The thesis explores training and inference of Gaussian processes in frequency domain. The multi fidelity setting is considered where a less costly low- fidelity function is used to help during the prediction of the high- fidelity samples.
Gaussian processes can have different kernels. In the project one specific type of kernels is explored, namely the Spectral mixture kernels. Additionally the BNSE initialisation is applied to the kernel.
In multi- fidelity setting one can include the information about low- fidelity samples in different ways. One idea could be to introduce an additional structure in the Gaussian process kernel. This structure on one hand separates low- fidelity inputs from the domain inputs x. On the other hand one models linear structure by introducing kernels modelling additive and multiplicative components.