Fig. 1. Structure of the low-latency deep-reinforcement learning algorithm based on DDPG strategy in the laser environment.

Fig. 2. Flow chart for stable mode-locked state monitoring.

Fig. 3. Experimental setup of UFL based on SA. WDM, 980/1550 nm wavelength division multiplexer; EDF, erbium-doped fiber; EPC, electrical polarization controller; FPGA, field-programmable gate array; SA, saturable absorber; OC, optical coupler; ISO, isolator; PD, photodetector.

Fig. 4. Characterization of the output when the laser is in the FML state. (a) Time-domain pulse output within 0.2 μs laboratory time. The pulse interval is 25.34 ns. (b) Frequency-domain signal characterization in 4 GHz bandwidth. The frep is 39.459 MHz. (c) The pulse autocorrelation and the result of fitting using sech2 function. (d) The spectrum of the laser output.

Fig. 5. Comparison of the time-domain and frequency-domain signal output by the laser in different polarization states. From left to right: FML state and Q switch mode-locked state. In each column, the top row shows the time-domain signal within 40 μs laboratory time, and the bottom plot is the frequency-domain signal within the 4 GHz bandwidth.

Fig. 6. Effect diagram of algorithm recovery after the laser loses mode-locked state due to motor vibration. (a) The convergence curve of the reward value in the last 100 rounds of stable mode-locked calculation model-training iterations. (b) The frep and power change of the laser output during the process of applying vibration to the laser and starting the recovery algorithm. (c) The recovery time statistics of 1500 vibration tests. (d) The output frep and power change of the system within 10 min under the condition of vibrating for 1.5 s per minute and running the mode-locked recovery algorithm all of the time.

Fig. 7. Changes in the reward value of the designed algorithm model in the system at different temperatures. (a)–(f) are the results of the last 100 rounds of training recorded from 15°C to 40°C.