摘要
为了克服一般神经网络学习方法易陷入局部极小值的缺陷,提出一种新的基于粒子滤波的神经网络学习算法.采用无迹卡尔曼滤波(Unscented Kalman Filter,UKF)产生粒子,以较少的粒子逼近状态的后验概率分布,搜索到经验风险函数的最小值.此方法适用于在线的、非线性的、非高斯的神经网络学习.仿真结果表明,该学习方法与同类方法相比,性能明显提高.
To overcome the weakness that current learning algorithms of Neural Network (NN) is likely to trap in local minimum, a novel algorithm for NN training based on particle filter is proposed. Unscented Kalman filter (UKF) is utilized to select particles, rendering smaller number of particles to approximate the posterior probability distribution, thus converging at global minimum of experimental loss function. The proposed algorithm is well suited to applications involving on-line, nonlinear and non-Gaussian signal processing. Simulation results show that its performance is markedly superior to those available.
出处
《武汉大学学报(工学版)》
CAS
CSCD
北大核心
2006年第6期86-88,共3页
Engineering Journal of Wuhan University
基金
航天"十五"预研基金项目(413160203)
关键词
粒子滤波
神经网络学习
UKF
particle filter
neural network learning
unscented Kalman filter