An Adaptive Multi-Kernel RBF Model Using State Matching

Radial basis function neural network has a strong capability of non-linear mapping for system identification. Especially, using the orthogonal least square method can generate a parsimonious structure to avoid overfitting problem effectively. Nonetheless, it is difficult to deal with dynamic systems by static models, which exist mainly in manufacture and life. Aimed at the non-stationary time series presented by dynamic systems, it is necessary to study on-line model with alterable or composite structure. This paper proposes a multi-kernel RBF neural network with some novel methods. The kernels are trained by transformations of sample set instead of just by the sample set like regular algorithms. And the on-line weight distribution of all kernels relies on the state matching method to trace the changes in the system. Finally, through some numerical experiments, the performance of proposed model is validated by comparing it to the other models.