Combined learning and pruning for recurrent radial basis function networks based on recursive least square algorithms

This paper discusses a way to combine training and pruning for the construction of a recurrent radial basis function network (RRBFN) based on recursive least square (RLS) learning. In our approach, a RRBFN is first trained using the proposed RLS algorithms. Afterwards, the error covariance matrix which is directly obtained from the RLS computations is used to remove some unimportant radial basis function (RBF) nodes. We propose two algorithms: (1) a “global” version which is suitable for low dimensional input space situation, and (2) a “local” version which can be applied in situations when the input dimension is large. In both cases, it is shown that the error covariance matrix, obtained from the RLS algorithms, can be used as a means for pruning the trained RRBFN. Simulation examples are presented to illustrate the proposed approaches.