Multiagent Deep-Reinforcement-Learning-Based Channel Allocation for MEO–LEO Networked Telemetry System

Numerous mega low-Earth orbit (LEO) satellite constellation plans have recently become a significant part of the future satellite communication era. Since the existing ground-based and geostationary Earth orbit (GEO)-based telemetry system is unsuitable for monitoring the working status of mega LEO constellations, the networked satellite telemetry system is used to achieve the full time, low delay telemetry in this article, which is a significant scenario of satellite Internet of things. In order to satisfy the data transmission requirements of extensive satellites, this article formulates the channel allocation problem, which aims at maximizing the total transmitted data value by allocating multiple medium-Earth orbit (MEO) beams in multiple time slots to serve multiple LEO satellites. Considering that the data generation states of LEO satellites are hybrid constant and stochastic, that the MEO satellites could allocate channels more timely than the ground mission center, and that the action space for channel allocation is too large, the multiagent deep-reinforcement-learning-based algorithm is adopted to solve the channel allocation problem. Furthermore, this article designs the connections of the output layer of the deepQnetwork so as to reduce the computation and storage overhead. Finally, the upper bound performance (UBP) of the channel allocation problem is analyzed and numerical simulation is performed to verify the effectiveness of our proposed channel allocation algorithm.