To solve the problem of convergence to a local optimum in the multi-layer feedforward neural network , a new disturbance gradient algorithm is proposed. Through introducing random disturbance into the training process, the algorithm can avoid being trapped into the local optimum. The random disturbance obeys the Boltzmann distribution. The convergence of the algorithm to the global optimum is statistically guaranteed. The application of the algorithm in RoboCup, which is a complex multi-agent system, is discussed . Experiment results illustrate the learning efficiency and generalization ability of the proposed algorithm.