基于无先导卡尔曼滤波的RBFN训练算法研究

提出了应用无先导卡尔曼滤波器（UKF）来训练径向基神经网络（RBFN）的新方法。与广义卡尔曼滤波器（EKF）和双重卡尔曼滤波器（DEKF）对函数的一阶近似不同，UKF对非线性函数采用二阶近似展开，而且最重要的一点是不必求取系统的雅克比矩阵，从而大大减小计算量。本文对时间序列预测及分类问题进行了仿真，结果证实了该方法的有效性和快速性。

Training radial basis neural networks with the unscented Kalman filter

A new method is proposed for training radial basis function networks (RBFN) using the unscented Kalman filter (UKF). In contrast to the extended Kalman filter (EKF) and the dual extended Kalman filter (DEKF), which extend the nonlinear functions using a first order approximation, the UKF uses a second order approximation. The most important consequence is that the algorithm does not require the Jacobi matrix of the system to be calculated, thus reducing the calculation complexity and resulting in considerable savings in time. Simulation results in the fields of chaotic time series prediction and classification problems demonstrate both the validity and faster speed of the proposed method.