点击次数：

论文类型：期刊论文

发表时间：2016-10-01

发表刊物：STRUCTURAL AND MULTIDISCIPLINARY OPTIMIZATION

收录刊物：SCIE、EI、Scopus

卷号：54

期号：4

页面范围：747-773

ISSN号：1615-147X

关键字：Efficient global optimization; Parallel computing; Multiple points infill criterion; Mutual information; Expected improvement

摘要：This work is focused on the parallel algorithms in efficient global optimization. Firstly, a multiple points infill criterion named EI&MI is developed, which adopts the entropy to precisely measure the uncertainty of Kriging surrogate, and then balances global exploration and local exploitation of the multiple points infill sampling criteria. Secondly, given the computational difficulties in Kriging with a large size of training data, a domain decomposition optimization strategy is proposed, which ensures a small size of training data. Several mathematical functions and one engineering problem are employed as testing examples. The results show that comparing with several other methods, the EI&MI has an obvious advantage in solving complex optimization problems under the large-scale parallel computing environment, and the domain decomposition optimization strategy could improve the stability of optimization without sacrificing the optimization efficiency.