EdgeRIC-Enabled RL-Based MAC Scheduling for srsRAN-Based Open RAN System

作者：Annu, Sikander Kathat, Abhilash S, P. Rajalakshmi · 发表于：2025 IEEE 102nd Vehicular Technology Conference (VTC2025-Fall) · 年份：2025 · DOI：10.1109/vtc2025-fall65116.2025.11310299 · 研究领域：Computer Science

Newer intelligent scheduling techniques and open-source RAN (Radio Access Network) architectures have made modern wireless communication systems more effective and adaptive. In this work, Edge RAN Intelligence Controller (EdgeRIC), an edge computing service framework, is integrated with a containerized deployment of the srsRAN stack to enable real-time network optimization using Reinforcement Learning (RL)-based scheduling strategies. Traditional scheduling policies such as Max Weight, Max Channel Quality Indicator (CQI), Proportional Fair (PF), and Round Robin (RR) were benchmarked against RL-based approaches implemented in muApps, utilizing Proximal Policy Optimization (PPO), Q-Learning, and Actor-Critic models. To enhance scheduling performance, we introduced additional state features (e.g., transmission bytes) into the Q-Learning model. We conducted real-time metric supervision using monitoring muApp, along with latency measurements between eNodeB (eNB) and User Equipments (UEs) via ZeroMQ (ZMQ) sockets. The Q-learning and actor-critic models were specifically chosen and designed to maximize performance in highly dynamic environments, where they demonstrated superior adaptability over conventional methods. In particular, the Q-Learning model achieved a higher throughput gain compared to PPO, while the Actor-Critic model achieved similar gains with lower computational complexity. The system was further scaled to support four concurrent UEs, validated using iperf traffic ge...