尺度可调的混合核RBF网络

Mixture Kernel RBF Network with Tunable Scales

针对传统核模型中采用单一核函数的局限性,利用两个核函数的线性组合得到混合核.在RBF网络的训练中,采取正交最小二乘的方法进行逐步回归建模.在学习每个神经元参数时,首先,用全局k均值聚类法得到数据样本的聚类中心,然后对每一个聚类中心,利用群搜索优化器搜索出最佳的尺度和混合核调节参数,误差最小的参数组合即为径向基函数参数.实验说明,新的RBF网络具有稀疏性好,泛化能力高等优点.

To overcome the limitation of one single kernel in the traditional kernel function model,a new type of mixture kernel is constructed by combining two kernel functions linearly.Each individual regressor in RBF network is trained term by term using orthogonal least squares algorithm.In the training phase,a global k-means cluster algorithm is used to decide the kernel centers.For each clustering centers,Group search optimizer（GSO） is utilized to get the proper scale and weight within the kernel.The optimal kernel parameters which minimize the training error are used to shape each regressor.The experiments results show that the new RBF network is sparser than some traditional RBF network with one single kernel.