Discussed the training algorithms and its features of multilayer feedforward neural networks, and derived a fast training algorithm based on the layer by layer optimizing procedures. In thisnew algorithm, the weight matrix（V） of the output layer and the output vector（B） of the previous layerare trea tedas two variable sets. An optimal solution pair （V*, B*p） is found to minimi ze the sumsquare erro r of the pat terns. B*p is then used as the desi red output of th e previo us layer, the optimalw eight ma trix and o utput v ecto r o f the hidden layer i s found with the same method as that used for the output layer. Computer simulatio n show s that the new algori thm is effective.