Reasoning about model accuracy

Although computers are widely used to simulate complex physical systems, crafting the underlying models that enable computer analysis remains a difficult job with only a small number of computer tools for support. Our goal is to mechanize this process by building an automated model management system that evaluates simplifying assumptions and selects appropriate perspectives. In this paper we present our initial results: a framework for dynamically changing model accuracy based on model sensitivity analysis. We show how to perform model sensitivity analysis efficiently when one model is a fitting approximation of the other. Finally, we discuss two implementations of our technique in programs that perform • • query-directed simplification by adding bounding abstractions, and • • discrepancy-driven refinement. Our programs use a mixture of qualitative and quantitative techniques with both symbolic and numeric computations.

Authors

• Daniel S. Weld

Keywords

No keywords are indexed for this paper.