Optimisation of Neural-Network Structure using Genetic Techniques

We present arguments that it is necessary to use structured neural networks for the solution of certain problem types. Structure is imposed on connectivity, activation functions and other parameters of the network to simultaneously optimise generalisation ability and compactness of the network. An analogy is made between the development of biological nervous systems from their genetic coding and the generation of artificial neural networks from a parametric description. Experiments are described which use genetic techniques to optimise network structure for a specific class of problem. Results are given which demonstrate the effectiveness of genetic optimisation of network specifications in comparison with other optimisation techniques. The parallel asynchronous implementation of genetic algorithms on a network of Sun’s is briefly described.