Automated Qualitative Domain Abstraction

Abstract

Automated problem-solving for engineered devices is based on models that capture the essential aspects of their behavior. In this paper, we deal with the problem of automatically abstracting behavior models such that their level of granularity is as coarse as possible, but still sufficiently detailed to carry out a given behavioral prediction or diagnostic task. A task is described by a behavior model, as composed from a library, a specified granularity of the possible observations, and a specified granularity of the desired results. The goal is to determine partitions for the domains of the variables (termed qualitative values) that are both necessary and sufficient for the task at hand. We present a formalization of the problem within a common relational (constraint-based) framework, present results regarding solutions to task-dependent qualitative domain abstraction, and devise methods for automatically determining qualitative values for a device model.