Long-Term Decision-Optimal Access Mechanism in the Large-Scale Satellite Network: A Multiagent Reinforcement Learning Approach

The large-scale low-Earth orbit (LEO) satellite network presents an obvious challenge for user access with obviously dynamical coverage resulting from the fast-changing locations of LEO satellites, i.e., time-varying overlapped range in spatial and temporal coverage by multiple overhead satellites, making existing studies low throughput for supporting various users with fluctuating access demands. In this article, we propose a multiagent deep deterministic policy gradient-based access (MADDPGA) mechanism for the large-scale satellite network, which allows each user to adjust its strategy autonomously by learning the changing network conditions in the long term. By solving a throughput-maximizing optimization problem, we develop a fully decentralized multiagent deep reinforcement learning (MADRL) algorithm by exploiting optimal dormancy probability (DP) and a well-designed weighted allocation strategy. The simulation results show that the proposed method can effectively improve the throughput performance compared with the Random algorithm and fixed DP algorithm.