Abstract: From the perspective of summarizing and analyzing the existing neural networks,it is discussed the structure,the character and the input/output relationship of the GMDH neural network model.Then,it is introduced a sort of neural network learning algorithm based on the GMDH model,described the main design idea and the realization process of this algorithm,and used simulated technique to research the problem of how to make the algorithm stop and how to find the optimal number for network layers.Through experiments,it is shown that the algorithm has strong self-organization ability,preferable generalization ability and stability.