Sequential Experiment Design for Contour Estimation from Computer Simulators

In many engineering applications, one is interested in identifying the inputs to a computer simulator that lead to a pre-specified output. In this talk we introduce statistical methodology that identifies the desired contour in the input space. The proposed approach has three main components. Firstly, a stochastic model is used to approximate the global response surface. The model is used as a surrogate for the underlying computer model and provides an estimate of the contour together with a measure of uncertainty, given the current set of computer trials. Then, a strategy for choosing subsequent computer experiments to improve the estimation of the contour is outlined. Finally, we discuss how the contour is extracted and represented. The methodology is illustrated with an example from a multi-class queueing system.