How priors of initial hyperparameters affect Gaussian process regression models

摘要: Abstract The hyperparameters in Gaussian process regression (GPR) model with a specified kernel are often estimated from the data via maximum marginal likelihood. Due to non-convexity of likelihood respect hyperparameters, optimisation may not converge global maxima. A common approach tackle this issue is use multiple starting points randomly selected specific prior distribution. As result choice distribution play vital role predictability approach. However, there exists little research literature study impact distributions on hyperparameter estimation and performance GPR. In paper, we provide first empirical problem using simulated real experiments. We consider different types priors for initial values some commonly used kernels investigate influence GPR models. results reveal that, once chosen, have no significant prediction, despite that estimates very true cases.