Seminar| Institute of Mathematical Sciences
Time: Friday, October 26th, 2024 , 10:00-11:00
Speaker: Deyu Ming, University College London
Abstract: Emulation, or surrogate modelling, plays a crucial role in computer model experiments by enabling fast uncertainty quantification (UQ), including sensitivity analysis and calibration, for computationally expensive simulators through mimicking their input-output relationships. In this talk, I will present recent advances in deep Gaussian process (DGP) emulation using Stochastic Imputation (SI), a flexible and efficient statistical framework for constructing emulators. I will focus on five key aspects of DGP emulation: expressivity, scalability, connectivity, dimensionality, and stochasticity, each illustrated with synthetic and real-world examples using our R package 'dgpsi', freely available on CRAN.
To begin, I will demonstrate how DGP emulators outperform conventional Gaussian process (GP) emulators in predictive performance, particularly with smaller training datasets generated through sequential design, and when the underlying simulators exhibit non-stationary behaviour. For scalability, I will show how fast inference for DGP emulators can be achieved with tens of thousands of data points using the Vecchia approximation. In terms of connectivity, I will illustrate how DGP emulators can be linked analytically to build surrogate models for simulators composed of sub-processes, enabling the emulation of a computation-heavy land-environment simulator that predicts time-series of carbon storage in trees across the UK. Regarding dimensionality, I will explain how effective dimensionality reduction can be achieved when an active subspace is inadequate, by transforming model dimensions into networks of DGP emulators. Finally, I will introduce the latest developments in DGP emulation within a generalised and scalable SI framework for stochastic computer simulators, including cases where replicates are available.