Deterministic annealing learning of the radial basis function nets for improving the regression ability of RBF networks

The deterministic annealing method for training the center vectors of RBF networks is proposed. The method is a soft-competition scheme and derived from optimizing an objective function using the gradient descent method. To some extent it can overcome the problems that the learning vector quantization algorithms with the winner-take-all scheme and the heuristic procedure have. The emulation experiment is given to validate the algorithm. The experimental results show that, compared to the error backpropagating algorithms of the multi-layer perception and the RBF network, it not only enhances learning precision and generalization ability, but also reduces learning time as well