Multi-UAV Dynamic Wireless Networking With Deep Reinforcement Learning

This letter investigates a novel unmanned aerial vehicle (UAV)-enabled wireless communication system, where multiple UAVs transmit information to multiple ground terminals (GTs). We study how the UAVs can optimally employ their mobility to maximize the real-time downlink capacity while covering all GTs. The system capacity is characterized, by optimizing the UAV locations subject to the coverage constraint. We formula the UAV movement problem as a Constrained Markov Decision Process (CMDP) problem and employ Q-learning to solve the UAV movement problem. Since the state of the UAV movement problem has large dimensions, we propose Dueling Deep Q-network (DDQN) algorithm which introduces neural networks and dueling structure into Q-learning. Simulation results demonstrate the proposed movement algorithm is able to track the movement of GTs and obtains real-time optimal capacity, subject to coverage constraint.