一种提高神经网络泛化能力的自适应UKF滤波算法

An Adaptive UKF Algorithms for Improving the Generalization of Neural Network

摘要: 给出了利用EKF(extended Kalman)滤波和UKF(unscented Kalman)滤波提高神经网络泛化能力的方法。针对UKF参数选取随意性的问题,采用移动开窗估计法对状态噪声和观测噪声协方差矩阵进行自适应估计,提出了一种新的提高神经网络泛化能力的自适应UKF算法。利用检测样本进行了验证,结果表明,利用EKF、UKF和自适应UKF算法训练神经网络都能提高其泛化能力,其中自适应UKF算法优于其他几种算法。

Abstract: Two methods of improving the generalization of neural network,using EKF and UKF algorithm,are described.In order to avoid the random selection of some parameters for UKF,a new adaptive UKF algorithm for training neural network,by using windowing residual vectors to adaptively estimate the covariance matrices of the observational vectors and the model errors,is established.The results show that EKF,UKF and adaptive UKF algorithms for training neural network can all improve the generalization of neural network,and the adaptive UKF algorithm is better than others.