A novel approach for optimizing neural networks with genetic algorithms

A novel approach for optimizing feed forward neural networks is proposed in this paper, the genetic algorithms is not based on the traditional criterion of minimized square error, however its fitness function is determined by the average risk. The method considered not only the errors between the network's outputs and the desired outputs, but also the risk caused by these errors, because the errors for different types of samples in training set may present different risks. The neural networks optimized by the proposed approach shows good performance on the samples both inside and outside training set.