Theory of Endorsements and Reasoning with Uncertainty, January 1984 - January 1986

摘要

By emphasizing the sources of uncertainty and its consequences, an intelligent reasoning system shows the ability to plan a course of action appropriate to one's uncertainty, the ability to explain one's actions, and the ability to determine degree of belief in alternatives given evidence. Other than numerical inferences and reason maintenance, another approach to parallel certainty inference, the theory of endorsements, is presented. The fundamental assumption of the theory of endorsements is that subjective degrees of belief are composites of reasons to believe and disbelieve (positive and negative endorsements). Two implementations of the theory of endorsements, SOLOMON and HMMM, are briefly described. GRANT, a knowledge system that finds sources of funding for research proposals, was developed to explore the utility of semantic matching in uncertain domains. The basic mode of inference used in GRANT is plausible inference. MU (Management of Uncertainty), a development environment which supports prospective reasoning, was used to reimplement a medical expert system, called MUM. MUM manages uncertainty by reasoning about evidence and its current state of belief in hypotheses. A prospective, constructive decision making system, called CDM, was developed to define and evaluate decisions autonomously. Contrast with the more static decision theoretic models, CDM is dynamic in that the algorithm gathers evidence as the decision evolves at the most opportune time. Three principles of design for knowledge acquisition are described and demonstrated their application is demonstrated in an architecture where knowledge about evidential combination and knowledge about control can be acquired from an expert.