点击次数：

论文类型：会议论文

发表时间：2017-01-01

收录刊物：CPCI-S

页面范围：1671-1675

关键字：Bayesian optimization; data parallel; gaussian; surrogate; adaptive

摘要：Bayesian optimization has been demonstrated as an effective methodology for the global optimization. However, it suffers from a computational bottleneck that the inference time grows cubically with the number of observations. In this paper, a Bayesian optimization based on the data-parallel approach is proposed to alleviate this problem. Firstly, an improved geometry motivated clustering algorithm is developed to partition the search space. Then, multiple local surrogate models are constructed. At last, an expected improvement acquisition function based on the adaptive surrogate model is developed. The experiment results for six test problems show that the proposed method converges to good results on a limited computational budget and gives significant savings in computational cost.