基于迭代容积卡尔曼滤波的神经网络训练算法-计算机科学.pdf

Vo 1. 43 No. 10 第 43 卷第10 期 计算机科学 2016 年 10 月 Computer Science Oct 2016 基于迭代容积卡尔曼滤波的神经网络训练算法 袁光耀1 胡振涛1 张谨1 赵新强1 付春玲2 (河南大学图像处理与模式识别研究所 开封 475004)1 (河南大学物理与电子学院 开封 475004)2 摘 要 针对现有应用非线性滤波算法对神经网络进行训练时存在精度不足的问题，提出了一种基于迭代容积卡尔 曼滤泼的神经网络训练算法。首先，将前馈神经网络各个节点的连接权值和偏置作为状态向量，建立前馈神经网络的 状态空间模型。其次，利用 Spherical-Radial 准则生成容积点，并依据 Gauss-Newton 迭代策略来优化量测更新过程中 获取的状态估计值和状态估计误差协方差，通过容积卡尔曼滤泼估计精度的改善，提升神经网络节点的连接权值和偏 置的训练效果。理论分析和仿真实验结果验证了所提算法的可行性和有效性。 关键词 前馈神经网络，状态空间模型，容和、卡尔曼i虑泼，Gauss-N ewton 迭代 中固法分类号 TP391 文献标识码 A DOI 10. 11896/j. issn. 1002-137)♀ 2016.10.048 Novel Neural Network Training Algorithm ßased on Iterated Cubature Kalman Filter YUANGl即19-yao1 HU Zhen-tao 1 ZHANG Jin 1 ZHAO Xin-qiang 1 FU Chun-lini CInstitute of Image Processing and Pattern Recognition ,Henan University ,Kaifeng 475004 ,China)1 (School of Physics and Electronics ,Henan University ,Kaifeng 475004 ,China) 2 Abstract In view of insufficient accuracy in the existing application of nonlinear filtering algorithm for neural network training ,a novel neural network training algorithm based on iterated cubature Kalman filter was proposed. Firstly ,the connection weights and bias of feedforward neural network are used as the state vector to establish the state space mσ del. Secondly ,using the Spherical-Radial standard to generate cubature points ,the state estimation and covariance ac- quired during the measurement update process are optimized based on Gauss-Newton iteration strategy. The training effect of neural network connecting weights and bias