一种基于扩展Kalman滤波器的神经网络学习算法

Learning algorithm for neural networks based on extended Kalman filter

为了解决前馈神经网络BP算法在处理非线性对象时 ,收敛速度慢 ,易陷入局部极值 ,需调节参数多等的缺陷 ,提出将扩展卡尔曼滤波 (EKF)算法引入神经网络的学习中 .把前馈网络的所有权值、阈值作为EKF算法的状态 ,网络输出作为EKF的观测 .同时为了防止滤波发散 ,对算法做了改进 .仿真结果表明 ,该算法比BP算法在收敛速度、抗噪能力方面都有明显提高 。

Since back propagation (BP) algorithm is defective in rapidity of convergence an d apt to trap into local extreme value, and it also has too many parameters to b e adjusted when it is applied to nonlinear objects, an extended Kalman filtering (EKF) algorithm is presented and used for training artificial neural networks ( ANN). It regards all the weight values and threshold values as the states, a nd the outputs of the network as the observing values for the Kalman filter in t he feedforward networks. Furthermore, the EKF algorithm is improved to prevent d ivergence. Simulation results show that the EKF algorithm is evidently superior to BP algorithm in the rapidity of convergence, the ability of resisting noise and the ability of generalization.