Structured genetic algorithm representations for neural network evolution

Evolutionary Algorithms used to generate Artificial Neural Networks have relied on both binary and real value representation approaches to encode connection weights in the chromosomes. This paper documents a study which examined how the utilisation of these two approaches affects the convergence of the Structured Genetic Algorithm when used to evolve Artificial Neural Networks. This study found that Structured Genetic Algorithms exhibited better performance when they utilised a real valued approach to encode the weights, especially when multiple control levels were utilised. Structured Genetic Algorithms which used real number encoding for the weights in their parametric level attained reduced training and testing errors. A reduction in the duration of the SGA simulations was also observed, though this diminished with each added control level. In contrast to this, Genetic Algorithms performed better with the binary encoding approach.