Traffic Prediction-based Multi-Agent HDRL for Cooperative Resource Allocation in Heterogeneous LEO Networks

This paper proposes a hierarchical multi-agent deep reinforcement learning (HDRL) framework for heterogeneous Low Earth Orbit (LEO) satellite networks. The framework addresses the timing mismatch between the execution and the decision by incorporating delay-aligned traffic predictions into the agent state, structuring the control so that an upper-level agent selects an active set of cells while lower-level agents allocate beams and channels under visibility constraints per-satellite. Using a shared reward composed of demand-weighted efficiency, a duplication penalty for identical cell-channel pairs, and a coverage bonus, the framework reduces early exploration loss and converges to higher rewards, while suppressing resource duplication and jointly improving efficiency and coverage—yielding a practical cooperative policy for heterogeneous LEO systems.