一种快速逼近的神经网络构造性算法

A Faster Convergent Constructed Algorithm for Neural Networks

选取合适的网络结构是前馈网络应用中首先遇到一个十分重要而又困难的问题。文章针对前馈网络的特点,提出了一种网络构造算法。该算法的主要特点是把二层网络的训练问题转化为求解线性方程组的最小二乘解,直接求得使误差最小的权值矩阵,并通过划分样本子集来创建子网,把子网输出映射到输出层,形成三层网络结构。由于避免了传统BP算法的梯度搜索过程,该算法能快速收敛。仿真结果表明,所提出的算法可快速、有效地逼近连续函数。

How to select appropriate network architecture is a very important and difficult problem for one to first sur-mount in applications.Considering the feature of Neural Network,an algorithm for constructing and training multi-layer neural network is presented in this paper.Its distinctive feature is using least-square algorithm to find the weights of a two-layer neural network,and selecting subsets of the pattern space creates hidden layers.The construction algorithm avoids iterative process of typical gradient algorithms ,such as BP algorithm,so it is quick to converge.In final a simulation example is given to show that the construction algorithm is fast and efficient method for function ap-proximation.