查看文章

The term neural network evolution usually refers to network topology evolution leaving the network's parameters to be trained using conventional algorithms. In this paper we present a new method for neural network evolution that evolves the network topology along with the network parameters. The proposed method uses grammatical evolution to encode both the network and the parameters space. This allows for a better description of the network using a formal grammar allowing the network architect to shape the resulting search space in order to meet each problem requirement. The proposed method is compared with other three methods for neural network training and is evaluated using 9 known classification problems and 9 known regression problems. In all 18 datasets, the proposed method outperforms its competitors.