Artificial fish swarm algorithm (AFSA) is a global optimization method proposed recently. After analyzing the disadvantages of AFSA, this paper introduced best-step operator and refined the prey behavior. An improved artificial fish-swarm algorithm for the RBF neural network and a model based on this method is developed. Finally the new algorithm is applied to the problem of expression recognition. The research indicates that the new algorithm has some advantages in terms of convergence performance, recognition rate and so on.