G-STAR: A new Kriging-based trust region method for global optimization

摘要

Trust region methods are an efficient technique to identify good solutions when the sampling effort needs to be controlled due to the cost of running simulation. Meta-model based applications of trust region methods have already been proposed and their convergence has been characterized. Nevertheless, these approaches keep the strongly local characteristic of the original trust region method. This is not desirable in that information generated at local level are “lost” as the search progresses. A first consequence is that the search technique cannot guarantee global convergence. We propose a global version of the trust region method, the Global Stochastic Trust Augmented Region (G-STAR). The trust region is used to focus the simulation effort and balance between exploration and exploitation. Such an algorithm focuses the sampling effort in trust regions sequentially generated by adopting an extended Expected Improvement criterion. This paper presents the algorithm and the preliminary numerical results.