EGO方法的训练算法及应用

Training Algorithms for EGO Method and Applications

针对高效全局优化(Efficient Global Optimization,简称EGO)方法的训练问题,选择元启发式(Meta-heuristic)算法、随机取样算法以及低频序列算法,并选用三个无约束、两个带约束解析优化算例以及两个气动优化算例,对这三类训练算法进行详细地比较研究,发现在元启发式算法中差分进化算法最具应用潜力,而低频序列算法可以有效降低EGO方法的随机性,其中Faure序列平均性能最优.

Three kinds of training algorithms for efficient global optimization （EGO） method are investigated. A kind of training algorithm based on low-discrepancy sequences is proposed to reduce randomness of EGO method. Performance of EGO method depends on a good training algorithm. Since training problems in EGO are non-convex and non-smooth, meta-heuristic algorithms, random algorithm and low-discrepancy sequences are chosen to address five benchmark optimization problems and two aerodynamic shape optimization problems. In these problems, differential evolution algorithm was found the best in meta-heuristic algorithms. Training algorithm based on low-discrepancy sequences can effectively reduce randomness of EGO method and Faure sequence has the best performance.