Genetic structure for NN topology and weights optimization

A structural genetic algorithm is proposed to optimize the neural network topology and connection weightings. This approach is to partition the genes of chromosome into control genes and connection genes in a hierarchical fashion. The control genes represented in bits are used to govern the layers and neurons activation and considered to be the higher level genes. Whereas the connection genes in the form of real values are the weightings and bias representations, regarded as the lower level genes. This inherent genetic variations enable multiple changes in lower level genes by a single change at the higher level genes. Such formulation of chromosome is found to be a phenomenal improvement over the traditional GA approach that without genes classification. As a result, the learning technique of the neural network is greatly improved. Simulation results have indicated that the proposed learning scheme requires the least iteration steps to reach a optimum network as compared to the uses of backpropagation and traditional non-structural genetic algorithms.