GROWING RADIAL BASIS FUNCTION NETWORKS USING GENETIC ALGORITHM AND ORTHOGONALIZATION

With the recent popularity of the neural networks for modeling nonlinear systems, there is a growing need for a systematic and autonomous way to simultaneously determine the optimal network size and parameters. This paper tries to meet such demand by presenting a new growing algorithm for the radial basis function networks (RBFN). In the proposed algorithm, a new hidden node is added to the RBFN in sequence while previously found hidden nodes remain intact. Genetic algorithm searches for each new hidden node after calculating the fitness of candidate hidden nodes via a computationally efficient orthogonalization procedure. The proposed algorithm is compared with conventional GA-based methods and non GA-based growing algorithms in the literature through benchmark examples to demonstrate its superior performance in model-building without any priori knowledge. Lastly, the proposed algorithm is successfully applied to a task of modeling the disk grinding process based on the actual process data collected from electronics industry.