72 An Automated Approach to Generate Neural Network Topologies

This paper presents a novel concept of using graph grammars to generate neural network topologies. The approach involves representing the neural network as a graph and defining graph transformation rules to generate the topologies. The method starts from a small network and then uses predefined grammar rules to build the topology. Each topology generated by the graph grammar is trained using a hybrid training algorithm that uses a combination of genetic algorithms and a quasi-Newton (BFGS) method. The error of the response predicted by each topology during training, is used to further evolve the topologies. The proposed methodology appears to generate promising responses and has the potential to be a powerful process modeler.