Sparse Approximation For Gaussian Process With Derivative Observations

We propose a sparse Gaussian process model to approximate full Gaussian process with derivatives when a large number of function observations t and derivative observations t' exist. By introducing a small number of inducing point m, the complexity of posterior computation can be reduced from O((t+t')(3)) to O((t+t')m(2)). We also find the usefulness of our approach in Bayesian optimisation. Experiments demonstrate the superiority of our approach.