Constrained Deep Reinforcement Learning for Joint Energy-Efficient Power Control and Link Switching in Hybrid FSO-RF Systems

Abstract

Hybrid free space optics-radio frequency (FSO-RF) backhaul systems offer a solution to high capacity, weather-robust wireless infrastructure; however, jointly optimising transmit power and link selection under dynamic atmospheric conditions remains challenging. These systems must balance outage avoidance, switching stability and power efficiency. Existing approaches either focus on link switching alone or apply machine learning for prediction; yet the controller does not learn the optimal policy because the decisions for power control and switching are rule-based. In this paper, a reliability-first, energy-aware deep reinforcement learning (DRL) control framework for hybrid FSO-RF operation is presented, where the agent jointly selects the free space optics (FSO) transmit power level and the active transmission link under weather-induced channel degradation. Incorporating a Light Gradient Boosting Machine (LightGBM) one-step-ahead visibility forecast into the agent’s state representation, the controller anticipates atmospheric degradation before it occurs rather than reacting to it. Over a 1 km link, the proposed policy achieves zero post-shield outage while more closely tracking the minimum feasible FSO power than the benchmark policies it is compared against - it outperforms Soft Actor Critic (SAC) in eight of ten paired runs, with a median improvement of 38.29 percentage points. It also draws the least average total transmit power of the power-adaptive policies evaluated, largely because of minimal usage of the RF link. Together, the results demonstrate that constrained deep reinforcement learning can serve as a practical and proactive control strategy for energy-aware hybrid FSO-RF backhaul under real meteorological conditions.