基于Kalman滤波的基本Elman网络训练新算法

A new training algorithm based on Kalman filter of essential Elman network

Kalman滤波算法应用于基本Elman网络学习时,收敛速度较快,但收敛精度往往不高;而基于梯度下降的BP算法可以以很高的精度实现输入输出的非线性映射,但在极值点处收敛速度缓慢.针对上述问题,提出一种将Kalman滤波算法应用于基本Elman网络的新学习训练算法.该算法结合Kalman滤波算法和基于梯度下降的BP算法的优点来训练网络,以基本Elman网络隐层单元输出作为非线性系统的状态变量,通过Kalman滤波算法实现状态变量的快速准确跟踪,然后通过梯度下降法修正权值以保证精度.另外,在训练过程中,通过增加训练样本的信息内容来提高网络收敛的精度.仿真结果表明了该算法的有效性.

In the application of essential Elman network, the Kalman filter algorithm has high convergence speed but low precision; on the contrary, the BP algorithm with gradient descend technique has high precision to map the output with input nonlinearly, but usually with low convergence speed at extreme point. Therefore, a new algorithm based on Kalman filter is proposed to accomplish essential Elman network training. The advantages of Kalman filter and BP algorithm are combined to optimize the new algorithm. The outputs of the hidden layer of the essential Elman network are set as the state varibles of a nonlinear system. Kalman filter algorithm is used for rapid tracking of the state variables and BP algorithm is used for more precise weights modification. Furthermore, some new learning information is added to improve the precision of the results during the learning process. The validity of the new algorithm is supported by the simulation results.