User Centric Satellite Handover for Multiple Traffic Profiles Using Deep Q-Learning

Multiple Low Earth Orbit (LEO) satellites have recently been launched in constellations to insure direct Internet access to users anywhere and at any time. Due to the high-speed mobility of LEO satellites, users undergo multiple handovers (HO)s during their service time, which has a negative impact on users' Quality of Service (QoS) if occurred in high frequency. Moreover, next-generation communication technologies are designed to support a wide spectrum of applications, including Artificial Intelligence, Virtual Reality, and Internet of Things (IoT). Thus, differentiating User Equipments (UEs) with different and varying Traffic-Profiles (TP) has become necessary due to each application's unique performance requirements. However, LEO satellites have limited onboard resources and the launched constellations ensure that each UE will be covered by more than one LEO satellite at any given moment, making it challenging to select the optimal satellite at any given time to assure the optimum QoS. Therefore, a satellite HO strategy has to effectively use the few available satellite resources and prevent network congestion while respecting the various resource requirements per TP. To address all the above requirements, we propose a user-centric Multi-Agent Deep Q-Network (MADQN) satellite HO strategy, that is the first in the state of the art to address the variety and diversity of UEs' performance requirements and generated traffic statistics. Our method showcases a significant achievement of approximately 60% reduction in HO rate and around 91% reduction in blocking rate compared to conventional single criterion approaches