Parallel Partial Gaussian Process Emulation for Computer Models with Massive Output

• Activity Number: 349
• Type: Topic Contributed
• Date/Time: Tuesday, August 11, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Mengyang Gu* and James Berger
• Companies: Duke University and Duke University
• Keywords: Gaussian Process ; Computer Model Emulation ; Space-time Coordinate ; Reference Prior ; Composite Likelihood

Abstract

We consider the problem of emulating (approximating) computer models (simulators) that produce massive output. Our testbed example is a computer model of volcanic pyroclastic flow, a single run of which produces up to 10^9 outputs over a space-time grid of coordinates. An emulator (essentially a statistical model of the simulator -- we use a Gaussian Process) that is computationally suitable for such massive output is developed, and studied from practical and theoretical perspectives. On the practical side, the emulator does unexpectedly well in predicting what the simulator would produce, even better than much more flexible and computationally intensive alternatives. Theoretical results are developed that suggest why this is so. Generalizations of the emulator are introduced that allow for a nugget and spatial smoothing. The main roadblock to implementation is estimating the correlation parameters of the Gaussian Process. We utilize reference priors and composite likelihood methods in this estimation.