Attention-Enhanced Meta-Reinforcement Learning Tackles 5G URLLC-eMBB Resource Conflicts

Denise Maddox — Tue, 27 Oct 2026 16:02:40 +0000

Researchers have proposed an attention-enhanced meta-graph reinforcement learning framework to manage the coexistence of two very different 5G services sharing the same radio resources. Ultra-reliable low-latency communications (URLLC) demand guaranteed low latency and reliability, while enhanced mobile broadband (eMBB) requires high throughput and fairness. The work, published as an open access Research article in the Journal of Big Data, addresses dynamic scheduling of orthogonal resource blocks in a single-cell environment under time-varying channels, bursty traffic and shifting traffic distributions.

The framework models resource competition at the MAC layer as a dynamic contention graph. A permutation-equivariant GATv2 encoder learns the interactions among users, allowing the scheduler to account for how competing flows affect one another. Model-agnostic meta-learning enables the model to adapt quickly to new traffic conditions rather than requiring lengthy retraining when the load pattern changes.

Scheduling proceeds in two stages. The algorithm first selects the most prioritized user, and the remaining resources are then allocated with a proportional-fair scheduling rule. This combination is intended to protect URLLC transmissions from violations while preserving reasonable throughput and fairness for eMBB users.

Under the nominal evaluation condition, the proposed method achieved a URLLC violation rate of 63.11%, compared with 90.54% for a DQN baseline and 79.52% for PPO, while attaining a Jain’s fairness index of 0.945. These reported figures come from the article’s abstract; the evaluation is described for a single-cell setting, and the paper is an early-access version that remains subject to further edits before the final Version of Record.

The study frames the URLLC-eMBB resource allocation problem as one of balancing competing guarantees: low latency and reliability on one side, throughput and fairness on the other. By representing contention as a graph and applying attention-based encoding, the approach aims to capture the structure of user interactions that simpler schedulers miss, while meta-learning addresses the variability of real traffic over time.

Subject of Research: A reinforcement learning framework for scheduling shared 5G radio resources between URLLC and eMBB services.

Article Title: Robust URLLC-eMBB coexistence in 5G networks: an attention-enhanced meta-reinforcement learning framework

Article References: Dai, W., & Li, L. (2026). Robust URLLC-eMBB coexistence in 5G networks: an attention-enhanced meta-reinforcement learning framework. Journal of Big Data. https://doi.org/10.1186/s40537-026-01586-3

Image Credits: AI Generated

DOI: 10.1186/s40537-026-01586-3

Keywords: 5G, URLLC, eMBB, reinforcement learning, resource allocation, GATv2, meta-learning, scheduling, Jain's fairness index, Journal of Big Data, Robust, URLLC-eMBB

deep reinforcement learning for 5G resource allocation – Science

Attention-Enhanced Meta-Reinforcement Learning Tackles 5G URLLC-eMBB Resource Conflicts