Reinforcement Learning for Power Management in Low-margin Optical Networks

Sze Ka Tse, Xiaoyang Zhao, Anita Chan, Di Tang, Aanchan Mohan, Carlos Natalino, Michal Aibin

Abstract: We explore the Q-learning and Proximal Policy Optimization (PPO) algorithms for solving the Routing, Wavelength, and Power Allocation (RWPA) problem. The results indicate that reinforcement learning can significantly optimize launch power in low-margin optical networks, with Q-learning achieving better results than PPO, delivering improvement in GSNR of over $175 \%$ and over $48 \%$ in small and large networks, respectively.

External IDs:dblp:conf/icton/TseZCTMNA24