摘要
针对传统径向基函数神经网络构造的网络分类器通常存在分类精度不高、训练时间长等缺陷,首先提出了一种改进的自适应聚类算法,用于确定分类器的隐含层节点。该算法通过筛选基于轮廓系数的优秀样本群,来寻找最佳初始聚类中心,避免了传统K-means算法易受初始聚类中心点影响,导致最终的分类效果严重偏离全局等情况的发生。其次,将该改进算法用于构造径向基函数神经网络分类器和快速有效地确定隐含层节点径向基函数中心及函数的宽度。最后,通过大量UCI数据集的实验和仿真,验证了改进算法在聚类时间、聚类轮廓系数及聚类正确率等方面具有优越性。同时,大量的仿真实验也证明了基于改进算法构造的RBF分类器具有更高的分类精度。
Owing to defects of lower classification precision and longer training time of Radial Basis Function Neural Network （RBF） classifier,a new self-adaptive clustering algorithm was produced firstly,which can be applied into construction of nodes in implicit layer.The new algorithm optimizes initial cluster centers by choosing good samples based on silhouette coefficients.It not only avoids the effects of initial centers in traditional k-means,but also avoids classification deviation.Secondly,the new algorithm was introduced into designing of RBF classifier.It can ascertain centers of radial basis function and its width efficiently.Finally,by a large number of tests and simulation,the new clustering algorithm was testified to be superior in clustering time,silhouette coefficients and accuracy rate.Besides,RBF classifier based on the advanced algorithm was proved to have higher precision.
出处
《计算机科学》
CSCD
北大核心
2014年第6期260-263,共4页
Computer Science
基金
山西省自然科学基金项目(2013011016-3)资助
