Scalable Bayesian optimization with generalized product of experts

Bayesian optimization (BO) is challenging for problems with large number of observations. The main limitation of the Gaussian Process (GP) based BO is the computational cost which grows cubically with the number of sample points. To alleviate scalability issues of standard GP we propose to use the generalized product of experts (gPoE) model. This model is not only very flexible and scalable but can be efficiently computed in parallel. Moreover, we propose a new algorithm gPoETRBO for global optimization with large number of observations which combines trust region and gPoE models. In our experiments, we empirically show that our proposed algorithms are computationally more efficient and achieve similar performance to other state-of-the-art algorithms without using any specialized hardware.