论文部分内容阅读
提出了一种适用于神经网络控制器在线学习的BP算法,它用共扼梯度因子λ确定学习的方向,通过奖惩系数γ调整学习步长η,从而具有高速收敛性、快速跟踪能力及较强的鲁棒性。最后通过一个4层前馈NN网络对标准控制序列进行学习仿真,证实了该算法的优越性
A BP algorithm which is suitable for on-line learning of neural network controller is proposed. It uses the conjugate gradient factor λ to determine the learning direction and adjusts the learning step η through the reward-punishment coefficient γ, so that it has high-speed convergence and fast tracking ability Robustness. Finally, a 4-layer feed-forward NN network is used to simulate the standard control sequences, which proves the superiority of this algorithm