Evolutionary Design of Radial Basis Function Neural Network for Data Modelling

ABSTRACT: In this thesis, implementation of Radial Basis a Function Neural Network (RBFNN) using genetic algorithm is described. The developed algorithm is used to model a certain dataset by training a RBFNN using some part of it, and then testing the performance of this RBFNN using the rest of data. The objective function of the proposed algorithm is to minimize the error between the computed output by the model and the target output given in the dataset. The genetic algorithm used in this thesis is an evolutionary algorithm that uses natural evolutionary process for selection and reproduction. An individual is constructed from the RBFNN parameters, which are hidden units, centers, weights, widths and bias associated with hidden units and output of RBFNN. Therefore, the fitness values are also assigned to all chromosomes as a result of getting the difference between the target output and the computed output by the RBFNN, in which a Gaussian function was used as an activation function. In experimental results, different tests were conducted in order to see the performance and correctness of the developed model. Since the number of hidden units plays an important role as well as weights, the intervals of weight values were adjusted accordingly and the number of hidden units was changed for different tests. As a result of conducted experiments, it is observed that the developed algorithm is successful in obtaining good results by minimizing the error. Keywords: Evolutionary algorithms, Radial Basis Functions, Data Modeling. ……………………………………………………………………………………………………………………………………………………………………………………………………………………