QoS-Aware Radio Resource Allocation in UAV-Assisted NOMA Networks with Multi-Agent Soft Actor-Critic Learning

In UAV-aided NOMA cellular networks, the joint optimization of user association, trajectory planning, and power allocation poses significant challenges due to high-dimensional continuous action spaces, dynamic user mobility, and stringent QoS requirements. We address this complex resource allocation problem by proposing a two-stage optimization framework integrating distance-aware clustering with multi-agent deep reinforcement learning. Specifically, we formulate a QoS and fairness radio resource optimization for throughput maximization in a UAV-aided NOMA cellular offloading network. We develop a distance-aware dynamic user clustering (DDUC) algorithm to ensure NOMA-compatible user grouping. We propose a multi-agent soft actor-critic (MASAC) framework to enable collaborative learning among UAV agents through experience sharing, incorporating tiered QoS penalty mechanisms and fairness constraints. Simulation results demonstrate that DDUC outperforms K-means by ensuring NOMA-specific constraints, while MASAC achieves superior scalability across different UAV fleet sizes and maintains approximately 99% QoS satisfaction while significantly improving the system throughput when comparing with other benchmarks by at least 67%.