ABSTRACT
Design and analysis of computer experiments focused on deterministic solvers from the physical sciences via Gaussian process (GP) interpolation. But computer modeling is common in the social, management and biological sciences, where stochastic simulations abound. Noisier simulations/observations demand bigger experiments/studies to isolate signal from noise, and more sophisticated GP models – not just adding nuggets to smooth over noise, but variance processes to track changes in noise throughout the input space in the face of that heteroskedasticity. The chapter focuses on reducing mean-squared prediction error, a global criterion. Targeting specific regions of interest, such as global minima or level sets is also of interest. In the face of noise, and particularly heteroskedasticity, simplicity and computational tractability are important features when adapting a method originally designed for noiseless/deterministic settings.
