面向复杂仿真元建模的序贯近邻探索实验设计方法  

Design method of sequential neighbor exploratory experimental for complex simulation metamodeling

作　　者：雷永林[1] 王言 于芹章 朱智 董微 朱一凡[1] LEI Yonglin;WANG Yan;YU Qinzhang;ZHU Zhi;DONG Wei;ZHU Yifan(College of Systems Engineering, National University of Defense Technology, Changsha 410073, China;The Key Laboratory for Complex Systems Simulation, Beijing 100010, China)

机构地区：[1]国防科技大学系统工程学院,湖南长沙410073 [2]复杂系统仿真总体重点实验室,北京100010

出　　处：《系统工程与电子技术》2021年第5期1232-1239,共8页Systems Engineering and Electronics

基　　金：国家自然科学基金面上项目(61273198)资助课题。

摘　　要：复杂仿真影响因素多,影响关系非线性,需要高效的元建模方法。元模型的建立不仅取决于所选定的元模型数学结构,还取决于仿真训练数据的充分性。数据充分性直接影响到仿真实验空间的设计与实验运行的计算复杂度,需要通过针对性的仿真实验设计方法来应对维数爆炸困境。序贯实验设计方法可以有效提高设计的效率,但传统上的序贯方法依赖于人工干预和关于目标系统的先验知识,不易自动化开展。因此,提出一种支持仿真元建模的序贯近邻探索实验设计方法,该方法综合运用梯度采样与随机近邻探索技术,在产生尽可能少的实验点的同时,能够自动提取关于实验对象的信息,有效提高复杂仿真元建模的实验效率。案例部分将所提方法与拉丁超立方抽样、传统序贯采样两种方法进行了对比,验证了所提方法的有效性。The complex simulation contains many influencing factors,and its influence relationship is nonlinear,which need an efficient meta modeling method.The establishment of the meta model depends not only on the mathematical structure of the selected meta model,but also on the sufficiency of the simulation training data.Data adequacy directly affects the design of simulation experiment space and the computational complexity of experiment operation,so it is necessary to deal with the dimension explosion dilemma through targeted simulation experiment design methods.Sequential experimental design method can effectively improve the efficiency of design,but the traditional sequential method relies on human intervention and prior knowledge about the target system,which is not easy to carry out automatically.Therefore,a sequential neighbor exploration experiment design(SNEED)method is proposed to support simulation meta modeling.SNEED combines gradient sampling and random nearest neighbor exploration technology to generate as few experimental points as possible and extract information about experimental objects automatically,which can effectively improve the experimental efficiency of complex simulation meta modeling.In the case part,the proposed method is compared with Latin hypercube and traditional sequential sampling methods to verify the effectiveness of the proposed method.

关 键 词：仿真元模型 序贯实验设计 近邻探索 拉丁超立方抽样 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]