Empirical sensitivity analysis for computational procedures

Abstract

Sensitivity analysis in computer science aims to improve stability in computer applications by considering uncertainty due to small perturbations in parameters. Mathematical and computational methods of sensitivity analysis are discussed. Advantages and disadvantages of both methods are addressed. A tool is developed to compute computational sensitivities. This tool was validated using three simple, well understood problems. The tool was then applied to a dynamic power grid system and an agent-based criminal computation. In the case of the power grid system, computational sensitivity analysis agrees with mathematical analysis in reporting stability and no excessive sensitivity. In the case of the criminal computation, the code is found to be unstable.