Multi-Agent Deep Reinforcement Learning for Wireless-Powered UAV Networks

International audienceUnmanned Aerial Vehicles (UAVs) have attracted much attention lately and are being used in a multitude of applications. But the duration of being in the sky remains to be an issue due to their energy limitation. In particular, this represents a major challenge when UAVs are used as base stations (BSs) to complement the wireless network. Therefore, as UAVs execute their missions in the sky, it becomes beneficial to wirelessly harvest energy from external and adjustable flying energy sources (FESs) to power their onboard batteries and avoid disrupting their trajectories. For this purpose, wireless power transfer (WPT) is seen as a promising charging technology to keep UAVs in flight and allow them to complete their missions. In this work, we leverage a multi-agent deep reinforcement learning (MADRL) method to optimize the task of energy transfer between FESs and UAVs. The optimization is performed by carrying out three essential tasks: (i) maximizing the sum-energy received by all UAVs based on FESs using WPT, (ii) optimizing the energy loading process of FESs from a ground BS, and (iii) computing the most energy-efficient trajectories of the FESs while carrying out their charging duties. Furthermore, to ensure highlevel reliability of energy transmission, we use directional energy transfer for charging both FESs and UAVs by using laser beams and energy beam-forming technologies, respectively. In this study, the simulation results show that the proposed MADRL method has efficiently optimized the trajectories and energy consumption of FESs, which translates into a significant energy transfer gain compared to the baseline strategies