The tensile strength, yield strength and elongation of AZ31 magnesium alloys on different annealed conditions are tested by mechanical properties experiments. A model of corresponding mechanical properties is built by applying artificial neural network, and it is optimized by a new method，namely all permutations and combinations training of parameters. The results show that the network model has an excellent performance, which is based on optimal parameters obtained from all permutations and combinations training. Compared with traditional model, whose parameters are obtained from conventional heuristic, the improved model has higher average correlation coefficient and lower average error. Therefore, it can predict the mechanical properties of AZ31 magnesium alloy on different annealed conditions more accurately.