Abstract

The instability of millimeter wave links in 5G-Advanced and 6G networks creates rapid channel fluctuations that traditional Active Queue Management mechanisms struggle to manage, leading to degraded network performance. To address this, we propose the Reinforcement Learning Queuing Optimizer (RLQO), a Deep Q-Network algorithm deployed at the gNB's Radio Link Control layer. RLQO intelligently manages buffer resources by learning optimal packet dropping decisions based on real-time traffic conditions, balancing throughput, latency, and buffer stability. Extensive simulations using ns-3 and the 5G-LENA module demonstrate that RLQO adapts to diverse application flows and Congestion Control Algorithms (CCAs) better than traditional approaches. The results confirm that RLQO significantly improves throughput and reduces the successful round-trip time (a measure of the delay observed for confirmation of reception of a transmitted packet) in fluctuation-prone environments, enhancing the user experience and optimizing node efficiency within the RAN, with possible implications for reducing infrastructure costs.