一种新型解决非光滑伪凸优化问题的神经网络方法  

New Neural Network Method for Solving Nonsmooth Pseudoconvex Optimization Problems

作　　者：喻昕[1] 马崇[1] 胡悦[1] 伍灵贞 汪炎林 YU Xin;MA Chong;HU Yue;WU Ling-zhen;WANG Yan-lin(Department of Computer and Electronic Information,Guangxi University,Nanning 530004,China)

机构地区：[1]广西大学计算机与电子信息学院

出　　处：《计算机科学》2019年第11期228-234,共7页Computer Science

基　　金：国家自然科学基金项目(61862004,61462006)资助

摘　　要：优化问题的研究一直以来深受科研工作者的关注,凸优化问题作为优化问题的一个重要部分更是成为研究重点,许多应用神经网络思想提出的模型已经被应用到实际问题中。然而,在机器学习、信号处理、生物信息学等领域中涉及的优化问题往往不是凸优化问题,而是伪凸优化及非凸优化的问题,因此解决后一类问题变得刻不容缓。针对目标函数是非光滑伪凸函数、约束函数由等式和不等式函数构成的优化问题,基于罚函数以及微分包含的思想,构建了一个新型的不含惩罚参数的单层神经网络模型。该模型的主要设计思路是根据已经提出的神经网络模型思想,为目标函数的梯度设计一个制约的函数,使其值始终保持在一个范围之内,再结合一个关于时间的函数,确保其值随时间变小。同时,考虑到不等式约束对状态解进入等式约束之前的收敛方向有影响,加入一个条件函数来限制它。与已提出的神经网络模型相比,所提模型具有结构简单、无须提前进行惩罚参数的计算、对初始点的位置无特殊要求等优势。而且,对于任意初始点,理论证明了状态解的有界性、状态解能够在有限时间内收敛到等式约束内并永驻其中、状态解能够在有限时间内收敛到可行域并永驻其中以及状态解最终收敛到优化问题的最优解。在MATLAB环境下,通过数学仿真实验,状态解都能快速地收敛到一个最优解。同时,用已经提出的类似神经网络模型解决同样的优化问题时,若罚参数或初始点选择不恰当则会导致状态解不能很好地收敛。这不仅验证了所提出的理论结果的正确性,同时也说明了所提网络具有更广泛的应用范围。The research of optimization problem is favored by researchers.As an important part of optimization pro-blem,convex optimization problem is the focus of research.Many models based on neural network are applied to practical problems.However,the optimization problems involved in machine learning,signal processing,bioinformatics and other fields are often not convex optimization problems,but pseudoconvex optimization and nonconvex optimization problems.Therefore,it is urgent to solve the latter kind of problems.To solve the optimization problem that the objective function is nonsmooth pseudoconvex function and constraint function is equality and inequality function,this paper constructed a new single-layer neural network model without penalty parameter based on the idea of penalty function and differential inclusion.The main idea of the design is that according to the proposed neural network model,a constrained function can be designed for the gradient of the objective function so that the value of the objective function is always kept within a range,and then a function about time is combined to ensure that its value decreases with time.At the same time,considering that inequality constraints affect the convergence direction of the state solution before it enters the equation constraint,a conditional function is added to restrict it.Compared with the proposed neural network model,it has the advantages of simple structure,no need to calculate penalty parameters in advance,and no special requirements for the position of the initial point.Furthermore,it is theoretically proved that for any initial point,the state solution can converge to the equality constraints in finite time and stay there thereafter,the boundedness of the state solution,the state solution can converge to the feasible region in finite time and stay there thereafter,and the state solution can finally converge to the optimal solution of the optimization problem.Under the environment of MATLAB,by mathematical simulation experiments,the state solution ca

关 键 词：伪凸函数 神经网络 微分包含 非光滑优化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]