一种改进的神经网络增量学习算法

A Modified Neutral Network Increment Study Algorithm

基于扩展KALMAN滤波器(Extended Kalman Filter)的神经网络是一类应用广泛的神经网络算法,但该算法在大数据量、抵抗噪声等方面还有相当的缺陷。本文从增量学习的角度出发,对扩展KALMAN滤波器算法进行了改进,同时借鉴周期算法的长处,引入部分增量训练机制(Partial incremental Training)和适当的隐层节点删减机制,使该算法在抵抗噪声等方面有了显著的提高。理论分析表明,该算法可以有效降低噪声数据的影响,提高神经网络算法的鲁棒性。

Neutral network based on extended Kalman filter is a widely applicable neutral network algorithm though it has problems in dealing with large data quantity and resisting noise. In this essay, the extended Kalman filter algorithm is modified from the point of incremental studying. The partial incremental training and proper obscure note cutting mechanism is introduced into the algorithm, considering the advantages of cycle algorithm, which makes it improved largely in resisting noise. Theoretical analysis manifests that the algorithm can reduce the impact of data noise effectively and improve the robustness of neutral network algorithm.