TSK网络结构设计:一个基于二阶运算的局部反馈网络

按照局部反馈和易于解释两个原则,本文设计了一个带反馈的二阶TSK网络。网络的结论部分以输入成分的二项式[1,2]作为神经元的导出域。二项式是神经网络领域少见的运算模式,导出域通常为输入与权向量的内积。网络的结论部分借助于延迟反馈引入动态特性,属于局部反馈,所以动态导数计算简洁。TSK网络表征一组If-Then规则,本文网络较采用独立变换[3]和基于G—FGS神经元[4]的网络易于理解。系统辨识的仿真证明本文网络可以很好地辨识动态系统,表明TSK网络拓扑结构设计中使用二阶运算是可行的。

Structure Design of TSK network: an Local Recurrent Network Based on Quadric Operation

In order to achieve two promising properties: local recurrence and interpreting ability, a new topology of recurrent TSK is presented in this paper, whose conclusion part has quadric form. The local recurrence of the conclusion part results in simple dynamic derivative, which makes it very convenient for training procedure. The if-then rules interpreted from the new network is more comprehensible than the other TSK networks that are based on independent transformations or G-FGS neurons. Simulation proves the favorable identification ability of our network for dynamic system, which means quadric operation is one feasible mode in the structure design of recurrent TSK networks.