Genetic algorithm for identifying self-generating radial basis neural networks

A self-generating algorithm for radial basis functions RBFs to satisfy a specified model error for a nonlinear system identification problem is proposed. Not only the weights, but also the variance, centre coordinates and number of RBFs are tuned. The genetic algorithms (GAs) approach is used for tuning both the model parameters. The proposed algorithm simulates the natural self-division mechanism used by the tiny creations such as germs and bacteria during the growth process. The used mechanism overcomes a number of difficulties associated with the classical genetic algorithms such as fitness function definition, patents selection for mating and the premature convergence to non-optimal global solution. Computer simulations are used for testing the proposed method to model nonlinear static and dynamic systems.