Evolutionary Design of Constructive Multilayer Feedforward Neural Network

This paper proposes an evolutionary design methodology of multilayer feedforward neural networks based on constructive approach. We elaborate an adjustable processing element as primitive neuron model. The neural layer can be constructed by assembling several neurons. The multilayer neural network can be finally constructed through cascading several neural layers. The constructive approach facilitates substantially to extract design specifications from a multilayer neural network. Based on the constructive representation of multilayer feedforward neural networks, we use a genetic encoding method, after which the evolution process is elaborated for designing the optimal neural network. The results of our experiments reveal that our methodology is superior to the error back-propagation algorithm both for its executing efficiency and performance. Key-Words: Evolutionary Design; Neural Network; Genetic Algorithm;