In this paper, a Surrogate-Assisted Grey Wolf Optimization (SAGWO) algorithm for high-dimensional and computationally expensive problems is presented, where Radial Basis Function (RBF) is employed as the surrogate model. SAGWO conducts the search in three phases, initial exploration, RBF-assisted meta-heuristic exploration, and knowledge mining on RBF. In the initial exploration, the Design of Experiments is carried out to generate a group of well-distributed samples based on which the original wolf pack and wolf leaders are sequentially identified to approximate the high-dimensional space roughly. The knowledge mining on RBF includes a global search that is carried out using the grey wolf optimization and a local search that is performed over a focused local region using a search strategy combining global and multi-start local exploration. In the proposed SAGWO, knowledge gained from the RBF model assists the generation of new wolf leaders in each cycle, and the positions of the wolf pack are iteratively changed following the wolf leaders, thus reaching balanced exploitation and exploration. The new SAGWO algorithm presents superior computation efficiency and robustness as demonstrated by comparison tests with ten representative global optimization algorithms on 30, 50 and 100 design variables.

本文提出了一种针对高维和计算量大的问题的替代辅助灰狼优化（SAGWO）算法，其中采用径向基函数（RBF）作为替代模型。SAGWO进行三个阶段的搜索，即初始探索，RBF辅助的元启发式探索和有关RBF的知识挖掘。在最初的探索中，进行了实验设计以生成一组分布良好的样本，基于这些样本，原始狼群和狼首领被依次识别，以大致近似高维空间。RBF上的知识挖掘包括使用灰太狼优化进行的全局搜索和使用结合了全局和多起点本地探索的搜索策略在聚焦的局部区域上执行的局部搜索。在提出的SAGWO中，从RBF模型获得的知识可在每个周期中帮助生成新的狼首领，并且随着狼首领不断地改变狼群的位置，从而达到平衡的开发和探索。