Abstract:
|
We explore ensembles of Gaussian processes (GPs) and connect them to standard notions of robustness, GP approximations and computational efficiency. Given a particular GP kernel, hyperparameter selection may be challenging, particularly in high dimensional problems. We evaluate the ability of ensembles of relatively inflexible GPs, which function as weak learners, to recover the true process while bypassing the hyperparameter selection problem. We also consider ensembles of GP approximations to better approximate a full GP while harnessing the computational benefits of the approximation techniques.
|