摘要
为了增强离线数据驱动的进化算法在小数据情景中的表现,削弱代理模型对数据集规模的依赖,提出基于多核数据合成的离线小数据驱动的进化算法(DDEA-MKDS).考虑到代理模型易因小数据陷入过拟合,通过经验公式与遍历法找出针对离线数据集的最优隐含层节点数,以简化模型结构.为了弥补数据量的不足,训练了3个不同核函数的径向基网络生成合成数据,通过轮盘赌法选择其中的部分数据与原数据集合并,使用新数据集训练代理模型.将DDEA-MKDS与其他5种流行的离线数据驱动的进化算法在6个单目标基准测试问题上进行对比,实验结果表明,所提算法在数据量极小的条件下能够取得良好的效果,寻优效率显著优于其他算法.
An offline data-driven evolutionary algorithm based on multi-kernel data synthesis(DDEA-MKDS)was proposed to enhance the performance of such algorithms in small data scenarios and weaken the dependence of surrogate model on data set size.The empirical formula and traversal method were used to calculate the optimal number of hidden layer nodes for the offline data set in order to simplify model structure by considering that the surrogate model is prone to overfitting due to small data.Three radial basis networks with different kernel functions were trained to generate synthetic data in order to make up for the lack of data.A part of synthetic data was selected by roulette to combine with original data,and new data set was used to train the surrogate model.The experimental results showed that DDEA-MKDS had good performance under the condition of small data by comparing with five states of the art offline data-driven evolutionary algorithms on six single objective benchmark problems,and its efficiency was significantly better than other algorithms.
作者
李二超
刘昀
LI Erchao;LIU Yun(College of Electrical and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China)
出处
《浙江大学学报(工学版)》
北大核心
2025年第2期278-288,共11页
Journal of Zhejiang University:Engineering Science
基金
国家自然科学基金资助项目(62063019)
甘肃省自然科学基金资助项目(24JRRA173,22JR5RA241)。
关键词
离线数据驱动
进化算法
小数据
代理模型
隐含层节点
合成数据
offline data-driven
evolutionary algorithm
small data
surrogate model
hidden layer node
synthetic data
