Ph.D. Dissertation, Indiana University, 2001.
David C. Wilson.
Case-based reasoning (CBR) is an artificial intelligence methodology that uses specific encapsulated prior experiences as a basis for reasoning about similar new situations. CBR systems rely on various ``knowledge containers,'' such as the case-base of prior experiences and similarity criteria for comparing situations and retrieving the most relevant cases. Explicit or implicit changes in the reasoning environment, task focus, and user base may influence the fit of the current knowledge state to the task context, which can affect the quality and efficiency of reasoning results. Over time, the knowledge containers may need to be updated in order to maintain or improve performance in response to changes in task or environment. In particular, maintaining the case-base---the traditional mainstay of knowledge underlying CBR systems---is essential for preserving and expanding the capability of a CBR system throughout its life-cycle.
This dissertation provides a first coherent picture of the overall case-base maintenance problem in CBR and develops new case-base maintenance techniques within that paradigm. The thesis presents a theoretical framework for describing case-base maintenance techniques according to the types of maintenance policies implemented by a given system. The framework serves to unify current maintenance practice, to point out areas for new fundamental research, and as a step toward recommending the best maintenance practices for varying system performance goals. In that context, the thesis goes on to make an examination and account of underlying regularity assumptions in the CBR process that directly affect maintenance activity.
The theoretical picture of case-base maintenance is then complemented with a presentation of new methods and experiments in applied case-base maintenance. The thesis first presents DRAMA, a case-based tool developed for aerospace design support at NASA, Ames, which provides facilities for initial case capture and subsequent refinement that directly exploit user knowledge. By monitoring users as they go about normal high-level design tasks, DRAMA automatically captures user design choices and rationale that can be used to provide proactive recommendations at both the design and design component levels. This helps to maintain the case-base through continuous support for case-authoring and design consistency, while significantly ameliorating the knowledge-engineering burden on system users.
Next, the thesis presents a practical model for transforming case-base implementations in the Metamorphoses project for representational maintenance. The thesis goes on to examine new methods for automatically maintaining case-bases by incorporating explicit performance concerns into measures of case-base competence in order to optimize case-base composition.
Finally, the thesis describes how the work developed for case-base maintenance can generalize across knowledge containers. The framework for case-base maintenance is extended and applied in the general knowledge container context for case-based reasoner maintenance. The thesis describes applied maintenance beyond the case base, and it presents an application of similarity maintenance in the context of the CBMatrix case-based recommender system for problem-solving support in Scientific Computing.
The whole provides a unifying framework and algorithms for constructing and maintaining CBR systems that may be used over extended periods of time and in changing environments---a valuable resource for implementers, maintainers, and users of CBR systems.
See http://www.cs.indiana.edu/~leake/INDEX.html for additional publications in the Artificial Intelligence/Cognitive Science report and reprint archive maintained by David Leake.