一种快速收敛的核CMAC神经网络模型  

Kernel CMAC Model with Fast Convergence

作　　者：孔敏[1] 朱大奇[1] 

机构地区：[1]江南大学控制科学与工程研究中心,江苏无锡214122

出　　处：《系统仿真学报》2006年第7期1938-1941,共4页Journal of System Simulation

基　　金：江苏省自然科学基金(BK2004021);教育部科学技术研究重点项目(105088)

摘　　要：为了提高CMAC(cerebellarmodelarticulationcontrollers)神经网络实时在线学习的快速性和准确性,在核CMAC的基础上引入了信度分配的概念。在核空间内,利用激活单元先前学习次数作为可信度,其误差校正值与激活单元先前学习次数的倒数成比例,设计了一种基于信度分配的核CMAC的训练规则,使网络的学习速度和网络逼近精度及建模能力得到提高。仿真结果表明基于信度分配的核CMAC结合了核CMAC与信度分配CMAC的优点,隐去了映射所带来的复杂运算,有较快的学习速度和训练精度及建模能力。In order to improve the speed and accuracy of the on-line learning neural network, a credit assignment concept based on kernel CMAC was proposed, where the credit was supposed to be the activated hypercubes previous learning times in the kernel space. The correcting amounts of errors are proportion to the inversion of the learned times of the addressed hypercubes. With this idea, the CMAC training regulation was designed in kernel space to improve the network learning speed and accuracy, and improve the modeling capability. The simulation result shows that the improved CMAC neural network in kernel space is faster and more accurate than the conventional CMAC.

关 键 词：CMAC神经网络 核CMAC 信度分配 可信度 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]