Aiming at the disadvantage of gradient instability and poor global optimality of the standard synchronous perturbation stochastic approximation (SPSA) algorithm, this paper proposes a improved SPSA algorithm based on double calculation step, which uses both stable convergence step and random differentiation step, on the basis of eliminating the gradient instability defect. The example shows that compared with the standard SPSA algorithm, the improved SPSA algorithm based on double calculation step has better global optimality.