Source-linked AI summary
Improving 5G AI-RAN MCS Selection by Predicting Retransmissions
Tamerlan Aghayev, Maxime Elkael, Michele Polese, Reshma Prasad, Salvatore D'Oro, Yunseong Lee, Koichiro Furueda, Tommaso Melodia
TL;DR
Reactive 5G NR link adaptation can lag behind fast channel changes because it relies on noisy, delayed measurements and HARQ feedback. NOSTRAdAMUS adds a predictive overlay that forecasts near-term retransmissions and corrects an existing policy’s MCS choice, improving goodput and reducing retransmissions across diverse evaluated scenarios.
Problem
Existing 5G NR link adaptation is inherently reactive, so delayed and noisy channel and HARQ information can become obsolete when channels change quickly.
Method
NOSTRAdAMUS uses short-term HARQ history and a machine-learning predictor to adjust the MCS selected by an existing LA policy without replacing it.
Results
Across diverse channels, configurations, mobility conditions, and OTA deployment, NOSTRAdAMUS improves average goodput by up to 71.5% and reduces retransmissions by up to 71.8%.
Takeaways & Limitations
The same predictive overlay generalizes across scenarios and complements both OLLA and SALAD by improving link recovery and reducing unnecessary retransmissions.
Abstract
from arXiv · showhide
Link Adaptation (LA) in 5G NR is inherently reactive, relying on channel measurements and HARQ feedback that may become quickly obsolete when the channel changes quickly. This data is also noisy, making it hard to track accurately, and has to be fed to real-time controllers with feedback-loop effects which are hard to troubleshoot. This explains why most practical deployments select simple but robust algorithms, which accept that the lag can leave the scheduler operating at overly aggressive or unnecessarily conservative rates, trading spectrum efficiency for predictable performance. In this paper, we improve on this status-quo with NOSTRAdAMUS, a predictive LA framework which adds foresight to existing algorithms without replacing or redesigning them. NOSTRAdAMUS predicts whether a retransmission will occur in the next radio frame from recent HARQ history, and applies corrections to the Modulation and Coding Scheme (MCS) selected by the underlying policy. We benchmark several ML models and show that Gradient Boosting achieves 82.9% accuracy overall with high-confidence interventions that are correct 94.2% of the time, and an inference latency of 5.5 μs. We train the model based on data collected Over-the-Air (OTA) on the X5G testbed, using the open-source OpenAirInterface (OAI) 5G stack, NVIDIA Aerial, and COTS O-RAN Radio Units and User Equipments. The model is then deployed as a dApp, which we evaluate OTA as well as on various channels with hardware-in-the-loop channel emulators. This includes 3GPP TDL and CDL channels, SISO and MIMO configurations, and pedestrian and vehicular mobility. Our evaluation shows that without retraining, and across this variety of scenarios, the dApp augments two SOTA LA algorithms, and increases goodput by up to 71.5% while reducing retransmissions by up to 71.8%. This demonstrates the robustness and generalization capabilities of our approach.
I. INTRODUCTION
5G NR link adaptation reacts to delayed, noisy channel and HARQ information, which can produce inefficient MCS choices under changing channels. NOSTRAdAMUS adds predictive retransmission corrections as a plug-and-play overlay without replacing existing LA policies.
- Motivation: Reactive LA acts only after errors occur, limiting its ability to respond when wireless channels change quickly.Fast fading, blockage, and delayed TDD feedback can leave the selected MCS too aggressive or too conservative.
- Motivation: 84% of selections remain below MCS 18, although MCS 18 achieves a median BLER of 8.6% within the 10% target.The static-UE example illustrates conservative operation caused by reactive adaptation.
- Research gap: Prior approaches redesign LA policies or improve their input information, rather than providing a general predictive overlay for existing policies.The paper positions NOSTRAdAMUS as complementary to these approaches.
- Approach: NOSTRAdAMUS predicts future retransmissions with machine learning and proactively corrects the MCS selected by a black-box host LA algorithm.The overlay can augment substantially different policies, including OLLA and SALAD.
- Contributions: Gradient Boosting achieves 94.2% correctness for high-confidence interventions, while evaluation reports up to 71.5% higher goodput and 71.8% fewer retransmissions.The system is evaluated across heterogeneous channels, antenna configurations, mobility conditions, and OTA experiments.
II. SYSTEM MODEL
The system model represents MCS selection as a policy balancing delivered information against retransmission-related resource costs. NOSTRAdAMUS overlays this process by using recent HARQ activity to adjust the host policy’s MCS decision.
- System operation: Each downlink slot schedules UEs, allocates resources, prioritizes pending HARQ retransmissions, and selects an MCS for each new transport block.The model uses frames and TDD-configured downlink slots as its scheduling units.
- Link adaptation: An LA policy maps available information—including CQI, SNR, BLER, previous MCS, and internal state—to an MCS.The selected MCS determines the transmission rate and reliability tradeoff.
- Link adaptation: Higher MCS values carry larger transport blocks but increase retransmission risk, whereas lower values carry fewer information bits over the same resources.This tradeoff motivates the long-run resource-efficiency objective.
- Objective: The objective measures long-run useful information delivered per allocated PRB across all transmission attempts.It accounts for successfully decoded payloads and the resources consumed by initial transmissions and retransmissions.
- Predictive overlay: NOSTRAdAMUS constructs an adjusted policy that uses machine-learning predictions from recent HARQ activity to modify the reference policy’s MCS decision.The overlay is intended to improve tracking despite noisy, non-instantaneous channel information.
III. NOSTRADAMUS
NOSTRAdAMUS overlays a reference link-adaptation policy with a predictor trained on recent first-transmission HARQ outcomes. Its confidence-gated predictions adjust the next frame’s MCS while abstaining when observations or timing are insufficient.
- Architecture: NOSTRAdAMUS constructs a new policy that adjusts the reference policy’s MCS decision using predictions from recent HARQ activity.The overlay preserves the reference policy while adding predictive corrections.
- Training signal: The predictor uses first-transmission HARQ outcomes because retransmissions also reflect HARQ chase combining and therefore capture channel state less accurately.Outcomes from later transmissions are discarded for learning.
- Prediction: NOSTRAdAMUS predicts whether new transport blocks will likely require retransmission in the next frame from a trailing window of recent BLER observations.The feature vector collects the H most recent windowed BLER values.
- Inference: Inference runs once per UE per frame, and the resulting prediction is held for all MCS decisions in the following frame.Each UE is evaluated independently using its own retransmission history.
- MCS correction: Confidence thresholds trigger configurable downward or upward MCS corrections, while intermediate-confidence predictions leave the reference decision unchanged.The correction is bounded by the minimum and maximum MCS values.
IV. TRAINING AND DEPLOYMENT
The model is trained and benchmarked using X5G testbed data and evaluated for both predictive quality and real-time inference suitability. Gradient Boosting performs best overall and is selected for deployment based on its predictive performance and latency.
- Data collection: Training uses real-world data collected on the X5G testbed with an OAI gNB, NVIDIA Aerial PHY, 2×2 MIMO, 40 MHz bandwidth, and an indoor walking user.The setup serves saturated downlink UDP traffic to a Samsung S25 UE.
- Model benchmark: The benchmark covers six models spanning probabilistic, distance-based, tree-ensemble, and neural families.The same feature vector is used across candidate models.
- Model benchmark: Tree ensembles dominate the benchmark, with Gradient Boosting performing best overall across accuracy, precision, recall, and F1.Figure 3 summarizes these four metrics across the candidate models.
- Deployment: 5.5 µs is Gradient Boosting’s lowest inference latency over the evaluated ONNX runs, supporting its selection for deployment.Latency is evaluated over 10,000 runs using the 99th percentile.
- Evaluation: 82.9% accuracy and 75.1% recall are achieved by the selected Gradient Boosting model on the full test set.The confidence gate intervenes on 54.1% of samples with 94.2% precision.
V. EXPERIMENTAL EVALUATION
The evaluation tests whether NOSTRAdAMUS generalizes across channel conditions, mobility, configurations, and deployment settings. Across these scenarios, it improves goodput or reduces retransmissions relative to OLLA and SALAD, with distinct benefits for each base policy.
- Static channel: NOSTRAdAMUS improves mean goodput by 27.1% over OLLA and 37.2% over SALAD on a single-tap flat channel.The OLLA comparison has nearly unchanged BLER (+4.6%), while the SALAD comparison reduces BLER by 71.8%.
- Cross-channel generalization: 5.5–21.3% goodput gains over OLLA and 5.1–20.4% over SALAD persist across unseen 3GPP TDL-A, TDL-B, and TDL-C channels without retraining.The evaluation uses a fixed 30 ns delay spread and no per-channel tuning.
- Frequency selectivity: As TDL-A delay spread grows from 30 to 300 ns, goodput gains persist from 5.5% to 53.7% over OLLA and from 5.1% to 21.6% over SALAD.The base policies lose 78.7% and 65.8% mean goodput, respectively, across that sweep, while the overlay remains useful as frequency selectivity degrades feedback.
- Policy-specific effects: NOSTRAdAMUS benefits OLLA by reducing first-attempt transmissions below MCS 5 by 73.4%, while it benefits SALAD by reducing BLER by 12.0–71.8%.For SALAD, BLER spike probability falls by 41.7% and spike duration by 38.4%.
- Mobility: Under CDL-C mobility, NOSTRAdAMUS shrinks the sub-MCS-5 region by 82.9% on UMa and 75.5% on UMi.It also raises the MCS ≥20 share by 108.3% and 27.0%, respectively, and restores high-rate operation faster for OLLA.
- Over-the-air evaluation: OTA evaluation reproduces the emulation trends: goodput rises 45.2% with OLLA and 14.4% with SALAD.The OLLA result has similar BLER (+3.1%), while the SALAD result reduces BLER by 27.7%.
VI. CONCLUSIONS & FUTURE WORK
NOSTRAdAMUS is a plug-and-play predictive overlay that uses short-term HARQ history to adjust existing link-adaptation policies. Across diverse emulated and OTA settings, it improves goodput and reduces retransmissions without replacing the underlying policies, while future work targets broader deployments and policies.
- Contribution: NOSTRAdAMUS adds foresight to existing LA policies without replacing or redesigning them.The framework proactively adjusts the MCS selected by a base policy.
- Approach: A lightweight Gradient Boosting predictor uses short-term HARQ history to anticipate retransmission activity and adjust MCS decisions.The same predictor augments both OLLA and SALAD across TDL and CDL channels, SISO and MIMO, mobility, and OTA deployment.
- Results: Across evaluated scenarios, average goodput increases by up to 71.5% and retransmissions decrease by up to 71.8%.The reported gains cover both OLLA and SALAD deployments.
- Future work: Future work will evaluate larger multi-UE deployments, uplink LA, and additional base policies including RL-based approaches.These directions extend evaluation beyond the demonstrated deployment settings and policies.