In this study, we propose an adaptive optics closed-loop control model based on the far-field index gradient, which can be used to stabilize the response matrix based on the recursive least square values. Further, the current system state can be rapidly self-learned using the far-field index gradient. The experimental results denote that the proposed model exhibits real-time online update characteristics; furthermore, the proposed model can adapt to the state of H-S subaperture lack of light or non-ideal centroid detection, which improves the control performance to some extent.