Automated Modeling for Answering Prediction Questions: Selecting Relevant Influences

The ability to answer prediction questions is crucial in reasoning about physical systems. A prediction question poses a hypothetical scenario and asks for the resulting behavior of variables of interest. Prediction questions can be answered by simulating a model of the scenario. Constructing a suitable model requires distinguishing relevant aspects of the scenario from irrelevant aspects. This paper provides criteria for making this distinction, and it presents an algorithm that uses these criteria to construct a suitable model for answering a given prediction question. The algorithm has been implemented in a modeling program called TRIPEL, arid the paper summarizes a preliminary evaluation of TRIPEL in the plant physiology domain.