Reinforcement Learning Combined With a Self-Retraining Strategy for Dual-Comb State Generation in a Single-Cavity Laser

The gain competition and collision of two asynchronous pulse trains increase the complexity of achieving fast dual-comb state (DCS) generation in a single-cavity dual-comb laser (SCDCL) under varying environmental conditions. Traditional heuristic algorithms merely search for a solution according to the dedicatedly designed objective function, and the solution cannot assist in future optimization once the environment is altered, resulting in poor robustness of the automatic DCS generation. Here, reinforcement learning combined with a self-retraining strategy is employed in a dual-wavelength SCDCL to achieve stable recovery in varying environments. The experimental results demonstrate that reinforcement learning combined with a self-retraining strategy enables the precise control of the SCDCL to generate two asynchronous pulse trains. The mean computation time for the SCDCL to generate DCS is ∼1.57 seconds under the control of reinforcement learning when the environment remains nearly steady, which is 63% of the state of the art. Benefiting from the self-retraining strategy, the mean recovery time of the proposed method is approximately half of that of the conventional heuristic algorithm when the environment changes. The proposed approach is nearly insensitive to the initial state, thereby offering a promising solution to overcome the impact of environmental variations on the SCDCL.