AdaptSky: A DRL Based Resource Allocation Framework in NOMA-UAV Networks

Unmanned aerial vehicle (UAV) has recently attracted a lot of attention as a candidate to meet the 6G ubiquitousconnectivity demand and boost the resiliency of terrestrialnetworks. Thanks to the high spectral efficiency and low latency, non-orthogonal multiple access (NOMA) is a potential access technique for future communication networks. In this paper, we propose to use the UAV as a moving base station (BS) to serve multiple users using NOMA and jointly solve for the 3D-UAV placement and resource allocation problem. Since the corresponding optimization problem is non-convex, we rely on the recent advances in artificial intelligence (AI) and propose AdaptSky, a deep reinforcement learning (DRL)-based framework, to efficiently solve it. To the best of our knowledge, AdaptSky is the first framework that optimizes NOMA power allocation jointly with 3D-UAV placement using both sub-6GHz and millimeter wave mmWave spectrum. Furthermore, for the first time in NOMA-UAV networks, AdaptSky integrates the dueling network (DN) architecture to the DRL technique to improve its learning capabilities. Our findings show that AdaptSky does not only exhibit a fast-adapting learning and outperform the state-of-the-art baseline approach in data rate and fairness, but also it generalizes very well.