Failing of Dempster's combining rule of interval-given probabilities

摘要

A more realistic approach to knowledge-based systems operating under uncertainty requires that the inference engine be able to aggregate inference weights (probabilities) not necessarily point- (precisely-) given but also interval-given. In the frame of the intensional approach, this aggregation is, as before, automatically realized by previous knowledge integration but in the enlarged ‘knowledge set’ of compatible distributions. On the other hand, any extensional approach requires the introduction of supplementary aprioristic rules as the Dempster's combining rule for interval-given probabilities. It is proved that the latter leads to output probability intervals which may essentially differ from reasonably expected ones.