一种基于自适应粒子滤波的多层感知器学习算法
来源期刊:中南大学学报(自然科学版)2013年第4期
论文作者:席燕辉 叶志成 彭辉
文章页码:1397 - 1402
关键词:多层感知器;粒子滤波;自适应粒子滤波
Key words:multi-layer perceptrons; particle filter; adaptive filter
摘 要:针对神经网络状态空间模型中系统噪声统计特性未知导致滤波发散或者滤波精度不高的问题,提出一种自适应的粒子滤波神经网络训练算法。该算法用粒子滤波估计网络的权时,利用序贯更新先验信息的序贯可信度最大化方法在线估计未知系统噪声方差。仿真结果表明:该自适应粒子滤波算法明显优于其他传统的神经网络训练算法,如扩展卡尔曼滤波、噪声可调的扩展卡曼滤波、普通粒子滤波等。
Abstract: To overcome the problems such as low filtering accuracy and divergence caused by unknown system noise statistics in state space neural network model estimation, an adaptive particle filter (APF) is proposed. By applying the sequential to estimate the variance of unknown system noise online, the particle filter is used to estimate the weights of the multi-layer perceptrons. The simulation results show that the APF algorithm outperforms the some conventional training algorithms such as the extended kalman filter (EKF), the EKF algorithm with evidence maximization and sequentially updated priors (EKFQ), and the general particle filter.
席燕辉1, 2, 3,叶志成1, 3,彭辉1, 3
(1. 中南大学 信息科学与工程学院,湖南 长沙,410083;2. 长沙理工大学 电气与信息工程学院,湖南 长沙,410077;3. 先进控制与智能自动化湖南省工程实验室,湖南 长沙,410083)
摘 要:针对神经网络状态空间模型中系统噪声统计特性未知导致滤波发散或者滤波精度不高的问题,提出一种自适应的粒子滤波神经网络训练算法。该算法用粒子滤波估计网络的权时,利用序贯更新先验信息的序贯可信度最大化方法在线估计未知系统噪声方差。仿真结果表明:该自适应粒子滤波算法明显优于其他传统的神经网络训练算法,如扩展卡尔曼滤波、噪声可调的扩展卡曼滤波、普通粒子滤波等。
关键词:多层感知器;粒子滤波;自适应粒子滤波
XI Yanhui1, 2, 3, YE Zhicheng1, 3, PENG Hui1, 3
(1. School of Information Science and Engineering, Central South University, Changsha 410083, China;2. Electrical and Information Engineering College, Changsha University of Science and Technology, Changsha 410077, China;3. Hunan Engineering Laboratory for Advanced Control and Intelligent Automation, Changsha 410083, China)
Abstract:To overcome the problems such as low filtering accuracy and divergence caused by unknown system noise statistics in state space neural network model estimation, an adaptive particle filter (APF) is proposed. By applying the sequential to estimate the variance of unknown system noise online, the particle filter is used to estimate the weights of the multi-layer perceptrons. The simulation results show that the APF algorithm outperforms the some conventional training algorithms such as the extended kalman filter (EKF), the EKF algorithm with evidence maximization and sequentially updated priors (EKFQ), and the general particle filter.
Key words:multi-layer perceptrons; particle filter; adaptive filter