Dual-Attention Deep Reinforcement Learning for Multi-MAP 3D Trajectory Optimization in Dynamic 5G Networks

5G and beyond networks need to provide dynamic and efficient infrastructure management to better adapt to time-varying user behaviors (e.g., user mobility, interference, user traffic and evolution of the network topology). In this paper, we propose to manage the trajectory of Mobile Access Points (MAPs) under all these dynamic constraints with reduced complexity. We first formulate the placement problem to manage MAPs over time. Our solution addresses time-varying user traffic and user mobility through a Multi-Agent Deep Reinforcement Learning (MADRL). To achieve real-time behavior, the proposed solution learns to perform distributed assignment of MAP-user positions and schedules the MAP path among all users without centralized user's clustering feedback. Our solution exploits a dual-attention MADRL model via proximal policy optimization to dynamically move MAPs in 3D. The dual-attention takes into account information from both users and MAPs. The cooperation mechanism of our solution allows to manage different scenarios, without a priory information and without re-training, which significantly reduces complexity.