基于信任度分配的小脑模型节点控制器改进算法及其收敛性分析  

A Credit-Assignment CMAC Algorithm and Analysis on Its Convergence

作　　者：张蕾[1] 曹其新[1] 李杰[2] 张春余[1] 

机构地区：[1]上海交通大学机器人研究所,上海200030 [2]威斯康星大学智能维护中心,密尔沃基wi53224

出　　处：《上海交通大学学报》2005年第3期377-380,385,共5页Journal of Shanghai Jiaotong University

基　　金：国家自然科学基金资助项目(50128504)

摘　　要：针对传统小脑模型节点控制器(CMAC)算法中的学习干扰现象,给出了一种基于信任度分配的CMAC改进算法(CA-CMAC).该算法将每个存储单元被激活次数的倒数作为该单元的信任度,误差的分配与该单元的信任度成正比.然后提出了信任度矩阵和信任度关联矩阵的概念,并根据线性方程组迭代理论,证明了改进算法在增量学习时的收敛性,给出了收敛条件并进行了验证.通过二自由度平面机器人臂逆动力学求解的仿真,比较了CA-CMAC与传统CMAC的性能,结果表明,CA-CMAC具有更快的收敛速度.To reduce learning interference in traditional cerebellar model articulation controller(CMAC), a Credit-Assignment CMAC(CA-CMAC)learning algorithm was proposed. The inverse of activated times of each memory cell is taken as the credibility, and the error correction is proportional to the credibility. Credit matrix and credit correlation matrix were presented. Based on the iteration theories of linear equations, the convergence property of CA-CMAC in incremental learning was analyzed. Convergent condition was given and validated. The simulation results of the inverse kinematics of 2-DOF planar robot arm prove that CA-CMAC converges more rapidly than traditional CMAC.

关 键 词：小脑模型节点控制器 信任度分配 学习干扰 收敛性 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]