Hierarchical Graph-Based Multi-Agent Reinforcement Learning for Resource Allocation in Heterogeneous Vehicular Networks

Vehicular networks face significant challenges in resource allocation due to heterogeneous Quality-of-Service (QoS) requirements and dynamic channel conditions. This paper proposes an efficient resource allocation framework for heterogeneous vehicular networks to address diverse QoS requirements in dynamic environments. We introduce a Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm enhanced by Graph Attention Networks (GATs), named MADDPG-GAT, to optimize sub-bands allocation and power control for Vehicle-to-Vehicle (V2V) links, while minimizing interference to Vehicle-to-Infrastructure (V2I) links. By constructing a hierarchical relationship graph to capture intra-group and inter-group interactions among vehicles, the proposed MADDPG-GAT leverages local and global embeddings for cooperative decision-making. The framework employs centralized offline training for efficient policy learning and decentralized online execution for real-time decision-making with low overhead, with normalized reward function to enhance convergence and training stability. Simulation results demonstrate that our approach outperforms baseline methods in terms of system performance and robustness.