An algorithm for automatic designation of the architecture and the weights of neural networks using gene expression programming (GEP) was presented. The fundamental ideas and procedures of the algorithm were discussed. The algorithm was improved to solve the problems of prematurity and lower variance rate. An application for neural networks designation was given. The experimental results indicate that the proposed GEP approach may evolve the architecture of neural network, and can obtain the weights more precisely. Compared to other conventional evolutional algorithms, GEP shows faster convergence.