Efficient Compositional Modeling for Generating Causal Explanations

摘要

Building adequate models that embody the simplifications, abstractions, and approximations that parsimoniously describe the relevant system phenomena for the task at hand is essential for effective problem solving. Compositional modeling is a framework for constructing adequate device models by composing model fragments selected from a model fragment library. This paper presents an implemented polynomial-time model composition algorithm for constructing adequate models that provide parsimonious causal explanations of the functioning of a device. To model important aspects of device function, we introduce expected behaviors, an abstract, causal accounts of what a device does. To focus the search for a model and guarantee efficient model construction, we introduce a new approximation relationship between model fragments, called a causal approximation. For efficient model fragment retrieval and model generation, we organize model fragments into various hierarchies. For efficient model validation, we use causal ordering and a new logarithm-based order of magnitude reasoning technique. We have implemented the compositional modeling algorithm and produced adequate models and causal explanations of how a variety of electromechanical devices function, based on a library of 20 component types and 150 model fragments.