Recurrent neural networks such as the fully interconnected Hopfield
network have the potential of being applied to many input output mapping
problems, especially those requiring the outputs to change with time,
such as a centeral pattern generator. However, due to the difficulty of
training, such networks have not been as extensively utilized as the
multilayer feedforward networks which can be trained using the back
propagation algorithm. In this thesis we explore the applicability
of the genetic algorithm as a search technique to find the network
parameters for recurrent neural networks. We also investigate the
potential of the genetic algorithm to determine not only the network
parameters, but also the number of neurons to use in the network
architecture.
Available in PDF format:
Available in postscript format: