计算机工程与应用 ›› 2007, Vol. 43 ›› Issue (24): 69-71.

• 学术探讨 • 上一篇    下一篇

基于比例UKF的神经网络及其应用

黄冬民   

  1. 西北工业大学,西安 710072
  • 收稿日期:1900-01-01 修回日期:1900-01-01 出版日期:2007-08-21 发布日期:2007-08-21
  • 通讯作者: 黄冬民

Neural network and its application based on the Scaled Unscented Kalman Filter(Scaled-UKF)

HUANG Dong-min   

  1. Northwestern Polytechnical University,Xi’an 710072,China
  • Received:1900-01-01 Revised:1900-01-01 Online:2007-08-21 Published:2007-08-21
  • Contact: HUANG Dong-min

摘要: 提出了一种利用比例无轨迹卡尔曼滤波(Scaled-UKF)进行神经网络权值估计的算法,该算法可以克服BP算法存在的学习速率缓慢、计算量大、容易使学习陷入局部极小等缺点。以Mackey-Grass混沌时间序列作为神经网络输入,运用比例UKF算法、UKF算法、BP算法仿真神经网络。结果表明,比例UKF算法较之BP算法具有更快的训练速度和更高的预测精度,且可以避免网络学习陷入局部极小;而相对于UKF算法,其变量分布可不限定为高斯型且能保证状态方差半正定。

关键词: 比例UKF, 神经网络, Mackey-Grass, 预测

Abstract: One algorithm based on the Scaled Unscented Kalman Filter(Scaled-UKF) is proposed to estimate the weights of the neural network,which can overcome the BP algorithm’s weaknesses of slow learning speed,large computational complexity,and easy convergence to the local minimum points.Taking the Mackey-Grass chaos time sequences as its input,the neural network is simulated with the Scaled-UKF,UKF and BP algorithm.The result of the simulation indicates that the Scaled-UKF algorithm has the faster training speed and higher forecast precision than the BP algorithm,and may avoid the network’s convergence to the local minimum points.Comparing with the UKF algorithm,the Scaled-UKF algorithm can guarantee positive semi-definiteness of the state covariance and its variable distribution may not be Gaussian-distributed.

Key words: Scaled Unscented Kalman Filter(Scaled-UKF), neural network, Mackey-Grass, forecast