Title: Calculating Empirical Sensitivity Metrics for Computer Experiments with Constrained Input Spaces
Computer experiments use computer simulators that are based on a mathematical model of a physical process as experimental tools to determine “responses” at a set of user-specified “input” values. Global Sensitivity Analysis (GSA) is the process of apportioning the variation in a statistical predictor of a computer experiment response to each of its respective inputs. There are a variety of existing approaches to GSA, many of which rely on simplified calculations that are possible when the simulator has a hyper-rectangular input space. In reality, input spaces are often constrained in some way, making the existing GSA methods difficult to compute. Thus, we propose an empirical method for GSA that averages simulator predictions over a grid of input space points that accounts for the constraints on the input variables. Given well-defined input restrictions, our method is fully automated and produces results that match well with theoretical values for examples having either constrained or unconstrained input spaces.
