Abstract
In recent years, several studies discussed the use of evolutionary algorithms, as more promising approaches, for automatic design of feedforward neural networks. Such methods are particularly useful for dealing with complex problems having large search spaces with many local optima. However, experimental evidence had indicated the inefficiency of these algorithms at fine tuning solutions. In this paper, we show how a pure mutation-based evolutionary algorithm could be used to find global basins of attraction by selecting the appropriate structure of a feedforward neural network and also to improve finer tuning capabilities of the designed network by using a specific self-adaptive procedure. Pertinences of this algorithm reside first in its structure selection method based on a competition between user chosen structures and second its adaptive weight perturbation with a constant mutation probability. It is a real pure evolutionary algorithm of design of multilayer neural networks with competitive and stable performances compared to popular algorithms as will be proven by simulations.