• Acta Optica Sinica
  • Vol. 43, Issue 21, 2114002 (2023)
Jiacheng Wu1,2, Meng Cai3, Yujie Lu1,3, Nanshun Huang4,*..., Chao Feng2,3 and Zhentang Zhao1,2,3|Show fewer author(s)
Author Affiliations
  • 1School of Physical Science and Technology, ShanghaiTech University, Shanghai 201210, China
  • 2Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai 201210, China
  • 3Shanghai Institute of Applied Physics, Chinese Academy of Sciences, Shanghai 201800, China
  • 4Zhangjiang Laboratory, Shanghai 201210, China
  • show less
    DOI: 10.3788/AOS230893 Cite this Article Set citation alerts
    Jiacheng Wu, Meng Cai, Yujie Lu, Nanshun Huang, Chao Feng, Zhentang Zhao. Reinforcement Learning for Free Electron Laser Online Optimization[J]. Acta Optica Sinica, 2023, 43(21): 2114002 Copy Citation Text show less
    Layout of the undulator system
    Fig. 1. Layout of the undulator system
    Comparison of power curves among SAC, TD3, and DDPG algorithms for FEL tuning tasks
    Fig. 2. Comparison of power curves among SAC, TD3, and DDPG algorithms for FEL tuning tasks
    Comparison of gain curves and optimized orbits among SAC, TD3, and DDPG algorithms in the FEL optimization. (a) Gain curves; optimized orbits along (b) x and (c) y directions
    Fig. 3. Comparison of gain curves and optimized orbits among SAC, TD3, and DDPG algorithms in the FEL optimization. (a) Gain curves; optimized orbits along (b) x and (c) y directions
    Comparison between the optimized spots of DDPG, TD3, and SAC algorithms and the initial spot. (a) Initial spot; (b) DDPG optimized spot; (c) TD3 optimized spot; (d) SAC optimized spot
    Fig. 4. Comparison between the optimized spots of DDPG, TD3, and SAC algorithms and the initial spot. (a) Initial spot; (b) DDPG optimized spot; (c) TD3 optimized spot; (d) SAC optimized spot
    ParameterValue
    Beam average energy /GeV1.5
    Peak current /A800
    Energy spread /%0.014
    FEL wavelength /nm3.72
    Average beam radius /μm50
    Undulator length /m3
    Period length /cm2.35
    Table 1. Main parameters of the simulation
    ParameterValue
    Actor learning rate0.0003
    Critic learning rate0.0003
    Neural network size256×512
    Batch size64
    OptimizerAdam
    Discount factor0.99
    Alpha learning rate0.0003
    Police noise0.2
    Table 2. Network parameter settings of SAC, DDPG, and TD3 algorithms
    Jiacheng Wu, Meng Cai, Yujie Lu, Nanshun Huang, Chao Feng, Zhentang Zhao. Reinforcement Learning for Free Electron Laser Online Optimization[J]. Acta Optica Sinica, 2023, 43(21): 2114002
    Download Citation