基于ε-不敏感准则和结构风险的鲁棒径向基函数神经网络学习

摘要

An ε- insensitive criterion and structure risk based Radius-Basis-Function Neural-Network (RBF-NN) modeling method is proposed. By introducing ε- insensitive criterion and the item of structure risk, the RBF-NN learning is transformed into the linear regression and Quadratic Program (QP) optimization issue. Compared with the traditional least-square-criterion based RBF-NN training algorithms, the proposed method is much more robust to noise data and small size of datasets. Through the simulation experiments on the synthetic and real-word datasets, the above virtues are confirmed.%该文提出一种新的基于ε-不敏感准则和结构风险的径向基神经网络(RBF-NN)建模方法.通过引入ε-不敏感准则和结构风险项,把RBF-NN训练转化为线性回归和经典的二次规划优化问题.和传统的基于最小平方误差准则的RBF-NN训练算法相比,提出的新方法对小样本数据集和噪声数据显示出较好的鲁棒性,具有较好的泛化能力.通过模拟和真实数据集进行仿真试验,上述优点得到了有效验证.