Federated Reinforcement Learning-Based UAV Swarm System for Aerial Remote Sensing. (29th April 2022)

• Record Type: Journal Article
• Authors: Lee, Woonghee
• Other Names: Khosravi Mohammad R Academic Editor.
• Is Part Of: Wireless communications and mobile computing. Volume 2022(2022)
• Publication Date: 2022-04-29
• Subjects: Wireless communication systems -- Periodicals; Mobile communication systems -- Periodicals; 621.38205
• DOI: 10.1155/2022/4327380
• Languages: English
• ISSNs: 1530-8669

Abstract

Abstract : In recent years, due to the development of technologies for unmanned aerial vehicles (UAVs), also known as drones, UAVs have developed rapidly. Because of UAVs' high mobility and computational capability, UAVs have a wide range of applications in Industrial Internet of Things (IIoT), such as infrastructure inspection, rescue, exploration, and surveillance. To accomplish such missions, it is more proper and efficient to utilize multiple UAVs in a swarm, rather than a single UAV. However, it is difficult for an operator to understand and control numerous UAVs in different situations, so UAVs require the significant level of autonomy. Artificial intelligence (AI) has become the most promising combination with UAVs to ensure the high autonomy of UAVs by establishing swarm intelligence (SI). However, existing learning methods for building SI require continuous information sharing among UAVs, which incurs repeated data exchanges. Thus, such techniques are not suitable for constructing SI in the UAV swarm, in which communication resources are not readily available on unstable UAV networks. To overcome this limitation, in this paper, we propose the federated reinforcement learning- (FRL-) based UAV swarm system for aerial remote sensing. The proposed system applies reinforcement learning (RL) to UAV clusters to establish the SI in the UAV system. Furthermore, by combining federated learning (FL) with RL, the proposed system constructs the more reliable and robust SI for UAV systems. We conducted diverse evaluations, and the results show that the proposed system outperforms the existing centralized RL-based system and is more suited for UAV swarms from a variety of perspectives.