In this article, first, the advantages of laser communication links are introduced. Then, end-to-end learning based on generative adversarial network improves the real-time and global optimization of the communication system. Finally, because of the lack of robustness in the system caused by the training set cannot contain all situations during the training process of the offline learning mode of the system, a self-organizing incremental learning method is introduced to improve online incremental training of the end-to-end system. Experiment results show that by implementing the self-organizing incremental learning method, the end-to-end communication system can simulate the channel effect efficiently, and it greatly improved the real-time optimization and robustness of the system.