A0625
Title: Linked deep Gaussian process emulation of model networks
Authors: Deyu Ming - University College London (United Kingdom) [presenting]
Abstract: Computer models can be expensive and quickly become computationally prohibitive with large simulations. Statistical emulators, such as Gaussian processes, are thus essential for accelerating these simulations, enabling efficient downstream analyses like sensitivity analysis, calibration, and optimization, especially when computational resources are limited. Modern scientific problems often span multiple disciplines, necessitating the integration of distinct computer models. Building statistical emulators for such networks of computer models is challenging due to each model's unique functional complexities, computation times, and programming environments. Linked deep Gaussian process (LDGP) emulation offers a powerful solution by conceptualizing a computer model network as deep Gaussian processes with partially exposed hidden layers. Stochastic imputation, which integrates the expectation-maximization algorithm with elliptical slice sampling, is developed to infer these partially exposed deep networks. Synthetic and empirical examples, including the Joint UK Land environment simulator (JULES) and LDGP emulators, augmented by sequential designs and automatic structural pruning, have been found to perform significantly better than conventional Gaussian process emulators in predictive accuracy and uncertainty quantification. The implementations of these examples are facilitated by the dgpsi package, which is publicly available for both R and Python users on CRAN and CONDA.
