Is case-based reasoning the same as nearest-neighbor classification?

They can share similarity methods, but case-based reasoning usually emphasizes solution adaptation, checking and learning from experience.

Does it need machine learning?

Not necessarily. Retrieval and adaptation can use hand-designed representations and rules.

Can failed cases be useful?

Yes. A well-documented failure can show what to avoid and under which conditions an approach did not work.

AI Glossary / AI and machine learning / Intermediate

Case-based reasoning

Definition

Case-based reasoning is an approach to problem solving that uses relevant past cases to help address a new situation. A system retrieves earlier experience, adapts a proposed solution, checks it and may retain the new experience for future use.

What it means

People often begin a difficult problem by asking whether they have seen something similar before. Case-based reasoning makes that process explicit. A case can include the original situation, the action taken, the outcome and contextual details explaining why the action worked or failed.

The objective is not merely to retrieve a document containing similar words. The retrieved experience must help form and evaluate a solution. A repair case involving the same symptom may be misleading if the equipment model, operating conditions or underlying fault differs.

Case-based systems range from relatively simple retrieval tools to systems with substantial adaptation knowledge. Their usefulness depends on the quality of the case base, the definition of relevant similarity and the ability to test a proposed solution. Past success provides evidence to examine, not a guarantee of future success.

How it works

A widely used description organizes the process into four activities: retrieve, reuse, revise and retain. Retrieval identifies potentially relevant cases using indexed attributes, similarity measures or other search methods.

Reuse transfers useful parts of an earlier solution to the new problem. This may require changing quantities, replacing unavailable components or accounting for a different constraint. Revision checks the proposed solution through analysis, simulation, expert review or appropriate real-world feedback and modifies it when necessary.

Retention stores useful new experience, including the actual outcome and conditions. Not every interaction deserves permanent inclusion. Duplicate, outdated or poorly documented cases can make future retrieval worse. The process is therefore a learning cycle with case maintenance, not a rule that every old answer should be copied into every new situation.

Three practical examples

1. A maintenance assistant retrieves earlier equipment faults with similar symptoms and highlights differences in model and operating environment before suggesting diagnostic steps.

2. A customer-support team reuses a resolved service case while checking that the product version and policy are still applicable.

3. A design team examines previous packaging solutions with similar dimensions and constraints, adapts a candidate and tests whether it satisfies the new shipping requirements.

Think of it this way

Think of an experienced craftsperson's notebook containing jobs, decisions and outcomes. A similar page can provide a starting point, but the craftsperson still checks what differs today. The limit is that a software system only benefits from distinctions that its representation and reasoning process can recognize.

Worked example

Suppose a toy retrieval score compares three attributes: equipment model, symptom and operating condition. Their weights are 0.5, 0.3 and 0.2. Case A matches the first two but not the third, so its score is 0.5 + 0.3 = 0.8. Case B matches only the symptom and condition, scoring 0.3 + 0.2 = 0.5.

Case A ranks higher under this chosen score. That does not mean it has an 80% chance of solving the problem. The number is a similarity score, and the weights are design choices requiring evaluation.

If the operating condition is critical to safety or feasibility, the system may need a hard constraint rather than a small similarity weight. Retrieval ranking and solution validity are different questions.

Illustrative case study

A fictional commercial-kitchen service team builds a case base from completed repairs. A retrieved case suggests replacing a sensor, but a technician notices that the new equipment uses a different control board.

The team adapts the diagnostic procedure, tests the proposed explanation and records the actual fault and repair result. It also updates the earlier case's applicability notes so future users can see the version difference.

The case illustrates the full cycle: experience starts the investigation, adaptation accounts for differences, checking determines whether the proposal works, and retention improves the record. Simply returning the old repair note would have completed only part of that process.

Common misconceptions and mistakes

The closest case must contain the correct solution. Similarity can overlook a decisive difference.
A similarity score is a calibrated probability. It often only ranks candidates under a chosen metric.
A larger case base is always better. Redundant, obsolete or unreliable cases can reduce usefulness.

Business use

Case-based reasoning can preserve practical organisational knowledge where examples and outcomes are more available than comprehensive rules. It is especially useful when staff can explain how a previous solution should be adapted.

Define what each case must record, who validates outcomes and when records become obsolete. Evaluate retrieval relevance separately from the success of adapted solutions. Provide clear access to the original evidence and important differences. A case base becomes a business asset through maintenance and feedback, not simply through accumulating a large archive.

Check your understanding

A system retrieves an old case and copies its answer without checking differences or outcomes. Which parts of a full case-based reasoning cycle are weak or missing?

Answer: Meaningful adaptation and revision are missing, and retention is incomplete if the new outcome is not evaluated and recorded.

Related terms

Similarity Search
Knowledge base
Nearest neighbor
Expert system
retrieval
adaptation
organisational learning

Learn next

Read similarity search and knowledge representation. Take a resolved support case and identify the facts that must match, the details that may be adapted and the evidence needed to confirm success.

Sources

Aamodt and Plaza, Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches (1994): https://journals.sagepub.com/doi/10.3233/AIC-1994-7104. Author paper copy: https://oak.conncoll.edu/parker/com316/papers/CBR.pdf. The retrieval weights, scores and repair business are fictional teaching examples.

Previous: Case-based planning Next: Case-control study