In order to reduce the waiting time and buffer resource consumption of conventional automatic repeat request technology, make the redundant symbol number of digital fountain codes adapt to the varying of channels, and guarantee the quality of the free-space optical communication at the same time, a Q-learning-based Raptor10 code decoding strategy is proposed. Reducing redundant symbols and decoding bit error rate are as the joint return target in the proposed strategy, and Raptor10 code redundant symbols received by receiver are dynamically adapted. As the decoding experience is accumulated continually, the communication performance is improved automatically. Simulation results show that, under the weak, moderate and strong turbulence conditions, the proposed strategy reduces 3% average redundant symbols compared with conventional Raptor10 codes with fixed redundancy rate and linear filtering adaption algorithm.