Causal and plausible reasoning in expert systems.

摘要

This research is directed at the development of a better understanding of the roles of causal and plausible reasoning in the management of uncertainty in expert systems. In earlier studies, these modes of reasoning were considered as separate issues, with a dissociation of the causal aspects from an assessment of the degree of likelihood. In the present study, the modes in question are analyzed from a unified point of view, yielding a new type of representation called structured rules that reflect both inference causation and strength. Causation is endorsed through a set of causal relationships known as roles, including "sufficient," "associational," "supportive," "weak" and "strong necessary," "contrary," and "exceptional," whereas inference strength is measured by a "conditional basic probability assignment" associated with the conclusion, much as the Bayesian conditional probability addresses uncertain rules.; Each causal role describes qualitatively a special form of inference. In addition, different roles, when combined for the same structured rule, produce a body of coherent knowledge represented locally. In this way, a normal associational relationship can be augmented with several supportive or exception conditions. In comparison to conventional rules, such localized structure facilitates more focused knowledge acquisition and simplifies the task of rule interpretation when a conclusion has to be modified at a later point in the inference process.; The Dempster-Shafer theory is selected as a basis for plausible reasoning because of its generality and ability to deal with incomplete information. The original theory is extended to support chains of reasoning and to combine conclusions from multiple rules when prior information is available. Practical solutions to the problem of reasoning with imprecise concepts are also developed. With these extensions, measures of beliefs can be expressed in the context of individual causal roles; during reasoning, these beliefs and the degree of ignorance are then propagated through chains of inferences. In addition, this work generalizes earlier results on evidential reasoning.; A prototype knowledge-based system for venture investment evaluation has been implemented on KEE. Its purpose is to demonstrate reasoning under uncertainly based on the extended theory and to experiment with the structured rules when expressed as frames and slots.