A0614
Title: Multi-fidelity Gaussian process modeling with boundary information
Authors: Matthias Tan - City University of Hong Kong (Hong Kong) [presenting]
Abstract: Time-consuming bi-fidelity simulations with a high-fidelity (HF) simulator and a low-fidelity (LF) simulator, where the HF simulator contains a vector of inputs not shared with the LF simulator, called the augmented input, frequently arise in practice. For such simulations, it is frequently known a priori that when the augmented input converges to any value in a subset of the boundary of its domain, the HF simulator output converges to the LF simulator output. This is a form of boundary information, i.e., prior information on a simulator's output at the boundary of the domain of the simulator's inputs. The standard autoregressive Gaussian process (GP) emulator can be constructed to approximate and replace the HF and LF simulators to reduce simulation time. However, this emulator does not satisfy boundary information. A solution will be presented to the problem of constructing a bi-fidelity GP emulator that satisfies the form of boundary information just mentioned. The proposed emulator called the boundary-modified autoregressive GP (BMAGP) emulator, is shown to outperform the standard autoregressive GP emulator with some examples.
