For the problems of low solution precision by the initial bat algorithm and falling into local optimum easily, an improved self-adaptive bat algorithm(SABA) is proposed, which combines the mechanisms of step-control and variation. Experiments are conducted on a set of 12 benchmark functions, and the results show that the proposed SABA has better performance than the particle swarm optimization(PSO) algorithm and initial bat algorithm(BA) in terms of accuracy and convergence speed.