一种局部回归神经网络的在线学习算法  

An on-line learning algorithm of neural networks of local recurrent

作　　者：江小平[1] 姚天任[1] 

机构地区：[1]华中科技大学电子与信息工程系,湖北武汉430074

出　　处：《华中科技大学学报（自然科学版）》2005年第5期1-3,共3页Journal of Huazhong University of Science and Technology(Natural Science Edition)

摘　　要：针对目前局部回归神经网络误差函数在线计算复杂的缺陷,利用信号流图(SFG)基本理论,通过分析信号流图(SFG)和转置信号流图(ASFG),将神经网络的误差导数的信号流图(SFG)和转置信号流图(ASFG)分别级联在原始信号流图(SFG)和转置信号流图(ASFG)上,构成单输出自回归神经网络.依据因果非线性时变系统流图计算仅仅与网络拓扑结构有关的理论,推导了一种与网络结构无关的在线后向BP学习算法,较好地解决了对任意结构的局部回归神经网络的在线学习问题.仿真结果表明了本算法的有效性.Using the signal-flow-graph (SFG) basic theories and analyzing the SFG and adjoint signal-flow-graph (ASFG), we connected the SFG of the gradient of the cost function and its ASFG respectively to the original SFG and ASFG, adjoined the overall systems to a target SFG, constituting a kind of alone-oupput recurrent neural networks which depends only on the functional relationships between the branch variables. The simulations of non-linear dynamic systems identification also were presented to assess the performance of the algorithm in the application of local recurrent neural networks.

关 键 词：局部回归神经网络 在线BP算法 信号流图 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]