Context Sensitive Evaluation of Uncertainty During Inference

In the application of Artificial Intelligence to industrial domains, uncertainty is almost always present in both measurement and inference. In measurement, uncertainty is manifested in the characterization of information as indicating an entity of a specific class or type. In inference, uncertainty is manifested when drawing conclusions from the presence of enabling conditions or data. These two types of uncertainty interact when a system is required to draw conclusions using probabilistic rules and uncertain data. While a number of numeric approaches to the description and combination of uncertain data exist, all are subject to the criticism that none effectively exhausts the range of uncertain reasoning that humans are capable of (Shafer and Tversky, 1985). Furthermore, numerical approaches cannot capture, in a single value, the reasons why a given fact or inference is to be believed; only how much it is believed. To express the reasons for believing a set of propositions, numerical uncertainty values must be supplemented by symbolic uncertainty values.