Regularized RBF- FA Neural Network to improve the generalization performance of function approximation

In order to improve the RBF (Radial Basis Function) Neural Network´s generalization performance, two main methods are proposed to improve the adaptive network structure and regularization. FA (Fuzzy Adaptive Resonance Theory, Fuzzy ART) is utilized as the preprocessor of the RBFN to compress the training data into a fewer number of clusters and give the adaptive network parameters determination approach. Regularization improves network generalization ability by adding penalty term to original cost function. L-curve method is applied to optimal regularization parameter. A simulation of the BDI (Baltic Dry Index) dataset, compared with the regularized BP Neural Network, RBFN based on k-means clustering method and RBF-FA method, demonstrates the proposed fusion regularized RBF-FA algorithm can get good generalization performance.