Alternative Theories of Inference in Expert Systems for Image Analysis.

Abstract : Reasoning with uncertainity is essential in most expert system applications to image understanding, both for bottom-up analysis of pixel data and for top-down utilization of general knowledge. A variety of alternative approaches have been proposed for inherence in expert systems: Bayesian probabilities, belief functions, fuzzy sets, non-monotonic logic, and others. They differ in the concepts they seek to address, in their normative justification, in computational feasibility and input burden, and psychological aptness for the purposes of experts and expert system users. The present report has two main objectives: (1) to clarify the strengths and weaknesses of various inference mechanisms for expert system applications, and (2) to develop alternative approaches which remedy shortcomings and retain strengths. We develop the top-level design of a new system, the Non-Monotonic Probabilist (NMP), which takes into account the actual practice of expert statisticians in probabilistic reasoning.