This paper trains the multilayer neural networks using dynamical programming.The weithts are adjusted layer by layer. The optimal weishts of each layer are calculated using OBPalgorithm proposed in the related reference. The globally optimal weights of each layer can be obtained by finite iterations.