Integrating Aerial and Ground Users: Deep Reinforcement Learning and Convex Optimization for Data Power Control in Cell-Free Networks

This paper addresses the power control problem in cell-free uplink networks, containing both aerial users and traditional terrestrial users in the same network. In particular, we study the data power control optimization problem to maximize the weighted sum of the spectral efficiencys (SEs), where each network user must satisfy its minimum SE demand. By adjusting the SE contribution weights for terrestrial and aerial users, we selectively prioritize certain user groups. Such model enhances current connectivity by addressing the needs and utilization behaviors of different user groups. In this context, we first introduce a centralized solution as the upper bound of solving the problem, which is mainly based on convex optimization. Then, we propose three decentralized solutions to solve the problem. The first solution focuses on coordinate descent techniques and still benefits from the support of convex optimization. Meanwhile, the other two proposed solutions utilize distributed deep reinforcement learning, harnessing advanced machine learning techniques to build adaptive, self-improving systems designed to improve autonomously within intricate environments. Moreover, in the context of our learning-based approaches, we outline practical signaling considerations for a decentralized model. In our model, each user acts as an independent agent and tries to find the best policy locally, while minimizing the need for information exchange. In our analysis, we explore the implications of weighing the performance of ground and aerial users in terms of SE. Finally, the numerical analysis reveals that the performance of our decentralized solutions closely matches that of the centralized and convex approach.