Choice of Initial Sample Size for the Sequential Analysis of a Computer Experiment

• Activity Number: 348
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Dexter C. Whittinghill*+
• Companies: Rowan University
• Keywords: computer experiments ; sequential analysis ; initial sample size ; maximin Latin hypercube ; expected improvement ; cross-validation

Abstract

Computer experiments, which comprise the evaluation of large, complicated mathematical models using computer code, continue to become more and more prominent in applied research as computers improve, and the body of literature on the design and analysis of computer experiments expands. While early research investigated the design and analysis for fixed sample sizes for the number of runs of the computer experiment, lately more and more investigators look at sequential experiments for optimization (Lam & Notz 2008, Williams, Santner & Notz 2000), determination of contours (Ranjan, Bingham & Michailidis 2008, Roy & Notz 2010) and calibration (Kumar & Notz 2010). However, guidance for what fraction of the budgeted sample size should be used for the initial sample size is largely unexplored. We will investigate the effect that the choice of different initial sample sizes has on the accuracy in optimization and the determination of contours.