The role of weight domain in evolutionary design of multilayer perceptrons

Among different models of neural networks multilayer perceptrons play an important role. Most training methods, including back-propagation concentrate on weight adjustment only. Still the performance of the network strongly depends on its architecture. In our paper the algorithm based on evolutionary programming is proposed. Unlike most other methods of this type, the genotype precision is being evolved together with the architecture and connection weights of the network. Iterative changes in the weight domain make the network structure rough at first so as to tune it later. Not only does it help to avoid inadequate weight precision, but also the search efficiency is increased. Different aspects of the weight set selection are investigated and discussed.