一种基于扩展Kalman滤波器的神经网络学习算法

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 ,同时还保证了一定的泛化能力

Learning algorithm for neural networks based on extended Kalman filter

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.