针对传统径向基函数(RBF)网络难以确定迭代停止条件的缺点,提出采用最小化留一误差来训练多尺度RBF网络的算法.分别使用全局k均值聚类算法和经验选择方法,构造RBF节点的中心和尺度参数备选项集合,利用正交前向选择方法逐步最小化留一误差,从而确定网络的每一项中心和尺度参数.实验结果显示,该算法能够自动终止新网络节点选择,不需要额外的迭代终止条件,与传统的RBF网络相比,能够产生稀疏性更高且泛化能力更好的径向基网络.%In order to circumvent the difficulty of pre-assigning a threshold to terminate the iterations in training traditional Radial Basis Function(RBF) network, a novel RBF network training algorithm is proposed. A global k-means clustering algorithm and empirical method are utilized to construct the candidate sets for centre and scales of regressors. At each regressor stage, the parameters of each term are selected by minimizing Leave One Out(LOO) criterion using orthogonal forward selection. Simulation results show that the new algorithm can be terminated fully automatically. Compared with the other RBF networks, this scheme is capable of producing sparser RBF network with much better generality.
展开▼