Reference: J. Rickel (1995), Automated Modeling of Complex Systems for Answering Prediction Questions. PhD thesis, Dept CS, Univ Texas at Austin. Masters' Thesis, Dept CS, Univ. Texas at Austin. Planning Operators. Tech Report AI96-242, Dept CS, Univ Texas at Austin. (1996).

Abstract: The ability to answer prediction questions is crucial in science and engineering. A prediction question describes a physical system under hypothetical conditions and asks for the resulting behavior of specified variables. Prediction questions are typically answered by analyzing (e.g., simulating) a mathematical model of the physical system. To provide an adequate answer to a question, a model must be sufficiently accurate. However, the model must also be as simple as possible to ensure tractable analysis and comprehensible results. Ensuring a simple yet adequate model is especially difficult for complex systems, which include many phenomena that can be described at many levels of detail. While tools exist for analysis, modeling is a creative, time-consuming task performed by humans.

We have designed algorithms for automatically constructing models to answer prediction questions, implemented them in a program called TRIPEL, and evaluated them in the domain of plant physiology. Given a prediction question and domain knowledge, TRIPEL builds the simplest differential-equation model that can adequately answer it and automatically passes the model to a simulator to generate the desired predictions. TRIPEL uses knowledge of the time scales on which processes operate to identify and ignore insignificant phenomena and choose quasi-static representations of fast phenomena. It also uses novel criteria and methods to choose a suitable system boundary, separating relevant subsystems from those that can be ignored. Finally, it includes a novel algorithm for efficiently searching through alternative levels of detail in a vast space of possible models. TRIPEL successfully answered plant physiology questions using a large, multipurpose, botany knowledge base (covering 300 processes and 700 plant properties) independently developed by a domain expert. Because its methods are domain-independent, TRIPEL should be equally useful in many areas of science and engineering.

Compressed Postscript: http://www.cs.utexas.edu/users/mfkb/papers/rickel-phd.ps.Z.