The paper presents a novel two-layer learning method for radial basis function (RBF) networks. At the lower layer, a regularised orthogonal least squares (ROLS) algorithm is employed to construct RBF networks while the two key learning parameters, the regularisation parameter and hidden node's width, needed by the ROLS algorithm are optimized using the genetic algorithm at the higher layer. Networks constructed by this learning method have superior generalisation properties, and the computational complexity of the method is reasonable. Nonlinear time series modelling and prediction is used as an example to demonstrate the effectiveness of this hierarchical learning approach.