A Bayesian Dynamic Model for Computer Emulation with Seasonality

• Activity Number: 638
• Type: Invited
• Date/Time: Thursday, August 4, 2011 : 10:30 AM to 12:20 PM
• Sponsor: IMS
• Author(s): Fei Liu*+ and Hongfei Li and Yasuo Amemiya
• Companies: IBM Thomas J. Watson Research Center and IBM Thomas J. Watson Research Center and IBM Thomas J. Watson Research Center
• Keywords: Computer model emulation ; Dynamic linear model ; Forwarding filtering backward sampling ; Gaussian process ; Markov chain Monte Carlo ; Time-Varying Autoregression

Abstract

Computer models are deterministic physics based models that are developed for complex experiments. They are often computationally expensive and thus it is desirable to develop suitable emulators, statistical surrogate models to the expensive to run computer models. For computer models with dynamic outputs, fitting the traditional Gaussian process models is not feasible due to the high dimensionality. A dynamic modeling strategy, based on multivariate time varying auto-regression, has been developed for higher-dimensional, time-dependent or functional outputs, which provides a computationally attractive alternative as well as a flexible modeling approach capable of emulating a broad range of stochastic structures underlying the input-output simulations. Motivated by the need in real applications to explicitly model the seasonality, we extend the dynamic modeling strategy to incorporate the underlying seasonal patterns in the time series outputs. We illustrate the approach with data from a weather simulation model. The general strategy will be useful for other computer model studies, in particular for those whose outputs are time series with seasonality.