finin@PRC.UNISYS.COM (06/27/89)
Dissertation Defense University of Pennsylvania COMPUTATIONAL VALUE ANALYSIS David Klein This dissertation presents Computational Value Analysis (CVA), a framework for explaining and refining choices among competing alternatives in the context of intelligent systems. CVA increases the transparency of value theory, a formal model of choice, to provide a framework for modeling choices that is both formal and transparent. The components of CVA include (1) an interpretation of value theory that provides an intuitive yet formally sound vocabulary for talking about choices, (2) a set of strategies for explaining choices, and (3) a set of strategies for refining choices. CVA at once addresses problems in artificial intelligence (AI) and in decision analysis (DA). From an AI perspective, CVA provides a general foundation for building formally justifiable, intelligible, modifiable systems for choosing among alternatives. A secondary contribution of the work to AI is a set of observations concerning formality and transparency; although previous approaches to modeling choices in a systems context generally have reflected a view of formality and transparency as competitive properties of representations, our experience developing CVA suggests that these properties are synergistic. Finally, the dissertation outlines a potential approach to employing other formal models in the context of intelligent systems. From a DA perspective, CVA addresses problems of transparency. First, CVA can potentially increase the acceptance of decision-theoretic advice by providing methods for justifying that advice in intuitive terms. Second, CVA provides a means for managing bias in parameter assessment; the framework provides users with an opportunity to observe the step-by-step effect of a parameter value on the final result, so that users' responses are less likely to be influenced by the fashion in which parameter-assessment questions are posed. Third, CVA can potentially reduce the demands on parameter-assessment methods by providing for the incremental repair of model parameters. Finally, the framework provides an approach to the problem of managing changing preferences over time. 4:30 pm, June 28, 1989 Moore 554 University of Pennsylvania COMMITTEE ------------------------- T.W. Finin (advisor) E.H. Shortliffe (advisor) N.I. Badler E.K. Clemons A.K. Joshi M.O. Weber