Part of Advances in Neural Information Processing Systems 31 (NeurIPS 2018)
Mojmir Mutny, Andreas Krause
We develop an efficient and provably no-regret Bayesian optimization (BO) algorithm for optimization of black-box functions in high dimensions. We assume a generalized additive model with possibly overlapping variable groups. When the groups do not overlap, we are able to provide the first provably no-regret \emph{polynomial time} (in the number of evaluations of the acquisition function) algorithm for solving high dimensional BO. To make the optimization efficient and feasible, we introduce a novel deterministic Fourier Features approximation based on numerical integration with detailed analysis for the squared exponential kernel. The error of this approximation decreases \emph{exponentially} with the number of features, and allows for a precise approximation of both posterior mean and variance. In addition, the kernel matrix inversion improves in its complexity from cubic to essentially linear in the number of data points measured in basic arithmetic operations.