Multi-Domain Virtual Network Embedding in LEO Satellite Networks: A Stackelberg Game Guided Multi-Agent Reinforcement Learning Approach

Low Earth orbit (LEO) satellite networks (LSNs) are a pivotal component of sixth-generation (6G) mobile communications, providing ubiquitous coverage, elastic bandwidth, and low latency for 6G services. Virtual network embedding (VNE) maps heterogeneous service requests onto the satellite substrate, enabling online resource allocation in dynamic constellations. However, existing centralized and distributed methods struggle to balance signaling overhead and real-time global optimality in dynamic, large-scale LSNs. To overcome these limitations, we propose a Stackelberg-guided multi-agent reinforcement learning (SG-MARL) algorithm for Multi-Domain VNE (MD-VNE). First, we formulate the problem as a Stackelberg game where a global leader coordinates domain followers via QoS-based lightweight price-quota signals. A cut-aware request decomposition strategy with a feasibility-aware action mask is devised to reduce the action space and enhance embedding efficiency. Furthermore, an alternating bi-level training scheme is developed to enhance learning stability and generalization for dynamic topologies. Simulations on Walker-Delta topologies demonstrate SG-MARL's superiority over baselines in acceptance ratio, revenue efficiency, latency, and resource utilization.