Reinforcement learning-based wavelength-dependent power control for WDM systems

In dynamic optical networks, wavelength-dependent power control is a challenging issue because it can dramatically affect lightpath quality of transmission. To address this issue, the authors proposed a reinforcement learning (RL) channel power equalization method to compensate EDFA wavelength-dependent gain in a single step. The optical power of the active WDM channels is monitored at the endpoints of the optical multiplex section (OMS) to learn the best policy for optimizing the variable attenuation elements of the reconfigurable optical add-drop multiplexer (ROADM). The proposed approach is validated experimentally on a three-span WDM experimental testbed, where a surrogate model of the RL environment significantly reduces the management effort required to collect samples. The applicability of the RL method to our experimental system demonstrates an average power difference reduction up to 87%, which was obtained for the 24-channel random allocation use case.