Optimizing The Structure Of Neural NetworksUsing Evolution Techniques
SPIRIDON D. LIKOTHANASSIS, Efstratios F. Georgopoulos, Dimitris A. Fotakis · WIT transactions on information and communication technologies · 1997
Evolutionary methods have been widely used to determine the structure of artificial neural networks. In this work, we propose a more efficient implementation which face the above problem, by using a neural network model as general as possible. So, we used a fully connected network, consisting of three parts, the input layer, the output layer and a number of hidden layers. This implementation has been proved, via simulations, that it optimazes the size of the network and gives better results compared with other existing methods, while the use of crossover results to more efficient networks. Furthermore, the random initialization has been proved more efficient, since it reduces significantly the number of the generations needed for the convergence of the algorithm. Finally, some aspects for the convergence of the parallel implementations of the algorithm are discussed.