Case Library Design Guidelines Transformed Case Library Case Library Transformer Rover Simulator Evaluation

Conversational case-based reasoning (CCBR) systems are used in many commercial applications. Their distinguishing behavior is that they interactively acquire features values from users (i.e., they incrementally build queries), and typically address problem diagnosis tasks. Users want CCBR systems to be precise (i.e., retrieve the correct case for solving their problem) and eecient (i.e., require a minimal amount of interaction before cases are retrieved). Case authors require substantial expertise to ensure that their case libraries satisfy these demands. Commercial CCBR vendors help by supplying documents describing guidelines for designing case libraries, but mastering them involves a challenging learning curve. Software is needed that helps authors to design and reene case libraries according to these design guidelines. This paper outlines a proposal for using machine learning and data mining techniques to support the process of case authoring by reening case libraries. Recent growth in the commercial case-based reasoning (CBR) industry has been dramatic. Although commercial CBR tools were rst introduced in the late 1980's, at least 15 companies now market CBR shells or provide related services, and some have hundreds of thousands of end users. However, CBR is not a \magic bullet" that eliminates the need for good knowledge acquisition techniques. Rather, it focuses developers on the process of case authoring 2 rather than knowledge engineering. Authoring cases is exceedingly diicult in some contexts and requires substantial skill, which can be expensive in time (e.g., long learning curves) or money (e.g., for training fees, consulting fees, or for purchasing case libraries). CBR vendors provide users with library design guidelines because novice case authors frequently generate poorly designed case libraries, which can lead to low user satisfaction, causing clients to abandon CBR products. Unfortunately, written guidelines, even if extracted according to careful knowledge engineering principles, are seldom eeective because, in addition to being large in number, they are imperative rather than descriptive: guidelines typically state characteristics of good case library designs rather than explain how to guarantee good CCBR behavior. Novice case authors need additional guidance to increase their productivity, preferably interactive assistance from the CBR system itself. This paper introduces a proposal for assisting case authors under the assumption that design guidelines are available. Our proposal uses machine learning and data mining techniques to transform case libraries so that they comply with the design guidelines. We describe the context of our approach, outline our proposal, and exemplify the case transformation process.