Self-Adaptation in Evolutionary Design of Neural Networks
Evolutionary computation as an alternative to the traditional methods of multilayer neural networks design has been widely applied. The results of many simulations show that evolutionary algorithm can outperform standard training strategies, including back-propagation and its modifications.
The algorithm, described in this paper, summarises the results of our work both in the field of recurrent and feedforward networks. Its main feature is the selfadaptation procedure applied to make the search for the network weights and architeture both effective and precise.