The stochastic parallel gradient descent (SPGD) algorithm is a widely used model-free optimization algorithm in adaptive optical systems. However, when affected by noise, the SPGD algorithm may incorrectly estimate the gradient, resulting in a significant decrease in its performance and limitations in its application in complex environments. To address this issue, this paper introduces the optimal estimation algorithm and adds a self-checking process to the SPGD algorithm, proposing a Kal-SPGD algorithm. The feasibility of the Kal-SPGD algorithm is verified through simulation and experimentation. The experimental results show that compared with the standard SPGD, the use of the Kal-SPGD algorithm can reduce the signal amplitude variance by up to 92.15%. The Kal-SPGD algorithm provides a more favorable control technique for adaptive optical systems.