一种改进的RAN网络学习算法

Training algorithm of RBFNN similar with RAN

本文讨论了一种类似于RAN(资源分配网络)网络学习算法的动态RBFNN学习算法。该学习算法在均值聚类初始化基础上,选取训练过程中误差最大的样本,根据RAN网络的新性条件,决定是否分配新的隐层节点,使用最小二乘法训练权值。最后通过对无机建筑材料成分分析的仿真表明该算法可以简化网络结构,实现样本正确分类,并获得较好的泛化性能。

An efficient dynamic training algorithm of RBFNN which is similar with RAN(resource-allocating network) for pattern recognition is presented in this paper The algorithm chooses the maximal error pattern during the training process after initialization,then according to the novelty of RAN,decides it to be a new hidden unit or to use it to alter the network parameters.The training of weight utilizes the least square method.Finally,simulation by the analysis of the composition of building materials shows that the algorithm proposed above could improve better performance of generalization,while only needs smaller architecture of network.