A1055
Title: Recent advances of deep Gaussian process emulation
Authors: Deyu Ming - University College London (United Kingdom) [presenting]
Abstract: Emulation, or surrogate modeling, is essential in computer experiments, enabling rapid evaluation of complex and computationally intensive simulators by replicating their input-output relationships. The aim is to present recent advances in deep Gaussian process (DGP) emulation using stochastic imputation (SI), a versatile and efficient statistical framework for emulator construction. Specifically, five key aspects of DGP emulation are discussed expressivity, scalability, connectivity, dimensionality, and non-Gaussianity each demonstrated through synthetic or real-world examples using the R package 'dgpsi', available on CRAN. It also illustrates how DGP emulators significantly outperform conventional GP emulators in predictive accuracy, particularly with smaller datasets derived from sequential designs and non-stationary simulators. For scalability, rapid inference capabilities are showcased for DGP emulators on large datasets using the Vecchia approximation. Regarding connectivity, it is shown how DGP emulators can analytically link subprocesses to construct network surrogate models. In dimensionality, the aim is to explain how effective dimensionality reduction can be achieved when an active subspace is inadequate by transforming model dimensions into networks of emulators. Finally, recent developments in generalized, scalable SI-based DGP frameworks designed to handle non-Gaussian simulator outputs, including scenarios involving replicate observations, are discussed.
