Doctor of Philosophy, The Ohio State University, 2009, Statistics

We propose a methodology for the simultaneous optimization of multiple goal functions evaluated by a numerically intensive computer model. In a black box multiobjective problem, the goal is to identify a set of compromise solutions that provide a minimally sufficient representation of the Pareto front in the most efficient manner. To reduce the computational overhead, we adopt a surrogate-guided approach where we perform optimization sequentially via improvement. Our algorithm relies on a multivariate Gaussian process emulator which uses a novel multiobjective improvement criterion called the expected Pareto improvement function to guide the sampling of points in the Pareto efficient region. We show that the algorithm is capable of approximating the Pareto front within a computational budget.

Committee: Thomas Santner PhD (Advisor); Peter Craigmile PhD (Committee Member); William Notz PhD (Committee Member) Subjects: Statistics