求解时变凸二次规划的变参积分动态学习网络  

Novel Varying-parameter Integral Dynamic Learning Network for Solving Constrained Time-varying Convex QP Problem

作　　者：陆荣秀 黄伟 杨辉 张智军 LU Rong-xiu;HUANG Wei;YANG Hui;ZHANG Zhi-jun(School of Electrical and Automation Engineering,East China Jiaotong University,Nanchang 330013,China;Key Laboratory of Advanced Control&Optimization of Jiangxi Province,Nanchang 330013,China;School of Automation Science and Engineering,South China University of Technology,Guangzhou 510640,China)

机构地区：[1]华东交通大学电气与自动化工程学院,南昌330013 [2]江西省先进控制与优化重点实验室,南昌330013 [3]华南理工大学自动化科学与工程学院,广州510640

出　　处：《小型微型计算机系统》2022年第9期1853-1861,共9页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61863014,61733005,61963015)资助;国家重点研发计划项目(2020YFB1713700)资助;江西省自然科学基金项目(20171ACB21039、20192BAB207024)资助。

摘　　要：凸二次规划问题在许多领域都有着广泛的应用,系统分析、组合优化等诸多科学问题和工程问题都可以表述为二次规划问题后求解.为了高效地求解线性等式约束下的时变凸二次规划问题,本文提出了一种基于误差积分设计的新型变参积分动态学习网络.该网络引入时变指数型设计参数,具有灵活和自适应调整的特点.理论分析证明变参积分动态学习网络在使用单调递增的奇激活函数时有全局收敛性质和较强的鲁棒性.除此之外,变参积分动态学习神经网络还具有灵活的控制策略和超指数级收敛速率.仿真实验结果表明,使用不同激活函数的变参积分动态学习网络比传统的微分神经网络(即梯度神经网络,零化神经网络和变参收敛微分神经网络)有更好的收敛性质.The convex quadratic programming(QP)problem is always the focus of engineering research.Many scientific and engineering problems,such as system analysis and combinatorial optimization,can be expressed as quadratic programming problems to solve.In this paper,according to an error-integral design equation,a novel varying-parameter integral dynamic learning network(VP-IDLN)is proposed and analyzed for solving time-varying convex QP problem constrained by a time-varying linear-equality.The VP-IDLN has time-varying exponential design parameters to obtain the flexibility and adaptive adjustment characteristics.Theoretical analysis verifies the VP-IDLN possess global convergence performance and strong robustness when using monotonically increasing odd activation functions.In addition,the VP-IDLN also has flexible control strategy and super exponential convergence rate.Simulink experiment results verifies that the VP-IDLN with different activation functions can have better convergence property than traditional differential neural networks,such as the gradient-based neural network(GNN),the zeroing neural network(ZNN)and the varying-parameter convergent differential neural network(VP-CDNN).

关 键 词：变参积分动态学习网络 时变问题求解 二次规划 收敛性 鲁棒性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]