Deep Reinforcement Learning Applied to Congestion Control in Fronthaul Networks

Fifth-generation wireless technologies embrace more flexible network architectures as a way of reducing deployment and operation costs while increasing user satisfaction. Centralized Radio Access Networks (C-RANs) play a fundamental role in this context, being envisioned for increased flexibility and lower cost of deployment. More recent C-RAN architectures assume packetized fronthaul links connecting radio units to baseband processors, a more cost-efficient solution relying on statistical multiplexing. This shared infrastructure scenario brings new challenges, including network congestion in the fronthaul links. Since current solutions may neither scale nor react in time for the microsecond-order delay requirements, this paper evaluates the adoption of machine learning-based techniques for congestion control in C-RAN scenarios. Deep Reinforcement Learning methods were evaluated through discrete-event simulations and compared with legacy TCP-based solutions. Promising results were found with satisfactory throughput level in all simulated scenarios, achieving low rates of average delay and packet loss compared with the TCP congestion control baseline.