Representativeness and Uncertainty in Classification Schemes

The choice of implication as a representation for empirical associations and for deduction as a mode of inference requires a mechanism extraneous to deduction to manage uncertainty associated with inference. Consequently, the interpretation of representations of uncertainty is unclear. Representativeness, or degree of fit, is proposed as an interpretation of degree of belief for classification tasks. The calculation of representativeness depends on the nature of the associations between evidence and conclusions. Patterns of associations are characterized as endorsements of conclusions. We discuss an expert system that uses endorsements to control the search for the most representative conclusion, given evidence. Tasks can be classified by the kinds of uncertainty that characterize them. Planning tasks, for example, arc characterized by uncertainty about the interactions of plan steps. Strategic plamling is further characterized by uncertainty about the intentions and actions of an opponent. Perception is characterized by too much data too noisy for bottom-up interpretation, and ambiguous with respect to top-down models. The subject of this article is classzficatzon, an important task for many AI systems. Most expert systems are classification problem solvers. They heuristically associate data with one or more known solutions; the problem is to match data with the solution that explains them best (Clancey, 1984). Uncertainty in ClassifiCation problem solving has two major sources. The first is that data may be inaccurate or incomplete, and the second is partial matching. This This work is supported by NSF Grant IST-8409623 and DARPARADC Contract F30602-85-C-0014 We wish to acknowledge Bmce McCanless and Marg Burggren for their patient assistance and Daniel Corkill for his advice on drafts of this article article is not concerned with the quality of data; we focus instead on uncertainty inherent in the design and behavior of classification systems. The partial matching problem has two forms, easily illustrated by the following common, empirical association: A person with a queasy stomach, fatigue, aching limbs, and a fever has flu in its early stages. Now consider a person with a marginal fever, complaining of poor appetite, headache, and a persistent twitch in his left eye. This case seems to exhibit manifestations not stated in the rule for flu and fails to display manifestations that are so stated. We are uncertain whether the person has flu for two distinct reasons: we cannot be certain that the actual symptoms fail to match the stated ones (Does LLmarginal fever” count as a fever? Does “headache” count as aching limbs?); and we cannot be certain that the rule for flu includes all and only the relevant manifestations of