Decision space partition based surrogate-assisted evolutionary algorithm for expensive optimization
Decision space partition based surrogate-assisted evolutionary algorithm for expensive optimization
复制标题
DOI：
10.1016/j.eswa.2022.119075
复制
发表时间：
2022-10
期刊：
影响因子：
--
通讯作者：
Yuanchao Liu;Jianchang Liu;Shubin Tan
中科院分区：
文献类型：
--
作者：
Yuanchao Liu;Jianchang Liu;Shubin Tan
在昂贵的优化中，功能评估是基于昂贵的物理实验或耗时的模拟。此外，目标的梯度并不容易得到。因此，如何处理昂贵的优化问题是一项具有挑战性的任务。本文提出了一种基于决策空间划分的代理辅助进化算法（DSP-SAEA）。在DSP-SAEA中，引入了一种两阶段搜索策略，将全局搜索和局部搜索无缝集成。在全局搜索阶段，提出了基于决策空间划分的全局搜索策略。在这种策略中，所有准确评估的点被聚到一组聚类中。因此，可以根据形成的聚类将决策空间划分为若干个区域。然后，在每个区域中构建代理模型。该算法将在建立的代理模型的帮助下同时搜索这些区域。结果可以得到分布在不同区域的几个有希望的点。在局部搜索阶段，将模型自适应选择策略与信任域局部搜索相结合。引入模型自适应选择策略，从局部代理模型池中自适应选择局部精英代理模型，精确地辅助信任域局部搜索。对基准问题和调频声波问题参数估计的实验结果表明，DSP-SAEA算法与现有的一些算法相比具有竞争力。
In expensive optimization, function evaluations are based on expensive physical experiments or time consuming simulations. Moreover, the gradient for the objective is not readily available. Therefore, it is a challenge task to deal with expensive optimization. In this work, a decision space partition based surrogate-assisted evolutionary algorithm (DSP-SAEA) is proposed for expensive optimization. In DSP-SAEA, a two-stage search strategy is introduced, where the global search and the local search are seamlessly integrated. In the global search stage, a decision space partition based global search strategy is proposed. In this strategy, all the exactly evaluated points are clustered into a set of clusters. Thus, the decision space can be partitioned into several regions based on the formed clusters. Furthermore, in each region, the surrogate model is constructed. The algorithm will search for these regions simultaneously with the help of the built surrogate models. As a result, several promising points distributed in different regions are able to be obtained. In the local search stage, a model adaptive selection strategy and the trust region local search are integrated. The model adaptive selection strategy is introduced to accurately assist the trust region local search, where the local elite surrogate model is adaptively chosen from the local surrogate model pool. Experimental results on benchmark problems and the parameter estimation for frequency-modulated sound waves problem demonstrate that DSP-SAEA performs competitively compared with some state-of-the-art algorithms.