Abstract
Much of the problem solving done by both novices and experts uses "case-based" reasoning, or reasoning by analogy to previous similar cases. We explore the ways in which case-based reasoning can help in problem solving. According to our model, transfer of knowledge between cases is guided largely by the problem solving process itself. Our model shows the interactions between problem solving processes and memory for experience. Our computer program, called the MEDIATOR, illustrates case-based reasoning in interpreting and resolving common sense disputes.
Introduction
In the mid-1980s, most artificial intelligence problem solvers functioned as "amnesiacs." They treated every new problem as a unique event, following the same exhaustive chain of reasoning steps even if they had solved a nearly identical problem moments before. Furthermore, these systems lacked the capacity to learn from their mistakes; a failure in logic would simply be repeated the next time the same conditions arose.
This work addresses the "inefficiency of uniqueness" by proposing that human-like expertise relies on a library of past cases. The challenge lies in formalizing how a machine can store and retrieve relevant past experiences from a massive memory, adapt those old solutions to fit new contexts, and learn from failures to avoid repeating them.
The Case-Based Reasoning Process
The authors present a three-step process model for Case-Based Reasoning (CBR) that integrates problem solving with memory management. This model moves away from traditional state-space search toward a retrieval-based architecture.
- Locate and Retrieve: Using "generalized episodes" (memory structures that hold commonalities between similar events), the system uses the current problem features as keys to "remind" itself of past cases.
- Evaluate: The system determines which of the recalled cases is the most applicable using an evaluation hierarchy, such as prioritizing cases with similar goal relationships.
- Transfer: Knowledge is mapped from the old case to the new one. This involves "plan instantiation," where roles are bound to new actors (e.g., replacing "USA" in a historical case with "Israel" in a current dispute).
As shown in the process model, CBR is inherently iterative. If a generated solution fails to confirm predictions, the system enters a failure-recovery cycle: explaining the results, assigning blame, and correcting the failure to update the knowledge base. This ensures that the system becomes more efficient and accurate over time.
Implementation: The MEDIATOR
The implementation of this model is MEDIATOR, a program designed to resolve common-sense and geopolitical disputes. MEDIATOR uses a memory organized by "MOPs" (Memory Organization Packets) to index cases by their features and their failures. The system was evaluated qualitatively using a case library of diverse disputes:
- Geopolitical Disputes: The Sinai dispute (Israel/Egypt), the Panama Canal dispute (US/Panama), and the Korean War.
- Common-sense Disputes: A "two sisters and an orange" quarrel and children fighting over a candy bar.
- Resource Disputes: Conflicts over mineral rights in sea beds and fishing rights.
Failure-Driven Learning in Action
A key insight of the MEDIATOR is its ability to use failures to prompt re-interpretation. For example, when attempting to resolve the Sinai dispute, MEDIATOR initially suggested "dividing the land equally," a strategy retrieved from its memory of the Korean War. When this was rejected, the system was reminded of the "orange dispute"—where a 50/50 split failed because the parties wanted different parts of the object (the fruit vs. the peel).
This analogy prompted MEDIATOR to realize that the underlying issue was a misunderstanding of goals: Israel desired the Sinai for security, while Egypt desired it for sovereignty. By then retrieving the Panama Canal dispute, MEDIATOR suggested a plan where Egypt received political control while Israel maintained military buffer zones—a solution mirroring the actual Camp David Accords.
Conclusion and Future Work
CBR demonstrates that reasoning is as much about interpretation as it is about calculation. By "filling in the blanks" based on past cases, AI can understand complex problems more holistically. However, several hurdles remain, including the development of more robust evaluation criteria for case feasibility and the ability to synthesize parts from multiple different cases into a single coherent solution. While moving away from search reduces some computational burdens, managing a richly indexed episodic memory introduces its own complexities as the library of experience grows.
Summary updated