Coherent beam combining (CBC) is a major method for fiber laser power scaling. For many channels or feedback or when cameras as sensors, however, analog feedback loops become inconvenient and digital controllers offer advantages in flexibility. Here, we investigate the feasibility of using end to end reinforcement learning (RL) methods to achieve CBC. This should enable to achieve the goal of maximizing the power without manually setting any locking strategy or system model. In deep RL a neural network is assigned random weights in the beginning and trained to become a usable control policy. The output of our neural network is a 21-component vector which estimates the value for 21 hypothetical changes of the control voltage. For best control, we should pick the maximum and change the output accordingly. To train this we split up the process in 300 step episodes. Between episodes, we update the weights of the neural network. As our test setup, we used a simple two-channel system combining 1550 nm femtosecond pulses. We trained the RL agent and after half a day of training were able to achieve CBC. In the beginning of training, the agent, it tries many different seemingly random actuator movements but eventually, it converges to a suitable policy. After training for two hours this algorithm was capable of maximizing the output power reliably without us injecting any explicit information about optics or dither locking. The power noise of the trained controller was about 1% RMS.

© (2019) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.

Henrik Tünnermann and Akria Shirakawa, "Reinforcement learning for coherent beam combination (Conference Presentation)," Proc. SPIE 10897, Fiber Lasers XVI: Technology and Systems, 108971D (Presented at SPIE LASE: February 06, 2019; Published: 13 March 2019); https://doi.org/10.1117/12.2509327.6013185247001.