近邻传播聚类算法的RBF隐含层节点优化.docVIP

近邻传播聚类算法的RBF隐含层节点优化 　　摘 要： 传统的RBF神经网络预测精度会由于随机选取隐含层中心节点不合适而导致算法效率低下和数值病态，为了提高RBF神经网络的效率，提出了一种用近邻传播AP聚类算法改进RBF神经网络的方法，并介绍了该方法的原理及建模步骤。由于采用的AP聚类算法属于自适应聚类学习算法，无需事先给定隐含层中心节点的个数，能够适用于不具有先验信息的预测。首先，利用AP算法根据训练样本的信息进行聚类迭代，从而确定RBF神经网络中隐含层的中心节点和节点数值，解决了RBF网络的中心取值问题。然后，把所有输入数据代入基于AP聚类算法优化的RBF神经网络中进行预测。由于AP算法无需预先指定聚类数目，所提方案能提高网络的学习精度和训练速度，利用所提优化方案对正弦函数进行逼近的仿真实验，结果表明该方案的逼近误差仅为0.005 5，在0.3噪声下能保持较好的预测精度。 　　关键词： 径向基函数神经网络； 近邻传播聚类算法； 隐含层； 逼近误差 　　中图分类号： TN711?34； TP398.1 文献标识码： A 文章编号： 1004?373X（2016）19?0016?04 　　Abstract： The prediction accuracy of the traditional radial basis function （RBF） neural network may result in lower algorithm efficiency and pathological numerical value due to the inappropriate random selection of the hidden layer center node， to improve the efficiency of RBF neural network， a method of using affinity propagation （AP） clustering algorithm to improve RBF neural network is proposed. The principle and modeling steps of the method are introduced. Since the adopted AP clustering algorithm belongs to the self?adapting clustering learning algorithm， it needn′t predefine the numbers of the hidden layer center nodes， and is applied to prediction without transcendental information. The AP algorithm is used for clustering iteration according the information of training sample， so as to determine the center node and node numerical value of hidden layer in RBF neural network， and solve the center dereferencing problem of RBF network. After that， all input data is taken in RBF neural network based on AP clustering algorithm for prediction. Since the use of AP algorithm needn′t predefine the clustering numbers， the proposed scheme can improve the learning accuracy and training speed of the RBF neural network. The approximate simulation experiment was performed for sine function with the proposed optimization scheme. The results show that the approximate error of the proposed scheme is only 0.005 5， and can keep good prediction accuracy under the noise of 0.3. 　　K