Large scale optimization by differential evolution with landscape modality detection and a diversity archive
Large scale optimization by differential evolution with landscape modality detection and a diversity archive
复制标题
DOI：
10.1109/cec.2012.6252911
复制
发表时间：
2012-06
期刊：
影响因子：
--
通讯作者：
T. Takahama;S. Sakai
中科院分区：
文献类型：
--
作者：
T. Takahama;S. Sakai
本研究报告了基于景观模态检测和多样性档案（LMDEa）的差分进化方法在CEC2012大规模全局优化特别会议上提供的基准函数集上的性能。差分进化算法为了保持搜索的多样性，采用了比待解问题中决策变量数量大得多的种群大小。然而，采用如此大的规模来解决大规模的优化问题是困难的，因为人口规模会变得过大，搜索效率会下降。在本研究中，我们提出使用小种群规模和大的多样性档案来解决大规模的优化问题。此外，我们还提出了通过观察搜索点的景观形态来控制比例因子，以保持多样性。要优化的问题的格局通常是未知的，并且在搜索过程进行时格局是动态变化的。在LMDEa中，对连接搜索点质心和搜索点的直线上的一些点进行采样。当采样点的客观值变化先减小后增大时，认为存在一个谷。如果只有一个山谷，景观是单峰的，采用较小的比例因子。否则，采用较大的比例因子。采样点在所有搜索点所跨越的区域内实现全局搜索，在最佳搜索点附近实现局部搜索。通过对基准函数的求解，验证了该方法的有效性。
In this study, the performance of Differential Evolution with landscape modality detection and a diversity archive (LMDEa) is reported on the set of benchmark functions provided for the CEC2012 Special Session on Large Scale Global Optimization. In Differential Evolution (DE), large population size, which is much larger than the number of decision variables in problem to be solved, is adopted in order to keep the diversity of search. However, it is difficult to adopt such large size to solve large scaled optimization problems because the population size will become too large and the search efficiency will degrade. In this study, we propose to solve large scale optimization problems using small population size and a large archive for diversity. Also, we propose simple control of scaling factor by observing landscape modality of search points in order to keep diversity. The landscape of a problem to be optimized is often unknown and the landscape is changing dynamically while the search process proceeds. In LMDEa, some points on a line connecting the centroid of search points and a search point are sampled. When the objective values of the sampled points are changed decreasingly and then increasingly, it is thought that one valley exists. If there exists only one valley, the landscape is unimodal and small scaling factor is adopted. Otherwise, large scaling factor is adopted. Also, the sampled points realize global search in the region spanned by all search points and realize local search near the best search point. The effect of the proposed method is shown by solving the benchmark functions.