基于小波采样理论的新型准则函数  

A Novel Cost Function Based on Wavelet Sampling Theory

作　　者：毛伟伟 张治国[2] 金晓宇 MAO Weiwei;ZHANG Zhiguo;JIN Xiaoyu(Fifth Department,Naval Petty Officer Academy,Bengbu 233012,China;School of Automation Engineering,University of Electronic Science and Technology of China,Chengdu 610054,China)

机构地区：[1]海军士官学校5系,蚌埠233012 [2]电子科技大学自动化工程学院,成都610054

出　　处：《电子科技大学学报》2024年第1期102-109,共8页Journal of University of Electronic Science and Technology of China

摘　　要：为解决在噪声环境下建模的过拟合问题,基于小波采样理论,提出一种适用于小波神经网络的新型准则函数,并设计了相应的训练算法。这种算法能够利用样本分布和误差训练输入和输出层权值,因此可以大大提高小波神经网络的学习效率。理论和试验表明,新型准则函数有力地保证了小波神经网络的泛化能力,其相应的算法具有全局收敛性,并对噪声变化具有良好的鲁棒性。In order to solve overfitting of modeling in noisy circumstance,a novel cost function with corresponding training algorithm is proposed for wavelet networks based on sampling theory.Since such an algorithm can use sample distributions and errors respectively to train input and output weights,learning efficiencies of wavelet networks are improved greatly.The theories and experiments show that this novel cost function can ensure generalizations of wavelet networks.Simultaneously,the new algorithm can converge globally and is robust to noise varying.

关 键 词：广义采样 神经网络 过拟合 小波采样 

分 类 号：TN911.97[电子电信—通信与信息系统]