the most similar past case to the new problem, reuses that solution with adaptations, checks whether it worked, and files the new experience away for next time. Knowledge is therefore concrete and example-based, not a set of abstract rules, which makes CBR natural for domains where experience matters more than theory.
History
Case-based reasoning grew from Roger Schank's work on dynamic memory and scripts in the early 1980s and was developed into a practical methodology by Janet Kolodner, whose 1992 introduction and subsequent textbook shaped the field. Aamodt and Plaza's 1994 paper provided the unifying Retrieve-Reuse-Revise-Retain (4R) cycle that remains the standard description. CBR became a recognized branch of knowledge-based and analogical AI, with enduring applications in diagnosis, law, design, and recommendation, and it connects to nearest-neighbour learning, rough sets, and instance-based methods.
When to use it
Use case-based reasoning when good solutions to past instances exist and are reusable, the domain is hard to capture in explicit rules or a statistical model, and explanations grounded in concrete precedents are valued — classic fits include medical diagnosis, legal reasoning (precedent), help-desk and technical support, design, and configuration. It is attractive when data accrue incrementally (each solved case is immediately useful) and when transparency matters, since a recommendation can be justified by the analogous case it came from. It depends critically on a meaningful similarity measure and on adaptation knowledge; with poor similarity metrics, no adaptation, or a sparse/biased case base it degrades, and for large smooth datasets a statistical learner may generalize better.
Limitations and considerations
- Quality hinges on the similarity measure and on adaptation knowledge, which can be hard to engineer.
- A biased or sparse case base yields poor or skewed recommendations.
- Retrieval can become slow as the case base grows without good indexing.
- May generalize worse than statistical models on large, smooth datasets.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Learns incrementally by simply retaining solved cases — no batch retraining.
- Transparent and explainable: recommendations are justified by concrete analogous cases.
- Works in domains that resist explicit rules or parametric models.
- Naturally handles a growing, evolving body of experience.
Applications
- Medical diagnosis and treatment recommendation from similar past patients.
- Legal reasoning and precedent retrieval.
- Technical support, help desks, and fault diagnosis.
- Engineering design and product configuration by analogy.
- Recommendation and planning where past solutions can be adapted.
How it works
How it works
Given a new problem (query), search the case base for the most similar stored case(s) using a similarity measure over the problem's features — often a weighted distance or nearest-neighbour metric, sometimes with indexing for speed. The retrieved case carries a solution that worked in an analogous situation.
Formula
c^* = \arg\max_{c \in \text{CaseBase}} \ \text{sim}(q, c)Adapt the retrieved solution to fit the new problem. Adaptation ranges from null (copy the solution as-is) to substitution of parameters, to transformational adaptation guided by domain rules. This step turns 'what worked before' into a candidate solution for 'now'.
Formula
\hat{s} = \text{adapt}\big(\text{solution}(c^*),\ q\big)Evaluate the candidate solution — by simulation, expert review, or real-world outcome — and repair it if it fails. Revision is where CBR confronts the fact that the past case was only similar, not identical, and corrects the proposed solution accordingly.
Formula
s = \text{repair}(\hat{s})\ \text{after evaluating outcome}Store the new problem together with its confirmed solution (and lessons learned) back into the case base. This is CBR's learning mechanism: the system improves incrementally as it accumulates experience, with no separate training phase.
Formula
\text{CaseBase} \leftarrow \text{CaseBase} \cup \{(q, s)\}Common pitfalls
- Treating retrieval (nearest case) as the whole method and skipping adaptation and revision.
- Using an unweighted or naive similarity measure that ignores which features actually matter.
- Letting the case base grow unmaintained, accumulating redundant or obsolete cases.
- Retaining unverified solutions, which propagates errors into future reasoning.
Frequently asked questions
How is CBR different from k-nearest neighbours?
Nearest-neighbour classification stops at retrieval — it predicts from similar instances. CBR goes further: it adapts the retrieved solution to the new problem (reuse), evaluates and repairs it (revise), and stores the result (retain). So kNN is essentially CBR's retrieval step without adaptation, revision, or explicit learning by case retention.
Does CBR need a training phase?
No. CBR has no separate training step; it reasons directly from stored cases and learns by retaining each newly solved case. This makes it well suited to domains where knowledge accrues incrementally and where retraining a model would be impractical.
What makes a good similarity measure for CBR?
One that reflects which problem features actually determine the right solution, usually a weighted combination of feature distances with weights set by domain knowledge or learned from data. A poor or unweighted similarity measure is the most common cause of weak CBR performance.
Related methods
Sources
Aamodt, A., & Plaza, E. (1994). Case-based reasoning: Foundational issues, methodological variations, and system approaches. AI Communications, 7(1), 39–59.
Open DOI referenceCopied.Kolodner, J. L. (1992). An introduction to case-based reasoning. Artificial Intelligence Review, 6(1), 3–34.
Open DOI referenceCopied.