摘要
根据函数逼近理论以及Weierstrass逼近定理,构造出一类以伯努利多项式的乘积为隐层神经元激励函数的三输入神经网络模型,即三输入伯努利神经网络。针对该网络模型,根据权值直接确定法以及隐层神经元数目与逼近误差的关系,提出了三个网络权值与结构双确定算法。数值实验显示,由这三个算法分别确定的神经网络在学习与校验方面都拥有优越的性能,同时也具有较佳的预测能力。
Based on the function approximation theory and the Weierstrass approximation theorem, a novel 3-input neural net activated by a group of products of Bernoulli polynomials （i. e. , 3-input Bernoulli polynomial neural net, 3IBPNN） was constructed in this paper. Furthermore, on the basis of the weights-direct-determination （WDD） method and the relationship between the number of hidden-layer neurons and the approximation error of the neural net, three different weights and-structure-determination （WASD） algorithms were built up for the constructed 3IBPNN. Numerical experiment results further prove that all of the 3IBPNNs determined respectively by the three proposed algorithms perform excellently in training, testing and prediction.
出处
《计算机工程与科学》
CSCD
北大核心
2013年第5期142-148,共7页
Computer Engineering & Science
基金
国家自然科学基金资助项目(61075121)
教育部高等学校博士学科点专项科研基金博导类课题(20100171110045)
关键词
伯努利神经网络
权值直接确定法
权值与结构确定法
算法
数值实验
Bernoulli polynomial neural net
weights-direct-determination method
weights-and-structure-determination algorithm
algorithm
numerical experiment
