Low-latency deep-reinforcement learning algorithm for ultrafast fiber lasers  

作　　者：QIUQUAN YAN QINGHUI DENG JUN ZHANG YING ZHU KE YIN TENG LI DAN WU TIAN JIANG 

机构地区：[1]College of Computer,National University of Defense Technology,Changsha 410073,China [2]College of Advanced Interdisciplinary Studies,National University of Defense Technology,Changsha 410073,China [3]National Innovation Institute of Defense Technology,Academy of Military Sciences PLA China,Beijing 100071,China [4]Beijing Institute for Advanced Study,National University of Defense Technology,Beijing 100020,China [5]Hefei Interdisciplinary Center,National University of Defense Technology,Hefei 230037,China

出　　处：《Photonics Research》2021年第8期1493-1501,共9页光子学研究（英文版）

基　　金：National Natural Science Foundation of China(62075240);National Key Research and Development Program of China(2020YFB2205804)。

摘　　要：The application of machine learning to the field of ultrafast photonics is becoming more and more extensive.In this paper,for the automatic mode-locked operation in a saturable absorber-based ultrafast fiber laser(UFL),a deep-reinforcement learning algorithm with low latency is proposed and implemented.The algorithm contains two actor neural networks providing strategies to modify the intracavity lasing polarization state and two critic neural networks evaluating the effect of the actor networks.With this algorithm,a stable fundamental modelocked(FML)state of the UFL is demonstrated.To guarantee its effectiveness and robustness,two experiments are put forward.As for effectiveness,one experiment verifies the performance of the trained network model by applying it to recover the mode-locked state with environmental vibrations,which mimics the condition that the UFL loses the mode-locked state quickly.As for robustness,the other experiment,at first,builds a database with UFL at different temperatures.It then trains the model and tests its performance.The results show that the average mode-locked recovery time of the trained network model is 1.948 s.As far as we know,it is 62.8%of the fastest average mode-locked recovery time in the existing work.At different temperatures,the trained network model can also recover the mode-locked state of the UFL in a short time.Remote algorithm training and automatic mode-locked control are proved in this work,laying the foundation for long-distance maintenance and centralized control of UFLs.

关 键 词：FIBER POLARIZATION algorithm 

分 类 号：TN248[电子电信—物理电子学] TP18[自动化与计算机技术—控制理论与控制工程]