摘要
在神经计算中神经网络的泛化特性是一个非常重要的内容．该文简述了小脑模型（ＣＭＡＣ——ＣｅｒｅｂｅｌａｒＭｏｄｅｌＡｒｅｉｃｕｌａｔｉｏｎＣｏｎｔｒｏｌｅｒ）的原理和学习算法，并用仿真方法讨论了在机器人使用的坐标变换关系（输入直角坐标值，输出机器手的关节角度）下ＣＭＡＣ的泛化性能：当泛化率为１∶１００时ＣＭＡＣ仍能正常工作．系统的精度虽能满足需要，但是进一步提高却受到限制．本文还讨论了影响精度的各种因素及可能的改进方法．
Generalization of neural network is a very important topic for coordinate transformation in neural computation. In this paper, we describe the principle of Cerebellar Model Articulation Controller(CMAC) including its learning algorithm, and discusse the generalization of CMAC through simulation of coordinate transformation (the input is position coordinate values and the output is articulation degrees of robot). The CMAC may still run well at generalization rate 1∶100. Several factors affecting the accuracy are also discussed.
出处
《自动化学报》
EI
CSCD
北大核心
1997年第4期475-481,共7页
Acta Automatica Sinica
基金
国家自然科学基金
"八六三"计划
攀登计划
关键词
泛化性能
小脑模型
CMAC
坐标变换
神经网络
Generalization, coordinate transformation, cerebellar model articulation controller (CMAC).
