When faced with a complex problem, traditional artificial intelligence often attempts to build a solution from first principles—either by evaluating a massive tree of logical rules or optimizing a global statistical model. However, the real world is frequently too messy, incomplete, or computationally complex for perfect planning.
An alternative approach mimics human cognition: instead of starting from scratch, we recall a similar past experience and adapt it to the current situation. This paradigm is known as Case-Based Reasoning (CBR). By treating historical experiences as discrete "cases," systems can bypass the need for exhaustive rule-scripting or first-principles simulation, computing local modifications rather than searching an entire combinatorial space of solutions.
The Core Mechanism: The CBR Cycle
At its heart, CBR operates on a cyclical process of leveraging historical data to navigate novel scenarios. As outlined by D. K. Chaturvedi in a comprehensive survey of intelligent systems, this cycle typically consists of four iterative steps: Retrieve a similar past experience, Reuse its solution, Revise the solution to fit new constraints, and Retain the new experience to update the system's knowledge base (Survey of Expert Systems and the Cognitive Approaches towards an Effective Tutoring System).
[ Novel Problem ]
│
▼
┌───────────────────────────┐
│ RETRIEVE │ ◄─── Find closest historical case
└─────────────┬─────────────┘
│
▼
┌───────────────────────────┐
│ REUSE │ ◄─── Apply past solution structure
└─────────────┬─────────────┘
│
▼
┌───────────────────────────┐
│ REVISE │ ◄─── Tweak parameters for new constraints
└─────────────┬─────────────┘
│
▼
┌───────────────────────────┐
│ RETAIN │ ◄─── Save new experience to Case Base
└───────────────────────────┘
This iterative cycle allows systems to handle highly dynamic or poorly defined environments. For instance, Yiding Dou et al. demonstrate how retrieving past disaster cases and revising their parameters allows for real-time multi-hazard impact assessments when ground-truth data is missing (A Case-Based Reasoning Framework Augmented with Causal Graph Bayesian Networks for Multi-Hazard Assessment of Earthquake Impacts).
Bypassing First-Principles Bottlenecks
Why prefer retrieval and adaptation over direct calculation? In many complex domains, the search space is too vast or the underlying causal models are unknown.
Early foundational work by Rajendra S. Wall et al. showed that generating sets of possible future scenarios by retrieving and adapting past tactical patterns reduces computational complexity in automated planning (Scenarios as an Aid to Planning). Similarly, Edwina L. Rissland et al. introduced Constrained Example Generation to demonstrate that an AI can focus on learning the appropriateness of specific modification strategies rather than mastering the entire domain's rules from scratch (Constrained Example Generation: A Testbed for Studying Issues in Learning).
This advantage becomes clear in highly technical domains:
- Scientific Discovery: In protein crystallization—a complex 25-parameter search space—Vanathi Gopalakrishnan et al. utilized a repository of historical experiments to suggest optimal parameters, bypassing the need for a complete physical model of the domain (The Crystallographer's Assistant).
- Workflow Composition: In software engineering, Rui-XING CHENG et al. applied this to web service composition, modifying existing workflows rather than synthesizing new ones from scratch (Using Case-Based Reasoning to Support Web Service Composition).
Problem Interpretation and Contextual Inference
CBR does more than just copy solutions; it helps an agent interpret ambiguous or incomplete inputs. By mapping a new situation to a known historical case, a system can infer missing details and build a more complete understanding of the problem's constraints.
Janet Kolodner et al. developed an early process model showing how an agent can use historical data to fill in missing information and recover from errors during complex dispute resolutions (A Process Model of Cased-Based Reasoning in Problem Solving).
This capacity for contextual inference is vital in pervasive environments. For example, Hong-Bo Ni et al. used CBR to infer a user's current task in smart homes by comparing real-time sensor data to stored historical cases, avoiding the need for rigid, pre-programmed rules (CDTOM: A Context-driven Task-oriented Middleware for Pervasive Homecare Environment).
Strategy Selection and Adaptation
Once a case is retrieved, the system must adapt the historical outcome to the current context. This can range from adjusting numerical parameters to selecting high-level strategic interventions.
In adversarial planning and negotiation, Katia P. Sycara demonstrated that initial solution candidates could be generated from a memory of past successes and failures, then refined through heuristic adaptation to resolve goal conflicts (Resolving Goal Conflicts via Negotiation).
In interactive entertainment, Anders Drachen et al. proposed using CBR to manage narratives dynamically by retrieving and adapting strategies from a library of expert human performances, allowing the AI to handle unexpected player behaviors without manual rule-scripting (Towards Data-Driven Drama Management: Issues in Data Collection and Annotation).
Furthermore, the authors of Towards Situated Explanation, Raja Sooriamurthi et al., applied this to explanation generation, showing that case-based retrieval provides a computationally efficient method for generating initial hypotheses by packaging successful information-seeking strategies from past contexts. For behavioral control, Michael W. Floyd et al. used CBR to let autonomous systems bypass incremental adjustments and immediately adopt behavior configurations that resolved user dissatisfaction in similar past contexts (Trust-Guided Behavior Adaptation Using Case-Based Reasoning).
Real-World Deployment and Decision Support
Because CBR relies on concrete examples rather than abstract mathematical abstractions, it is highly effective for human-in-the-loop decision support and industrial automation.
In customer support, William Cheetham detailed a conversational CBR system at General Electric that allowed call center agents to dynamically diagnose appliance issues by retrieving similar past cases, bypassing rigid if-then logic (Case-Based Reasoning for General Electric Appliance Customer Support). In food safety, Yongmei Li et al. used CBR to automate corrective actions by mapping current safety deviations to a library of expert-approved historical interventions (Automated Management of Computer Management Procedure of Critical Control Point Information on HACCP System).
The paradigm also scales to modern data-driven challenges. Xiao‐Yun Tian showed that prioritizing analogical similarity over global statistical optimization makes systems robust for identifying rare patterns, such as enterprise financial risk (Internet Big Data Information Analysis and Power Intelligent Automation Risk Prediction Based on Case Based Reasoning). Even in modern multi-agent code generation, Md. Ashraful Islam et al. demonstrate that retrieving and analyzing similar problem-solving patterns grounds large language models, aligning their logic with proven algorithmic structures before implementation (CodeSim: Multi-Agent Code Generation and Problem Solving through Simulation-Driven Planning and Debugging).
Key Foundational Work
Constrained Example Generation: A Testbed for Studying Issues in Learning (1981) — contributed 1 concept elements to this direction.
Scenarios as an Aid to Planning (1982) — contributed 1 concept elements to this direction.