Reusable influence diagrams

摘要

Influence Diagrams have been recognized as a suitable formalism for building probabilistic expert systems. Nevertheless, the most part of applications consists in stand-alone systems, concerning a very limited domain. On the other hand, Artificial Intelligence research has outlined Blackboard Architectures as the basis for building expert systems in which several knowledge sources, in general built with different formalisms, cooperate to the solution of a complex task. This paper addresses the use of influence diagrams as knowledge sources of such a system, and particularly faces the problem of reusing the same influence diagram in different inference phases. We will show that, specialy in planning tasks, the modularity requirement of keeping the knowledge sources separated, may imply that an influence diagram must call another influence diagram to solve itself and to maintain the coherence of the whole set of decisions underlying the plan. Conditions for the correctness of this concatenation of knowledge sources will be provided, and an example from the medical domain of therapy planning for Acute Myeloid Leukemia will be shown, as an implemented prototype exploiting these ideas.