A network training method for star pattern recognition was designed by combining a classific Radial Basic Function(RBF) neural network and star pattern samples. Firstly,the star pattern abstraction method was discussed and a triangulation based on star magnitudes was induced to connect the stars which probably appear in the same field of view.By taking extrated angular distances as the characteristic of star pattern, a star pattern sample set with completion, translation and rotation invariance was established. Then, RBF neural network was studied to recognize the star patterns. RBF network training method was classified as sequence learning and batch learning. Some typical algorithms that could represent the two methods were studied on their advantages and disadvantages,and a new training method was designed based on the specialty of above star pattern sample sets.Experiments indicate that the designed method is more appropriate than those typical algorithms. Several star images were simulated through software, which was regarded as the observatory data and entered into the trained RBF neural network to test. The experiment results show that the network can recognize all the star patterns successfully.