In this paper, the limitations of conventional BP algorithm was analyzed, and to fasten the learning velocity of neural network and enhance its generalization capability, the APSO (adaptive particle swarm optimization) algorithm was introduced into BP network for the optimization of its weights and thresholds. To overcome its 'early maturity', the variance operation was made on particles owing bigger fitness values according to certain probability, and the inertia weights were dynamic changed during the iteration process. The algorithm was simulated on a nonlinear function regression problem, which shows that the improved PSO-BP algorithm has faster learning velocity and better generalization ability than conventional BP network. The simulation results prove that the APSO-BP algorithm can get over the limitations of the conventional BP network and is superior to it.