摘要
基于GaussNewton法的前馈神经网络虽然可以达到局部二阶收敛速度,但网络结构中如果结点个数过多,会造成过模拟;网络结点过少,又会导致不收敛。为了优化神经网络结构,尝试引入重置算法(EarlyRestartAlgo rithm),并将其应用于GaussNewton前馈神经网络,提出基于重置的GaussNewton变结构前馈神经网络。对比实验表明,重置算法的引入有效地解决神经网络的结构优化问题,优化后的神经网络具有良好的收敛性与稳定性。
Feed forward neural network based on Gauss-Newton algorithm will converge with order two in local area, but it will be over-modified with excess neural nodes or not converge with insufficient neural nodes. In order to optimize the neural network structure, the Early Restart Algorithm is introduced and applied to the Gauss-Newton Feed Forward Neural Network. The comparative experiment results demonstrate that the Early Restart Algorithm can solve the structure optimization problem of Neural Network effectively, and the revised neural network performs well in convergence and stability.
出处
《南昌大学学报（理科版）》
CAS
北大核心
2004年第4期341-344,共4页
Journal of Nanchang University(Natural Science)
基金
国家自然科学基金资助项目(10271025)
关键词
重置算法
神经网络
结构优化
early restart algorithm
neural network
structure optimization
