Generating explanations of device behavior using compositional modeling and causal ordering

TL;DR: It is shown how two techniques from the modeling research can be effectively combined to generate natural language explanations of device behavior from engineering models, and allows models that are more scalable and less brittle than models designed solely for explanation.

Abstract: Generating explanations of device behavior is a long-standmg goal of AI research in reasoning about physical systems. Much of the relevant work has concentrated on new methods for modeling and simulation, such as qualitative physics, or on sophisticated natural language generation, in which the device models are specially crafted for explanatory purposes. We show how two techniques from the modeling research--compositional modeling and causal ordering-- can be effectively combined to generate natural language explanations of device behavior from engineering models. The explanations offer three advances over the data displays produced by conventional simulation software: (1) causal interpretations of the data, (2) summaries at appropriate levels of abstraction (physical mechanisms and component operating modes), and (3) query-driven, natural language summaries. Furthermore, combining the compositional modeling and causal ordering techniques allows models that are more scalable and less brittle than models designed solely for explanation. However, these techniques produce models with detail that can be distracting in explanations and would be removed in hand-crafted models (e.g., intermediate variables). We present domain-independent filtering and aggregation techniques that overcome these problems.