Evolutionary neural network design

This thesis describes the application of genetic algorithms to the task of neural network design. The neural network model being configured is a radial basis function network. The network uses a single internal layer of locally-tuned processing elements. Each processing element in the internal layer uses a Gaussian sensitivity response to weight matrix output layer. function to determine its ranges of the inputs. An adaptive connects the internal layer to the Given a suitable number of processing elements in the internal layer, a suitable width for the Gaussian functions, and a suitable learning rate, this particular network model converges faster than traditional learning methods. The genetic algorithm searches for optimal values of the these three parameters by evolving network designs which are increasingly better at solving a particular mapping problem. Simulation results are presented for applications in real-valued function approximation and prediction of chaotic time series. a compared to back propagation networks. Results are compared to back propagation networks.