一种新的快速小波网络学习算法  

New and Quick Learning Algorithm of Wavelet Network

作　　者：耿道华[1] 徐桂云[1] 王恩元[1] 

机构地区：[1]中国矿业大学

出　　处：《煤矿机械》2006年第9期81-83,共3页Coal Mine Machinery

基　　金：江苏省博士后科研基金资助项目(2004035);中国矿业大学科技基金资助项目(2005B005)

摘　　要：针对小波网络学习算法具有训练时间长、收敛缓慢,以及训练时频繁调整隐含层与输出结点之间的权重,难以确定合适的步长等缺点,提出了一种把梯度下降方法和Kalman方法有机结合的快速学习算法。该算法使用梯度下降方法调整尺度和平移系数,用Kalman方法调整权重,以动态非线性系统和混沌系统为实例做了仿真,并与其它方法做了比较。结果表明该算法能够对动态非线性系统的输入输出快速学习和建模,优于其它小波网络的学习算法。At present the learning algorithm of wavelet network has many disadvantages, such as long training time, slow convergence, frequent modification of weights between hidden layer and output layer and hard determination of right step. Aiming at those, a fast learning algorithm which combines gradient descent method with Kalman method is put forward. This algorithm uses gradient descent method to adjust the measure and the pan coefficient and uses Kalman method to adjust weights. In the end, take dynamic nonlinear system and chaos system as examples to emulate and compare it to other methods. The results show that this algorithm can model input and output learning kernel of dynamic nonlinear system quickly, which is superior to other learning methods of wavelet network.

关 键 词：小波网络 梯度下降 卡尔曼滤波 反向传播算法 

分 类 号：TB115[理学—数学]