Controle adaptativo de pesos de QoS e equidade no escalonamento de downlink 5G via aprendizado por reforço profundo

Abstract

Downlink resource scheduling in 5G networks carrying mixed traffic requires balancing three conflicting objectives: aggregate throughput, fairness, and deadline compliance for latency-sensitive flows. Conventional schedulers optimize one axis at the expense of the others, while quality-of-service (QoS)-aware designs rely on weights fixed at design time that do not track varying network conditions. This work presents the DRLQoSScheduler, a downlink scheduler whose latency and reliability urgency weights and fairness target are tuned at run time by an agent trained with the deep deterministic policy gradient (DDPG) algorithm. Rather than allocating resource blocks directly, the agent applies indirect parametric control to a QoS-weighted proportional fair metric, preserving the interpretability of the allocation and avoiding the per-slot optimization required by conventional approaches. The allocation is computed in closed form, with complexity linear in the product of the number of resource blocks and the number of users. The scheduler was implemented in the Vienna 5G System Level Simulator and evaluated on a three-tier heterogeneous network across four scenarios varying user density, mobility, and the latency-sensitive traffic deadline, with 30 independent runs per configuration. Across 36 two-tailed t-tests against three native baselines, the proposed scheduler is never outperformed with statistical significance and prevails in 25, with deadline compliance favorable in all 12 comparisons of that metric. Under saturation, at eight times the reference density with unchanged topology, it is the only one of the five evaluated schedulers ranking second in all three metrics simultaneously.