Case-based representations for procedural knowledge
Richard Alterman, Roland J. Zito-Wolf · 1993
Case-based (CBR) refers to the class of memory-based problem solving methods that emphasize the adaptation of recalled solutions (explanations, diagnoses, plans) over their generation from first principles. CBR is emerging as one of the major methodologies for constructing intelligent systems. One of the primary problems in CBR has been the lack of a comprehensive framework for organizing and evaluating the many case organization and representation methods that have been proposed. This dissertation presents a theory of representation for reasoning. Formal models of procedure and case-based reasoning are developed and used to evaluate two current organizational proposals, individual cases and microcases. A new case organization, the multicase, is presented. Its advantages over the existing methods are quantified using the formal model. An implementation of multicases in the FLOABN system is presented. FLOABN acquires procedures for operating mundane electronic and mechanical devices, such as telephones and photocopiers. The core of FLOABN is a multicase-based adaptive planner, SCAVENGER. The multicase organization of episodic memory is used to support several important functions: retrieval at different levels of detail, reconstructive memory, incremental learning, routinization, and the task-guided interpretation of instructions. The multicase provides a structure by which both procedural and incidental knowledge can be stored and accessed efficiently. Empirical evidence from FLOABN is presented that suggests that the trends observed in the formal model also apply to case bases of moderate size. The thrust of this work is that differences in case-base organization have significant effects on the performance of CBR systems. The formal analysis of case organizations is a first in the CBR field. It is hoped that this work will further the development and analysis of case-base organizations and help CBR practitioners make more informed choices about them. One conclusion of this work is that, for the class of procedures studied, acquisition may be more effectively conceived of as the example-based construction of networks than as the recall individual example(s). While this might be interpreted as a negative result, it offers the possibility of a fruitful exchange of methods and problems between the two areas.