Abstract: Traditional training algorithms for radial basis function (RBF) neural networks usually start with a predetermined hidden layer structure, which is selected by using a priori knowledge and based on previous experience. The parameters of RBF networks are trained by using traditional clustering and the least squares method. These training algorithms are always based on the local search method and often suffer from being trapped at structural local minima. A new method for training RBF structural parameters by using a genetic algorithm is put forward, of which a new clustering method named the sorting MSS (most scattered set) clustering method and a traditional least square method are incorporated. This method can not only prevent the result of the network from being trapped at local minima but also highly improves the computational efficiency. It gives satisfactory results when this algorithm is applied to traffic flow forecasting.

Key words: GA, MSS means clustering , RBF neural networks