Combined Global and Local method for stochastic simulation optimization with an AGLGP model

Surrogate methods, motivated from expensive black box simulations, are efficient approaches to solve stochastic simulation optimization problems. However, estimating an appropriate surrogate model can still be computationally challenging when the data size gets large. In this paper, we propose a new optimization algorithm based on a previously proposed Additive Global and Local Gaussian Process model (AGLGP). This algorithm leverages the global and local features of an AGLGP model and can automatically switch between a global search (for a promising region) and a local search (within the promising region). The algorithm proceeds by globally narrowing down the search space sequentially, but allows it to escape from the current search region. We numerically illustrate the mechanism behind the algorithm in an example.