【作者】 鞠宪龙；

【导师】 沈艳；

【作者基本信息】 哈尔滨工程大学 ， 系统理论， 2011， 硕士

【摘要】 递归神经网络学习算法一直是神经网络方向研究的热点,并且其应用也引起了科研爱好者的广泛关注。动态递归神经网络由于内部有自反馈,表现出很强的动态映射能力。目前训练二阶对角递归神经网络多采用DBP算法,本文针对该算法中的辨识精度和收敛速度等问题做进一步的深入研究和探讨。首先,详细地介绍了递归神经网络结构,并且给出神经网络系统辨识的基本原理和网络辨识模型。其次,针对训练二阶对角递归神经网络采用的梯度搜索算法中存在的问题,提出改进的梯度下降学习算法,并且给出了这种改进算法的收敛性证明。然后,给出训练二阶对角递归神经网络的三种算法：DBP算法、改进DBP算法和RPROP算法和实现步骤,并将三种算法训练后的二阶对角递归神经网络用于非线性系统辨识,仿真结果表明：基于DBP算法的辨识精度不够理想且收敛速度慢,针对该算法存在的问题,采用了改进的DBP算法,辨识效果好于DBP算法；但是DBP算法和改进的DBP算法受梯度大小的影响较大。最后,首次将RPROP算法引入用于训练二阶对角递归神经网络的权值,仿真结果表明：将RPROP算法与改进的DBP算法和DBP算法相比,RPROP算法的非线性系统辨识精度和收敛速度都要优于DBP算法和改进的DBP算法。更多还原

【Abstract】 The learning algorithm of recurrent neural network has always been the hot spot of the neural network research, and its application also arouse entensive attention from research enthusiasts. Because dynamic recurrent neural network has the internal self-feedback, the result shows the highly dynamic capability. At present, second-order diagonal recurrent neural network is almost trained by using DBP algorithm, in this paper we will make the further study and discussion about the issues of identification accuracy and convergence speed in gradient descent algorithm.First, the paper introduces the structure of recurrent neural network in detail; and the basic principle of the neural network system identification and network identification model is given.Secondly, as the issue for the gradient search algorithm in second-order diagonal recurrent neural network, an improved gradient descent learning algorithm is proposed, and gives the convergence of improved algorithm.Thirdly, the realization process of DBP algorithm, improved DBP algorithm and RPROP algorithm are given. when identifying nonlinear systems based on second order diagonal recurrent neural network, The simulation results show the identification accuracy and convergence speed based on the DBP algorithm is bigger and slow, in view of the faults of algorithm, improved DBP algorithm is adopted, identification effect is better than the DBP algorithm. But both DBP algorithm and improved DBP algorithm are largely influenced by the gradient.Finally, primarily, RPROP algorithm is applied to train the SDRNN, the simulation results show that compared with the DBP algorithm and improved DBP algorithm, RPROP algorithm for nonlinear system in identification accuracy and convergence speed are superior to the DBP algorithm and improved DBP algorithm.更多还原