A Self-Adaptive Growing Method For Training Compact Rbf Networks

Radial Basis Function (RBF) network is a neural network model widely used for supervised learning tasks. The prediction time of a RBF network is proportional to the number of nodes in its hidden layer, while there is also a positive correlation between the number of nodes and the predication accuracy. In this paper, we propose a new training algorithm for RBF networks in order to construct high accuracy networks with as few nodes as possible. The proposed method starts with an empty network, selecting a best node from candidates iteratively until the training error reduces to a threshold or the number of nodes reaches a limit. Then the network is further optimized with a supervised fine-tuning method. Experimental results indicate that the proposed method could achieve better performances than traditional algorithms when training same sized RBF networks.