A comparative study of deep reinforcement learning algorithms in ultrafast mode-locked fiber lasers

Trending with the recent artificial intelligence development, deep reinforcement learning (DRL) algorithms have found vibrant applications for ultrafast mode-locked fiber lasers in practice. However, there has been lacking a comprehensive and detailed study regarding the algorithms’ performance in these laser systems up to date. In this paper, we present a comparative study of three popular DRL algorithms, the Double Deep Q Network (DDQN), Twin Delayed Deep Deterministic policy gradient (TD3), and Soft Actor-Critic (SAC), which are theoretically implemented in nonlinear-polarization-rotation-based mode-locked fiber laser systems. In our DRL performance simulations, the reward function is designed specifically to maximize the tendency towards mode-locking by the introduction of kurtosis in frequency domain. The simulated training processes indicate that the TD3 and SAC are outperforming DDQN in the multi-input context. It has been found that while DDQN has difficulty in handling multiple inputs and tends to have issues with convergence, the TD3 is the most stable and efficient, and the SAC is more advantageous in searching for various states. The acquired pulse evolution diagrams of the mode-locked states also have confirmed the feasibility of training ultrafast mode-locked fiber lasers with these DRL algorithms. Finally, an experimental ultrafast fiber laser system based on the active Ho:ZBLAN fiber and TD3 algorithm has been successfully demonstrated.