Abstract

Several formalisms for representing and reasoning with uncertain knowledge have been proposed in the artificial intelligence literature. Unfortunately, analyses of the adequacy of each formalism to different types of problems have seldom appeared, and designers are often forced to make arbitrary choices about how to model uncertainty in their domain. In this paper, we present an experimental approach to comparing uncertainty management techniques in the light of a specific problem to solve. We model a problem tailored on a real‐world application using three major techniques, namely, probability theory, Dempster‐Shafer's theory, and possibility theory, and discuss the results. We also propose a new qualitative way of analyzing the behavior of the three techniques that highlights some interesting assumptions. The experiment has been performed using PULCINELLA, a tool for propagating uncertainty based on the local computation technique of Shafer and Shenoy that can be specialized to each of our target uncertainty formalisms.