Radio Resource Allocation Optimization for Delay-Sensitive Services Based on Graph Neural Networks and Offline Dataset

Delay-sensitive services, particularly Ultra-Reliable and Low-Latency Communications (URLLC) services, are envisioned as one of the primary applications for 5G and forthcoming 6G networks. However, guaranteeing the stringent Quality of Service (QoS) requirements through efficient resource allocation remains a formidable challenge. While Reinforcement Learning (RL) presents a promising solution, conventional online RL methodologies typically necessitate direct interactions with practical wireless networks during training, posing potential risks to URLLC services if suboptimal actions are taken. In contrast, this paper introduces a novel offline data-driven optimization framework for designing radio resource allocation algorithms. Inspired by the dynamic nature of the wireless network inherently captured in the historical operation data traces, our proposed framework takes advantage of offline RL to optimize resource allocation policies solely from offline datasets, thus avoiding the risks associated with online training on practical networks. Moreover, to effectively characterize dynamic fluctuations in the number of active users, our framework incorporates Graph Neural Networks (GNNs), where the GNN-based architecture can enhance the policies' scalability and generalization capabilities, enabling the learned policies to adapt to evolving network conditions. Building upon this framework, we propose a novel radio resource allocation algorithm, namely GNN-based offline RL (GNN-ORL). Extensive experimental results validate the performance of the proposed GNN-ORL algorithm. Specifically, GNN-ORL exhibits both stable convergence and excellent generalization capabilities, and outperforms baseline schemes in meeting the QoS requirements of URLLC services.