The invention discloses a multi-layer perceptron training method based on a pigeon swarm optimization algorithm. Firstly, the basic parameters of the pigeon swarm algorithm are initialized, and the position and range of an individual are randomly initialized. Then the position and velocity of pigeon are updated according to the formula in the compass operator phase to expand the parameter optimization range of multilayer perceptron. Next, the surface operator is used to speed up the convergence rate and dig the local optimization. In each iteration, the function used to evaluate the fitness value of the individual is the mean square error of the multilayer perceptron corresponding to the parameters carried by the individual; In the process of training the multilayer perceptron, the invention and the traditional BP algorithm can overcome the characteristics that the multilayer perceptron is easy to fall into the local optimal solution and the searching process is slow, can dig out the multilayer perceptron parameters of the optimal performance as far as possible, and improves the performance of the multilayer perceptron.