Stackelberg Game-Assisted MATD3 for User Association and Resource Allocation in SAGIN Slicing

Abstract

With the advent of the sixth-generation (6G) era, the rapid growth of access points (APs) and service types in space–air–ground integrated networks (SAGIN) has posed significant challenges to efficient user association and joint resource allocation in multi-layer heterogeneous environments. To address this issue, we develop a three-layer SAGIN slicing framework composed of ground base stations (BSs), unmanned aerial vehicles (UAVs), and low earth orbit (LEO) satellites, where a shared UAV resource pool is introduced to enhance aerial resource coordination. A weighted multi-objective optimization problem is formulated to jointly support high-throughput, low-latency, and wide-coverage services. Due to its non-convex nature, a Stackelberg game pricing (SGP)-based multi-agent deep reinforcement learning (MADRL) framework, namely MATD3-SGP, is proposed. where the original problem is decomposed into user association and resource allocation subproblems. Specifically, a stackelberg game pricing mechanism is designed for user association, and the uniqueness of the stackelberg equilibrium is proven. while the resource allocation problem is modeled as a partially observable Markov decision process (POMDP), and a multi-agent twin delayed deep deterministic policy gradient (MATD3)-based slicing scheme with centralized training and decentralized execution is developed. Simulation results show that the proposed MATD3-SGP framework consistently outperforms benchmark schemes in terms of system utility, throughput, latency, and coverage ratio. In particular, the system utility is improved by approximately 12.19%, 26.32%, 75.37%, and 147.74% compared with MADDPG-SGP, MASAC-SGP, MAPPO-SGP, and the hard-slicing method, respectively.