A divide-and-conquer learning approach to radial basis function networks

This paper presents a new divide-and-conquer based learning approach to radial basis function (RBF) networks, in which a conventional RBF network is divided into several RBF sub-networks. Each of them individually takes an input sub-space as its input. The original network's output then becomes a linear combination of the sub-networks' outputs with the coefficients adaptively learned together with the system parameters of each sub-network. Since this approach reduces the structural complexity of a RBF network by describing a high-dimensional modelling problem via several low-dimensional ones, the network's learning speed is considerably improved as a whole with the comparable generalization capability. The empirical studies have shown its outstanding performance on forecasting two real time series as well as synthetic data. Besides, we have found that the performance of this approach generally varies with the different decompositions of the network's input and the hidden layer. We therefore further explore the decomposition rule with the results verified by the experiments.