Relevance Reasoning to Guide Compositional Modeling

The ability to choose an appropriate manner in which to model a given device is crucial in making a compositional modeling [3] approach successful. In compositional modeling, a system is provided with a library of composible pieces of knowledge about the physical world, called model fragments, each representing a conceptually distinct phenomenon such as a physical process or one aspect of a component behavior. Given a specific query about a device, the system chooses among those model fragments to compose a model of the device that is most adequate to answer the query. Selection of appropriate model fragments can be viewed as a special case of a more general problem of reasoning about relevance of knowledge to a given goal. In this paper we pursue this view by applying a general framework for reasoning about relevance to the problem of model fragment selection. We show that heuristics for model selection can be usciully stated as irrelevance claims. Employing such a framework allows one to state both general and domain-specific heuristics about relevance declaratively, as opposed to building them into the control structure of the system. Given relevance heuristics stated in the language, our relevance reasoning system can immediately make use of them to control the model formulation process, enabling us to experiment easily with different heuristics.