Case-based reasoning (CBR) is a subfield of artificial intelligence that solves new problems by referencing a library of previously encountered situations, or 'cases'. Instead of deriving answers from abstract rules or statistical patterns, a CBR system retrieves the most similar past case, adapts its solution to fit the current problem, and then stores the new experience for future use. This methodology mirrors human problem-solving, where people often rely on memories of specific past events rather than first principles.
The approach gained prominence in the late 1980s and early 1990s, largely through the work of researchers such as Roger Schank and Janet Kolodner at Xerox PARC and other institutions. Schank's dynamic memory theory, which posited that human memory is organized around episodic experiences, provided a cognitive foundation for CBR. Kolodner's 1993 book, 'Case-Based Reasoning', became a standard reference, and the field's annual International Conference on Case-Based Reasoning (ICCBR) has been held since 1995.
The CBR Cycle
A standard CBR system operates through a four-step cycle, often abbreviated as the '4R' process: Retrieve, Reuse, Revise, and Retain.
- Retrieve: Given a new problem description, the system searches its case library for cases with similar features. Similarity is typically computed using weighted feature matching, where each attribute of a case (e.g., a patient's symptoms or a machine's error codes) is compared with the corresponding attribute of the new problem. Distance metrics like Euclidean distance or cosine similarity are common, and indexing schemes help narrow the search.
- Reuse: The retrieved case's solution is transferred to the new problem. In simple systems, this may be a direct copy. In more complex ones, the solution is adapted based on differences between the old and new contexts. For example, if a past case solved a mechanical failure with a specific repair, the system might adjust the repair procedure for a slightly different machine model.
- Revise: The proposed solution is evaluated, either through simulation or real-world application. If it fails, the system diagnoses the cause of the failure and modifies the solution accordingly. This step often involves human feedback or a domain-specific repair algorithm.
- Retain: The new problem and its confirmed solution are added to the case library as a new case. This learning step ensures the system improves with experience, growing more accurate and efficient over time.
Applications and Domains
CBR has been deployed in a wide range of practical domains. One of the earliest and most successful applications was in help-desk and customer support systems, such as the Compaq SMART system (1990s), which used CBR to diagnose computer hardware problems. In medicine, CBR systems like CASEY (1989) and later tools have supported diagnosis and treatment planning by matching patient records to historical cases. The legal domain has also used CBR to retrieve precedents, with systems like HYPO (1987) at Carnegie Mellon University helping lawyers find relevant court decisions.
Other notable applications include:
- Engineering design: Reusing past design solutions for new components or systems.
- Robotics: Enabling robots to adapt past action sequences to new environments.
- E-commerce: Recommending products based on similar past purchases.
- Fraud detection: Identifying suspicious transactions by comparing them to known fraud cases.
Relationship to Other AI Approaches
CBR is often contrasted with rule-based expert systems, which encode knowledge as explicit if-then rules. While rule-based systems require knowledge engineers to extract and formalize expertise, CBR systems can be built more directly from historical records, reducing the knowledge-acquisition bottleneck. However, CBR systems face challenges in defining effective similarity measures and in maintaining large case libraries, which can become unwieldy and slow.
With the rise of machine learning and deep learning, CBR has sometimes been viewed as an older paradigm. Yet it remains relevant in hybrid systems. For instance, modern large language models (LLMs) have been combined with CBR to ground responses in specific past examples, improving factual accuracy and explainability. Research in neural networks has also explored 'neural case-based reasoning', where case representations are learned as embeddings, allowing similarity to be computed in a learned feature space. This integration has been explored at institutions like MIT CSAIL and Stanford AI Lab.
Strengths and Limitations
CBR offers several advantages. It is transparent, as solutions can be traced back to specific prior cases, making it easier to explain decisions to users. It learns incrementally, requiring no retraining of a global model. It is well-suited to domains where knowledge is incomplete or where exceptions are common, as it can handle 'edge cases' by retrieving the closest match.
Its limitations include a reliance on the quality and coverage of the case library; if a problem is unlike any stored case, the system may produce poor results. Similarity assessment can be brittle, and adaptation remains a hard problem, often requiring hand-crafted rules. Performance can degrade as the library grows, necessitating case-retention policies and indexing strategies. As of the mid-2020s, CBR is less prominent than deep learning in mainstream AI research, but it continues to be used in specialized industrial and academic settings, particularly where explainability and data efficiency are critical.