多层前向神经网络的RLS修正训练算法

AModified RLSAlgorithm for Multilayer Feedforward NeuralNetworks

文献［４］提出一种训练多层前向神经网络的快速学习算法—ＲＬＳ算法，与标准ＢＰ算法相比有较高的学习效率，但该方法的主要缺陷是存在数值稳定问题和鲁棒性不强的问题。提出了一种修正的基于递推最小二乘算法（ＲＬＳ）的多层前向神经网络的快速学习算法，证明了算法的数值稳定性，对两个系统进行了辨识，并与ＲＬＳ训练算法和标准ＢＰ算法进行了比较，仿真结果显示了所提方法的鲁棒性和有效性。

Paper [4] proposes a quick learning algorithm —the recursive least square (RLS), a m ethod for the m ultilayerfeedforward neuralnetw orks. Com pared with standard BPalgorithm , this algorithm has a high learning efficiency, w hile it's stablization and robustness are not high. This paper proposes a m odified learning algorithm based on theRLSm ethod, and dem onstratesthe m odified algorithm 'sstablization.The algorithm hasbeen applied to the identifications of two system s. The sim ulation results dem onstrate the robustness of the system and the effectivenessofthe proposed m ethod.