Probabilistic Similarity Networks 论文

摘要

I address practical issues concerning the construction of normative expert systems--expert systems that encode knowledge within a decision-theoretic framework. In particular, I examine the similarity network and partition, two extensions to the influence diagram. A similarity network is a tool for building an influence diagram, whereas a partition is a tool for assessing the probabilities associated with an influence diagram. Both representations encode asymmetric forms of conditional independence that are not represented conveniently in an ordinary influence diagram. Similarity networks and partitions exploit these forms of conditional independence to facilitate the construction and assessment of influence diagrams for problems of diagnosis. The representations aided considerably the construction of Pathfinder, a large normative expert system for the diagnosis of lymph-node diseases (the domain contains approximately 60 diseases and 110 disease findings). In an early version of the system, I encoded the knowledge of the expert using an erroneous assumption that all disease findings were conditionally independent, given each disease. When the expert and I attempted to build an influence diagram for the domain to capture the dependencies among the disease findings, we failed. Using a similarity network, however, we were able to construct the influence diagram for the entire domain in approximately 40 hours. Furthermore, using the partition representation, the expert was able to decrease the time required to assess a probability--on average--by almost one order of magnitude. Most important, through a comparison procedure based in decision theory, I found that the improvements in diagnostic accuracy afforded by the more sophisticated model of the domain were well worth the additional effort that we had invested to build the revised version of the system. In this work, I examine in detail the theoretical properties of similarity networks and partitions, and discuss the application of these representations to the construction of Pathfinder. This research suggests strongly that, by identifying specific forms of conditional independence, and by developing representations that exploit these forms of independence for knowledge acquisition, knowledge engineers can construct normative expert systems for domains of larger scope and greater complexity than the domains previously through to be amenable to the decision-theoretic approach.