An efficient sequential learning algorithm for growing and pruning direct-link RBF (DRBF) networks

This paper extends the sequential learning algorithm GAP-RBF to the direct link radial basis function (DRBF) networks, and describes a modified GAP-RBF learning algorithm used to train DRBF networks. The modified algorithm reserves the Growing and Pruning Criterion, defined in the GAP-RBF, but the decomposed extended Kalman Filter (DEKF) is used, instead of EKF in the original GAP-RBF algorithm, to adjust the centre, width and bias of the hidden neurons, and the weights of direct links of DRBF. A function approximation is used as the benchmark problem, in which the network is trained to approximate the rapidly changing continuous function referred to as "SinE". The simulation result shows that, when the target function has linear items, modified algorithm has a better generalization performance than RBF algorithm, and DRBF networks using modified algorithm are more compact.