Introduction
In the field of Case-Based Reasoning (CBR), representing "how-to" or procedural knowledge presents a unique challenge. Unlike simple factual cases, a procedural episode consists of a sequence of interrelated steps. As CBR systems scale, the method chosen to store and organize these episodes determines whether the system remains efficient or becomes overwhelmed by its own memory. This paper provides a formal and empirical evaluation of three primary organizational strategies, proposing the "multicase" approach as a superior solution for balancing storage efficiency and retrieval speed.
Organizational Strategies
To understand the trade-offs in procedural representation, the authors compare three distinct models:
- Individual Cases: The entire episode is stored as a single, monolithic unit. Solving a new problem involves retrieving a complete past sequence and adapting it.
- Microcases: Each step within an episode is treated as an independent case. Procedures are reconstructed step-by-step by searching a global pool of all known steps.
- Multicases: Episodes are segmented into steps, but these steps remain indexed according to their position within a shared procedural hierarchy. This structure partitions the search space, allowing the system to focus only on relevant options for the current step.
The Multicase Advantage
The core insight of the multicase approach is that the structure of the procedure itself serves as a powerful index. While "Individual Cases" suffer from massive redundancy—storing the same common steps repeatedly across different episodes—and "Microcases" suffer from a lack of context—forcing the system to search through every "Step 2" ever recorded—the Multicase approach provides the best of both worlds. By knowing the system is at a specific point in a procedure, it can ignore the vast majority of irrelevant cases in memory, maintaining the storage efficiency of micro-segmentation without the retrieval overhead.
Empirical Performance
The authors validated their theoretical models using the FLOABN system, which simulates device operations. The empirical data confirmed that as the number of processed episodes increases, the retrieval effort (measured in cumulative decisions) for Microcases grows significantly faster than for the other methods.
Key results from the study include:
- Storage Efficiency: The multicase method required the least memory, avoiding the redundancy trap of individual cases.
- Retrieval Speed: Microcases exhibited a retrieval effort an order of magnitude larger than multicases.
- Scalability: While individual cases performed well in retrieval for small datasets, the multicase approach remained stable and efficient as the library of episodes grew, avoiding the "exponential explosion" of storage costs.
Conclusion
This research demonstrates that the organization of procedural knowledge is not merely a implementation detail but a critical factor in system performance. By leveraging the inherent structure of procedures to index case segments, the multicase approach provides a scalable framework for industrial-strength CBR systems. Future work will extend this analysis to include the costs of case adaptation and more complex, non-linear procedural structures.
Summary updated