A novel learning objective function using localized generalization error bound for RBF network

A major issue of pattern classification problems is to train a classifier with good generalization capability. In this paper, a novel training objective function using the localized generalization error model (L-GEM) is proposed for a RBF network. The weight parameter of a RBF network is calculated to minimize its localized generalization error bound. The proposed training objective function is compared with well-known training methods: minimizing training error, Tikhonov Regularization and Weight Decay. Experimental results show that RBF networks trained by minimizing the proposed objective function consistently outperform other methods.