Reinforcement Learning-Driven UC-CFmMIMO for UAVs: A Subband Allocation Framework

This is the authors' abstract. We don't add key points for this paper.

Aerial vehicular networks are emerging as a critical use case in sixth-generation (6G) wireless systems, demanding ultra-reliable, low-latency connectivity across dynamic, three-dimensional environments. To meet these stringent requirements, user-centric cell-free massive MIMO (UC-CFmMIMO) has gained traction as a promising architecture, offering uniform coverage and high resilience. However, integrating UAVs introduces unique challenges in dynamic frequency resource allocation due to altitude-dependent propagation, mobility, and interference management with dynamic subcarrier spacing. To address this, we formulate a multi-objective resource allocation problem targeting frequency assignment, Spectral Efficiency (SE), and Quality of Service (QoS) preservation. We propose a Deep Deterministic Policy Gradient (DDPG) Reinforcement Learning (RL) approach that balances interference mitigation, SE, and fairness under strict bandwidth constraints. Additionally, we compare DDPG with two bio-inspired metaheuristics Binary Gray Wolf Optimizer (BGWO) and Binary Moth Flame Optimizer (BMFO) to assess their effectiveness. Results show that UC-CFmMIMO can meet 6G performance targets in aerial networks, with RL demonstrating superior adaptability in complex, altitude-varying environments.