What's Hot in Case-Based Reasoning

Ashok K. Goel, Belén Díaz-AgudoAAAI 2017 · DOI

Abstract

Case-Based Reasoning (CBR) is undergoing a significant research renaissance, evolving beyond simple data retrieval toward sophisticated adaptation, explanation, and "lazy learning." This paper summarizes the current state of the field, highlighting how CBR is being integrated with deep learning and big data to solve complex problems in robotics, law, and creative design. By mimicking human cognition—solving new problems by remembering and adapting solutions from similar past experiences—CBR offers a unique alternative to traditional rule-based systems and black-box machine learning models.

Introduction

In the era of "big data" and deep learning, traditional rule-based expert systems often struggle with the "knowledge acquisition bottleneck"—the difficulty of extracting exhaustive rules from human experts. Case-Based Reasoning (CBR) offers a robust alternative by mimicking human cognition: solving new problems by retrieving and adapting solutions from similar past experiences.

However, the field faces modern challenges: How do we extract cases from unstructured data like text or images? How can we make AI decisions transparent and trustworthy? This paper, originally presented at AAAI-17, serves as a "state of the union" for CBR, identifying how the community is evolving to address these gaps and stay relevant alongside modern machine learning paradigms.

The CBR Methodology

CBR is not a single algorithm but a general methodology defined by the "4R" cycle:

  1. Retrieve: Find similar past experiences from memory.
  2. Reuse: Apply a past solution to the current context.
  3. Review: Evaluate the new solution.
  4. Retain: Store the new experience for future use.

Recent methodological shifts have expanded this cycle to include:

  • Hybridization: Combining CBR with deep learning, Bayesian networks, and commonsense reasoning.
  • Qualitative Reasoning: Moving away from purely numerical similarity scores toward qualitative measures (e.g., in robot soccer applications).
  • Adaptation Discovery: Using "big data" to automatically discover the rules needed to adapt an old case to a new problem, rather than hard-coding them.
  • Diagrammatic Reasoning: Developing techniques to retrieve and reuse information stored as images or drawings rather than just text or feature vectors.

Application Domains

As a survey and summary paper, this document references a wide variety of application domains where CBR is currently being deployed, rather than introducing a single new benchmark. Notable examples include:

  • IBM’s Watson: A high-profile example of memory-based reasoning in practice.
  • Robotics: Specifically humanoid robot soccer (RoboCup), where qualitative reasoning is essential.
  • Recommender Systems: Using past user preferences to guide new suggestions.
  • Professional Domains: Legal reasoning, medical diagnosis, and engineering design, where precedent and historical cases are foundational.

Key Insights and Theoretical Advantages

CBR vs. k-Nearest Neighbors (k-NN)

While often confused, CBR goes much further than k-NN. While k-NN is primarily used for classification, CBR encompasses adaptation and reasoning, allowing the system to modify a retrieved solution to fit a unique new situation.

The "Lazy Learning" Advantage

Unlike "eager" learners (like neural networks) that build a general model during training, CBR is a "lazy" learner. It defers processing until a specific problem arises, allowing it to learn incrementally from a very small set of examples without requiring massive retraining phases.

Explainability by Design

Because CBR relies on concrete past examples (cases), it is naturally suited for Explainable AI (XAI). It is often easier for a human to trust a decision if the AI can say, "I'm suggesting this because it worked in this specific, similar situation," rather than providing a weight in a hidden layer.

Computational Creativity

Recent research explores using CBR for computational creativity by "blending" elements from multiple different cases to generate entirely novel solutions, moving the field into the realm of generative design.

Practical Results and Tool Maturity

The paper highlights the practical dominance of CBR, citing that it has led to more practical applications than almost any other AI family, with the notable exceptions of expert systems and standard machine learning. The emergence of robust, domain-independent tools like jCOLIBRI and myCBR has lowered the barrier for engineers to build case-based systems. Furthermore, contrary to the idea that CBR is only for small data, recent work shows that CBR can actually help manage "big data" by using it to generate more accurate adaptation rules.

Limitations and Future Work

The authors identify several "grand challenges" that the field must still overcome:

  • Raw Data Acquisition: There is a pressing need for better ways to automatically turn unstructured data (raw text, multimedia, diagrams) into structured "cases."
  • Scaling to Massive Memories: As case bases grow to "big data" proportions, maintaining and searching them efficiently becomes a significant computational hurdle.
  • Cognitive Modeling: There is still a gap between how AI retrieves cases and how the human brain performs "remindings." Bridging this could lead to more intuitive AI.
  • Widespread Distribution: The authors call for better distribution and standardization of tools to make CBR technology as accessible as modern deep learning frameworks.

Summary updated

BibTeX
@inproceedings{goel2017what,
  author = {Ashok K. Goel and Belén Díaz-Agudo},
  title = {{What's Hot in Case-Based Reasoning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.1609/aaai.v31i1.10643},
}
Figure