动力学神经网络模型的构建及其稳定性研究  

Construction of dynamics neural network model and its stability study

作　　者：李立平[1] 韩兵欣[2] LI Liping;HAN Bingxin(Shijiazhuang Tiedao University Sifang College,Shijiazhuang 051132,China;Shijiazhuang Tiedao University,Shijiazhuang 050000,China)

机构地区：[1]石家庄铁道大学四方学院,河北石家庄051132 [2]石家庄铁道大学,河北石家庄050000

出　　处：《现代电子技术》2017年第3期167-170,共4页Modern Electronics Technique

基　　金：国家自然科学基金项目(60774050)

摘　　要：针对静态人工神经网络具有在反映系统动态行为时网络结构复杂、不能很好地反映系统动态性能的缺点,提出一种由带有积分器和可调反馈系数的神经元构成的新型动力学神经网络模型。该网络比以前的动态网络即递归网络或在此基础上改进的网络能更好地反映系统的动态性能,网络的结构更加简单,训练过程加快,从而使系统能够更好的运行。利用梯度下降法研究了该网络的权值调整算法,并通过李雅普诺夫稳定性判据讨论了这种新型动力学神经网络的稳定性条件。该网络研究为反映系统的动力学行为提供了更好的模型结构和理论算法,为神经网络的发展提供了新的研究方向。Since the static artificial neural network has the complicated network structure while reflecting on the system dy?namic behavior,and can′t reflect on the system dynamic performance better,a new dynamics neural network model composed of the neuron with adjustable feedback coefficient and integrator is proposed. The new neural network can better reflect on the system dynamic performance than the previous dynamic network(recursion network)or the network improved on the basis of it,has simpler network structure and faster training process to make the system run better. The gradient descent method is used to study the weight adjustment algorithm of the network. The stability condition of the new dynamics neural network is discussed ac?cording to the Lyapunov stability criteria. The study of the network provides a better model structure and theory algorithm for re?flecting on the system dynamics behavior,and a new research direction for the development of the neural network.

关 键 词：动力学神经网络 梯度下降法 李雅普诺夫稳定性判据 权值调整算法 

分 类 号：TN711.34[电子电信—电路与系统]