System identification -- the process of finding a mathematical model of a dynamical black-box system -- is a routine and difficult problem faced by engineers in a variety of domains. This dissertation describes the program PRET, which automates this process by smoothly integrating a collection of highly heterogeneous reasoning modes: qualitative reasoning, qualitative simulation, numerical simulation, geometric reasoning, constraint reasoning, SLD-based resolution, reasoning with abstraction levels, declarative meta-level control, and a simple form of truth maintenance. Given observations, hypotheses, and specifications, the program constructs an ordinary differential equation (ODE) model of the target system. Unlike other modeling programs that map structural or functional descriptions to model fragments, PRET combines hypotheses about the mathematics involved into candidate models and then intelligently tests them against observations about the target system.
PRET's ODE theory is represented declaratively using a logic-based formalism. Typically, in the hierarchy from more-abstract to less-abstract models, the model of choice is the one that is just detailed enough to account for the properties and perspectives that are of interest for the task at hand. A human modeler's reasoning about a candidate model for a given physical system takes place at an abstract level first and resorts to more detailed reasoning later in the modeling process. PRET's reasoning framework is designed to mimic these strategies. The main goal of the work described in this thesis was to design and implement a knowledge representation and reasoning framework that allows a computer program to reason about physical systems and candidate models in such a way as to quickly find the right model at the right abstraction level. This knowledge representation framework was designed to allow easy formulation of knowledge and meta knowledge relative to various abstraction levels.
When modeling physical systems, human engineers make use of various existing modeling techniques. Likewise, PRET integrates various heterogeneous reasoning modules. In order to intelligently orchestrate these different reasoning modes, PRET makes use of abstraction levels and meta-level control techniques that allow PRET to reason about which techniques are appropriate in which situations. This is a first attempt at dynamically controlling multimodal reasoning systems, a task that has recently been identified as an important open research problem by the multimodal reasoning community.
Reinhard Stolle Last modified: Mon Jul 17 16:39:00 PDT 2000