ENTROPY-DRIVEN INFERENCE AND INCONSISTENCY

Probability distributions on a set of discrete variablesare a suitable means to represent knowledge about theirrespective mutual dependencies. When now thingsbecome evident such a distribution can be adapted tothe new situation and hence submitted to a soundinference process. Knowledge acquisition and inferenceare here performed in the rich syntax of conditionalevents. Both, acquisition and inference respect asophisticated principle, namely that of maximumentropy and of minimum relative entropy. The freedomto formulate and derive knowledge in a language of richsyntax is comfortable but involves the danger of contradictionsor inconsistencies. We develop a method howto solve such inconsistencies which go back to theincompatibility of experts’ knowledge in their respectivebranches. The method is applied to the diagnosis inChinese medicine. All calculations are performed in theEntropy-driven expert system shell SPIRIT.....