Abstract: Deep feed-forward neural networks are well applied in classification and regression problems, but network performance is greatly affected by their structure and hyper-parameters. To achieve high performance neural networks, a modified genetic algorithm is designed firstly, which modifies the selection strategy. Then, the modified genetic algorithm is employed to optimize the number of network layers, the number of nodes in each layer, and the learning rate and weights, which are coded by binary coding and real number coding strategy respectively. For the modified selection strategy, in 2n  indivi- duals from the combination of parent population with offspring population, some top fitness individuals are selected and some worse fitness individuals with a high probability are also selected to achieve better diversity and avoid falling into local optimum. dropout method is introduced to avoid the overfitting training data of network. Seven datasets (Ring, Breast cancer，Twonorm, Heart，Blood，Ionosphere，Monk) are used in the experiments. The results show that, compared with the algorithms in related literatures, the modified genetic algorithm has higher performance neural networks.

Key words: deep feed-forward neural network, modified genetic algorithm, network structure optimization, hyper-parameter optimization