RBF神经网络的一种鲁棒学习算法

A Robust Learning Algorithm for RBF Neural Networks

用定标鲁棒代价函数代替传统的二次型指标 ,并结合改进的遗传算法 ,搜索近最优径向基函数神经网络 ( RBFNN)的结构和参数 .实验结果表明该训练方法比其他方法具有更强的鲁棒性 ,可提高 RBFNN的泛化能力 ,自动消除训练数据中的噪声 ,再现训练数据中的潜在规律 .

The near optimal structures and parameters of the radial basis function neural networks (RBFNNs) are found by replacing the quadratic loss function with a scaled robust loss function, and also incorporating with improved genetic algorithm. The experimental results show that the learning algorithm proposed is of stronger robustness than other ones, and the generalization ability of the RBFNNs would be improved. The noise mixed in the training data would be eliminated automatically, while the underlying trend in the training data would reappear.