多层前向网络的逼近机理与拓扑结构学习方法

An Approximation Mechanism of Multilayer Feedforward Network and Its Learning Algorithm of Topology

对多层前向网络的最小二乘逼近机理进行了系统的分析，指出隐层节点函数特性的特定选择是构成网络有效逼近能力最关键的因素。分析了增加隐层数和增加隐节点数在改进网络逼近效果方面不同的作用机理，给出了前向网络拓扑结构学习的通用算法和其对应的神经生物学机制。

A systematic analysis is made on the least square approximation mechanism of multilayer feedforward neural networks,and the following conclusion is drawn that the types of node functions in hidden layers are the most important factor for forming good approximation capability of networks Besides these,the different acting mechanisms of adding number of hidden layers and adding number of hidden nodes for improving approximation results of neural network is analyzed,and a general learning algorithm of network topology is suggested,the corresponding mechanism of neurobiology is supposed