Graph Convolutional Multi-Agent Reinforcement Learning for UAV Coverage Control

Benefiting from the convenience of aerial flying, unmanned aerial vehicles (UAVs) can provide linear accessibility and dynamic adjusted coverage scheme to ground users, thus promising to act as removable base stations (BSs). Therefore, it necessitates the UAVs to work as a group or team and cooperate with each other, since the embedded computational and energy resources of UAVs lead to relevantly limited coverage range. In typical UAV groups, neighboring UAVs can create connections, forming a dynamic local network, to which UAVs are encouraged to stay connected due to the limited gateway resources. The connection form and networking mode of UAVs group have the characteristics of graph, which can help improve the group performance. Connections bring information sharing while the dynamic characteristics of the connections make the network topology changing with the UAVs' moving. Taking these into consideration, we utilize a graph convolutional based multi-agent reinforcement learning (MARL) method for UAVs group controlling. The proposed method is able to capture and take advantage of the mutual interplay between UAVs, so as to effectively improve the signal coverage as well as fairness and reduce the overall energy consumption in the meantime. Extensive simulation results verify the effectiveness of the proposed method.