| 激励函数可调的NARX神经网络 |
| |
| 引用本文： | 李明,杨成梧. 激励函数可调的NARX神经网络[J]. 江南大学学报(自然科学版), 2006, 5(4): 445-448 |
| |
| 作者姓名： | 李明 杨成梧 |
| |
| 作者单位： | 南京理工大学,动力学院,江苏,南京,210094 |
| |
| 摘 要： | 提出了一种隐层神经元激励函数可调的具有外部输入的非线性回归(NARX)神经网络,它在进行权值调整的同时,还对各隐层神经元激励函数的参数进行自适应调节;并推导出激励函数参数的学习算法,从而使NARX神经网络更符合生物神经网络.通过系统辨识的仿真实例,说明了隐层神经元激励函数对网络性能的影响,还验证了文中提出的NARX神经网络具有更快的收敛速度,并且能有效地避免算法陷入局部最小.
|
| 关 键 词： | 具有外部输入的非线性回归神经网络 激励函数 学习算法 |
| 文章编号： | 1671-7147（2006）04-0445-04 |
| 收稿时间： | 2006-01-16 |
| 修稿时间： | 2006-04-08 |
NARX Networks with Trainable Activation Functions |
| |
| LI Ming,YANG Cheng-wu. NARX Networks with Trainable Activation Functions[J]. Journal of Southern Yangtze University:Natural Science Edition, 2006, 5(4): 445-448 |
| |
| Authors: | LI Ming YANG Cheng-wu |
| |
| Affiliation: | College of Power Engineering, Nanjing University of Science and Techonology, Nanjing 210094, China |
| |
| Abstract: | The paper proposes a new kind of NARX network, which has trainable activation functions in addition to only trainable weights in the conventional NARX networks. A learning algorithm based on first-order gradient descent was used to adjust the parameters of activation functions. The simulation results show that the shape of activation functions has large effect on the performs of NARX networks, and the new NARX network has a faster convergence rate. |
| |
| Keywords: | NARX network activation function learning algorithm |
| 本文献已被 CNKI 维普 万方数据 等数据库收录！ |
|
