摘要
对复合输入动态递归网络作了改进 ,提出一种新的动态递归神经网络结构 ,称为状态延迟输入动态递归神经网络 (State Delay Input Dynamical Recurrent Neural Networks)。这种具有新的拓扑结构和学习规则的动态递归网络 ,不仅明确了各权值矩阵的含义 ,而且使权值的训练过程更为简洁 ,意义更为明确。网络增加了输入输出层前一步的状态信息 ,使其收敛速度和泛化能力与其他常用网络结构相比 ,均有明显提高 ,增强了系统实时控制的可能性。本文将该网络用于机器人定位监督控制系统中 ,通过利用神经网络建立起被控对象的逆模型 ,与传统 PD控制器结合 ,确保了控制系统的稳定性 ,有效地提高系统的精度和自适应能力。
A new neural network model named State Delay Input Dynamical Recurrent Neural Network (SDIDRNN) is presented. The model with new topological structure and learning algorithm has explicit significance for weight matrixes and makes training process of weights become more distinct and straightforward. Speed of learning and convergence and ability of generalization are improved by inputting the prior state knowledge of nodes in input layer and output layer, which make it possible for real time control. In this paper, the new neural network is applied to the system of supervisory control for position control of robot systems. By establishing inverse model of the controlled object with the new network and combining it with conventional PD controller, the stability or robustness of the system is ensured, and precision and adaptability is improved effectively. Simulation results show the efficiency and superiority of the new neural network.
出处
《机械科学与技术》
CSCD
北大核心
2003年第2期229-232,共4页
Mechanical Science and Technology for Aerospace Engineering
基金
国家自然科学基金项目 (5 9975 0 0 1)
北京市自然科学基金项目 (30 12 0 0 3)资助
关键词
动态递归网络
监督控制
机器人
Dynamical recurrent neural networks
Supervisory control
Robot
