Genetic Algorithm for Multilayer Radial Basis Function Networks Based on Generalized Inverse Matrix

In order to solve the problem that maxtrice are nearly singular when using regular least squares method, a method of generalized inverse matrix was employed to obtain the weight vectors of each layer in the multilayer radial basis function network, which was introduced to the genetic algorithm for training multilayer radial basis function networks. By using real function approximation, chaotic time series modeling and forecasting simulation experiments, the algorithm was verified. The results show that the generalized inverse matrix method is much more better than regular least squares method on the approximation precision, which can be up to a 1 to 2 orders of magnitude.