MODEL CONSTRUCTION OPERATORS

Expert systems can be viewed as programs that construct a model of some system in the world so that it can be assembled, repaired, controlled, etc. In contrast with most conventional computer programs, these models represent processes and structures by relational networks. Control knowledge for constructing such a model can be described as operators that construct a graph linking processes and structures causally, temporally, spatially, by subtype, etc. From this perspective, we find that the terminology of blackboard expert systems is not specific to a particular set of programs, but is rather a valuable perspective for understanding what every expert system is doing. This paper reviews different ways of describing expert system reasoning, emphasizing the use of simple logic, set, and graph notations for making dimensional analyses of modeling languages and inference methods. The practical question is, how can we systematically develop knowledge acquisition tools that capture general knowledge about types of domains and modeling methods? Examples of modeling operators from ABEL, CADUCEUS, NEOMYCIN, HASP, and ACCORD demonstrate how diverse expert system approaches can be explained and integrated by the model construction perspective. Reworked examples from TEIRESIAS, XPLAIN, and KNACK illustrate how to write metarules without using domain-specific terms, thus making explicit their model construction nature. Generalizing from these observations, we combine the system-model and operator viewpoints to describe the representation of processes in AI programs in terms of three nested levels of domain, inference, and communication modeling. This synthesis reveals how the use of relational networks in computer programs has evolved from programmer descriptions of computational processes (such as flowcharts and dataflow diagrams) to network representations that are constructed and manipulated by the programs themselves.