Bayesian Treed Gaussian Process Models With an Application to Computer Modeling

Motivated by a computer experiment for the design of a rocket booster, this article explores nonstationary modeling methodologies that couple stationary Gaussian processes with treed partitioning. Partitioning is a simple but effective method for dealing with nonstationarity. The methodological developments and statistical computing details that make this approach efficient are described in detail. In addition to providing an analysis of the rocket booster simulator, we show that our approach is effective in other arenas as well.