Representative evolution: a simple and efficient algorithm for artificial neural network evolution

In this study a new evolutionary algorithm, i.e., representative evolution (RE), for evolving artificial neural networks (ANN) is proposed. Unlike most of the evolutionary algorithms, the RE uses population information for generating variations in individuals of a population. An evolutionary system, i.e., RENet, based on the RE for evolving feedforward artificial neural networks with weight learning is described. The RENet uses three operators (i.e., one crossover and two mutations) sequentially. If one operator is successful, no other operator is applied. The RENet is applied to a benchmark character recognition problem. It can produce very compact ANN size with a small classification error.