Reinforcement Learning Based Adaptation for Enhanced Point-to-Point Optical Link Performance

With introduction of 5G, fiber to home, IoTs and growing data centers, world needs higher bandwidth to keep the world moving and connected. Along with high speed, reliability of data transfer is most important factor because of VoIP applications, Banking transactions, Stock Exchange live updates etc. All this communication needs low latency, fast and reliable data transfer. High speed connectivity is achieved by strong backhaul optical network. Fiber capacity is increased by higher rate modulations. In past decade it has been witnessed that bandwidth per wave has increased from 10G per wave to 100G, 200G, 400G, 800G, 1200G, 1600G per wave. Higher baud rates are prone to errors and reach is shorter. We propose a reinforcement-learning (RL) based framework to enhance the performance of a point-to-point optical link by dynamically adapting link parameters in response to varying channel conditions and system impairments. This method will increase the quality of signals and go for longer reach. We model the optical link as an environment in which an RL agent chooses actions (e.g., modulation format, forward error-correction rate, launch power) based on observations of current link state (e.g., Q-Factor, Signal-to-Noise Ratio, Dispersion/ Polarized Mode Dispersion (PMD) metrics, BER estimate). Through simulation and real-world emulation, we show that the RL agent converges to a policy that improves throughput and reliability compared to fixed or heuristically tuned parameter settings. This work demonstrates the potential of model-free learning methods in optical communication links, providing a path toward self-optimizing optical systems.