Multi-Agent Deep Reinforcement Learning Based Handover Strategy for LEO Satellite Networks

Abstract

The high rotation speeds and mega-constellations of low earth orbit satellites (LEO SATs) cause the inter-satellite frequent handovers (HOs) problem which can lead to substantial performance degradation. This letter proposes a novel distributed multi-agent deep Q-network based SAT HO strategy for the LEO SAT networks to simultaneously minimize the number of HOs and maximize the throughputs and the visible times of UEs while satisfying the quality-of-service (QoS) constraints of all UEs. The proposed HO scheme allows UEs to independently and simultaneously perform the HO decision makings based on their own local information, which enables to immediately adapt to the dynamic changes of the LEO SAT network environments. The numerical results demonstrated that our proposed HO strategy achieves the lowest average HO rate and the highest achievable throughputs compared to other conventional HO strategies, while ensuring a higher QoS guarantee time ratio.