Toward real-time onboard resource scheduling in LEO satellite networks: a decentralized multi-agent reinforcement learning approach

Low Earth Orbit (LEO) satellite systems are pivotal for achieving global 6G connectivity, yet they face significant challenges in user access and handover due to their dynamic topologies and limited onboard resources. A core issue lies in the joint optimization of beam hopping selection, user scheduling, and power allocation, which is an NP-hard mixed-integer nonlinear programming problem with complex coupling across time, frequency, space, and power domains. To address this, the paper formulates the problem as a cooperative Decentralized Partially Observable Markov Decision Process (Dec-POMDP) and proposes a novel decentralized multi-agent framework, called Beam-level Agentic Resource Scheduling (BARS). In this framework, each beam acts as an autonomous agent that makes its own decisions based on localized observations, while implicit coordination among agents is achieved through shared team rewards, interference penalties, and power feedback, eliminating explicit inter-agent communication. The training framework, built upon the Recurrent Multi-Agent Proximal Policy Optimization (RMAPPO) algorithm, directly optimizes group delay and packet loss rate as objectives. Simulation results demonstrate that the BARS framework achieves superior delay performance and lower packet loss than centralized and conventional baselines. Notably, its inference overhead does not scale with the number of beams, making it highly suitable for real-time onboard deployment in large-scale LEO satellite networks.