Joint Relay Selection and Power Allocation in Free-Space Optical Communication with Reinforcement Learning

This paper studies the problem of joint relay selection and power allocation in Free Space Optical(FSO) communication based on Reinforcement Learning (RL). This is a challenging problem since the action space is hybrid, consisting of continuous action(power allocation) and discrete actions(relay selection).Instead of discretizing the continuous action into a set of discrete options and applying conventional RL methods, such as deep Q learning (DQN), we propose a model-free Joint Relay Selection and Power Allocation algorithm (MF-JRSPA),which operates directly in a hybrid action and can be implemented without prior knowledge of the system model, aiming to optimize system capacity. Simulation results demonstrate that MFJRSPA algorithm has higher system capacity than the baseline, DQN, and PDDL algorithm under the peak power and total power constraints.