Surrogate modeling for Bayesian optimization beyond a single Gaussian process