Computer Science editorial
LOLLA: Deep Reinforcement Learning for Closed-Loop Link Adaptation Towards a GPU-Accelerated AI-RAN
The core problem
Outer-loop link adaptation (OLLA) is a widely deployed mechanism in 5G NR for tracking channel variations and compensating for residual SINR estimation errors. However, OLLA relies on first-order, single-bit feedback (ACK/NACK) and a fixed staircase update rule, which degrades performance significantly under high-mobility and fast-varying channel conditions. The core limitation is that OLLA cannot access rich PHY/MAC telemetry—such as channel quality indicators, Doppler estimates, or buffer states—that could inform more agile adaptation.
This paper introduces LOLLA (Learned Outer-Loop Link Adaptation), a deep reinforcement learning framework that replaces the conventional OLLA staircase with a learned, continuous SINR offset. The offset is conditioned on rich PHY/MAC telemetry inaccessible to OLLA and modulates the SINR-to-MCS lookup table, thereby preserving 3GPP-compliant MCS selection while provably subsuming the conventional OLLA update rule. The framework is realized as the first closed-loop AI-native control dApp on a GPU-accelerated 5G NR stack, targeting end-to-end control latencies under 500 microseconds.
Innovation
Evaluations under 3GPP TDL channel models demonstrate that LOLLA achieves 15% to 92% throughput gains over conventional OLLA across Doppler frequencies up to 400 Hz. The gains are most pronounced under high-mobility scenarios where OLLA's first-order feedback struggles to track fast channel variations. LOLLA attains a Pareto frontier that strictly dominates OLLA across all evaluated reliability targets (BLER from 1% to 15%), meaning that for any given reliability level, LOLLA provides higher throughput than OLLA.
The learned policy generalizes to unseen channel models, indicating that the policy has learned robust adaptation strategies rather than overfitting to specific channel conditions. Furthermore, LOLLA scales to eight concurrent user equipments (UEs) under shared-resource scheduling, demonstrating its applicability in multi-user scenarios. The end-to-end control latency remains under 500 microseconds, which is critical for real-time operation in 5G NR.
Key quantitative results are summarized in the table below:
| Metric | OLLA | LOLLA | Improvement |
|--------|------|-------|-------------|
| Throughput gain (Doppler up to 400 Hz) | Baseline | 15–92% higher | 15–92% |
| BL
Why it matters
The results confirm that replacing the OLLA staircase with a learned continuous offset yields substantial performance improvements, particularly in high-mobility scenarios. The use of a Lagrangian BLER constraint allows automatic enforcement of reliability targets without manual penalty calibration, which is a significant practical advantage over conventional OLLA where the step size and up/down thresholds must be manually tuned. The provable subsumption of OLLA ensures that LOLLA can never perform worse than OLLA in the worst case, providing a safety guarantee.
The GPU-accelerated implementation achieves sub-500 µs control latency, enabling closed-loop operation at the timescales required for 5G NR link adaptation. The uplink formulation, where the gNB directly observes decoding outcomes, ensures simulation-to-deployment parity, which is often a challenge in applying reinforcement learning to real-world systems. The generalization to unseen channel models and scaling to eight concurrent UEs suggest that LOLLA is a practical candidate for deployment in AI-RAN architectures.
However, the paper does not discuss potential overheads of the GPU-accelerated stack or the energy consumption of the deep reinforcement learning inference. Future work could explore transfer learning to further reduce training time and investigate the impact of imperfect telemetry. Overall, LOLLA represents a significant step towards AI-native radio access networks, demonstrating that deep reinforcement learning can be integrated into 5G NR with strict latency and reliability constraints.
Who should read this
Opening member content…