Improving 5G AI-RAN MCS Selection by Predicting Retransmissions
I. I NTRODUCTION In 5G NR, the scheduler selects a Modulation and Coding Scheme (MCS) for every Downlink (DL) Transport Block (TB), aiming to optimize spectral efficiency against an inherently dynamic wireless channel. This selection is usually made by a two-loop Link Adaptation (LA) process: an inner loop (ILLA) maps the Channel Quality Indicator (CQI) of the User Equipments (UEs) to an initial MCS; and an outer loop (OLLA) adjusts that mapping with a dynamic offset based on Hybrid Automatic Repeat reQuest (HARQ) feedback [1]. HARQ provides acknowledgment (ACK/NACK) feedback on whether a transmitted TB was successfully decoded, so that
Goodput 100
0.3 75
0.2
50
0.1
25
0
0
Goodput (Mbps)
BLER (boxplot) 0.4
BLER
Abstract—Link Adaptation (LA) in 5G NR is inherently reactive, relying on channel measurements and Hybrid Automatic Repeat reQuest (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 realtime 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 Machine Learning (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 (RUs) and User Equipments (UEs). The model is then deployed as a dApp, which we evaluate OTA as well as on various channels using the same testbed with hardware-in-the-loop channel emulators. This includes 3GPP TDL and CDL channels, single and multi antenna configurations, and pedestrian and vehicular mobility. Our evaluation shows that without retraining, and across this variety of scenarios, the dApp augments two state-of-the-art 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.
Usage (%)
arXiv:2609.09324v1 [cs.NI] 8 Sep 2026
Tamerlan Aghayev∗ , Maxime Elkael∗ , Michele Polese∗ , Reshma Prasad∗ , Salvatore D’Oro∗ , Yunseong Lee† , Koichiro Furueda† , Tommaso Melodia∗ ∗ Institute for Intelligent Networked Systems, Northeastern University, Boston, MA, U.S.A. {aghayev.t, m.elkael, m.polese, re.prasad, s.doro, t.melodia}@northeastern.edu † SoftBank Research Institute of Advanced Technology, Tokyo, Japan {yunseong.lee, koichiro.furueda}@g.softbank.co.jp
40 20 0
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19
MCS
Fig. 1. Reactive MCS selection for a static UE. The loop selects an MCS below 18 in 84% of the time, while MCS 18 is selected in less than 12% of frames despite achieving a median BLER of 8.6%, within the 10% target.
failed ones can be retransmitted. Although this LA architecture is well-established in the industry, it is inherently reactive as actions are taken only after errors have occurred. Unfortunately, being reactive is known to result in suboptimal MCS selection when the channel varies fast [2], [3]. For example, when the most recent CQI (reported periodically) indicates good channel conditions, ILLA maps to a high MCS. However, if the channel degrades suddenly due to fast fading or blockage, a high MCS might lead to decoding errors. In the meantime, the outer loop applies an adjustment derived from earlier positive HARQ feedback and therefore cannot immediately account for this change. This is especially a problem in Time Division Duplexing (TDD) deployments, typical in 5G, as HARQ and CQI feedback are delayed, since the UE needs to wait for Uplink (UL) slots to report about the reception in DL. As a result, the selected MCS can be too aggressive for the current channel, causing failed transmissions and HARQ retransmissions. The same also applies when the channel improves but the reactive loop remains at a conservative MCS because its adjustment still reflects earlier decoding failures, leaving available link capacity unused. Figure 1 shows that in our Over-the-Air (OTA) indoor deployment, this problem occurs even for a static UE in Line-of-Sight (LOS) with the Radio Unit (RU). We use the industry-standard OLLA algorithm [1], parameterized to target 10% of BLER. We observe that even
with this easy scenario, the reactive loop spends 84% of the time below the highest reliably supported operating point of MCS 18, which achieves a median BLER of 8.6%. This delay becomes particularly problematic in fast mobile scenarios [2]. Recent work bridges this gap by redesigning LA algorithms, from enhanced OLLA updates [4], [5] to learned MCS selection using bandits, supervised learning, and Reinforcement Learning (RL) [3], [6]–[8], as well as recent adaptive methods such as SALAD [9] and policy-based LOLLA [2]. Another line of work predicts channel information used by LA through CQI forecasting [10] or Channel State Information (CSI) prediction [11]. Learned and enhanced-loop approaches replace or redesign the LA policy, while channel prediction only improves the information available to it. Neither provides a general way to add predictive foresight to an existing, well-understood loop without changing how it makes MCS decisions. One must therefore either replace the loop or accept its lag. We present NOSTRAdAMUS, a predictive overlay framework that adds foresight to existing LA algorithms, thus improving their effectiveness without replacing them. Rather than learning a new MCS policy from scratch, NOSTRAdAMUS treats the existing LA algorithm as a black-box, and proactively corrects the MCS it selects. Corrections are derived from Machine Learning (ML) predictions of future retransmissions. The NOSTRAdAMUS overlay abstraction decouples the prediction from the LA design, allowing the same predictive mechanism to augment substantially different LA policies, from OLLA to the State-of-the-art (SOTA) SALAD algorithm [9]. To our knowledge, NOSTRAdAMUS is the first predictive framework for LA designed as a plug-and-play overlay and demonstrated end-to-end with commercial radio hardware. Main contributions. We formulate the predictive LA overlaying problem as a binary classification over historical HARQ data, enabling the scheduler to anticipate short-term decoding errors before reactive adaptation can respond. We then benchmark six ML models and select Gradient Boosting [12], whose high-confidence interventions are correct 94.2% of the time. We deploy NOSTRAdAMUS as a dApp [13] and validate it on the X5G testbed [14] with the open-source OpenAirInterface (OAI) stack, commercial RUs and UEs. We demonstrate NOSTRAdAMUS by applying the same predictor to the OLLA and SALAD algorithms. We evaluate it across heterogeneous 3GPP Tapped Delay Line (TDL) and Clustered Delay Line (CDL) propagation models, Single Input, Single Output (SISO) and Multiple Input, Multiple Output (MIMO) configurations, pedestrian and vehicular mobility, and OTA experiments, achieving respectively up to 71.5% and 60.9% higher average goodput than OLLA and SALAD, while reducing retransmissions by up to 71.8%. II. S YSTEM M ODEL We consider the downlink of a 5G gNB serving a set of UEs U = {1, . . . , NUE }. Time is divided into frames indexed by t ∈ Z≥0 , each containing a set Dt of downlink slot indices as per the configured TDD pattern (e.g., 6DS3U). We index downlink slots so that each slot n ∈ Dt belongs to
a unique frame t. At each downlink slot n, the gNB selects a set Un ⊆ U of UEs to serve, allocates radio resources, schedules pending HARQ retransmissions, and determines an MCS for each UE scheduled for a new TB. We consider HARQ priority scheduling, where pending HARQ retransmissions are scheduled before new-data transmissions. We start by formalizing the link adaptation problem, in which the gNB needs to select the MCS for each scheduled UE. For UE u ∈ Un scheduled in downlink slot n ∈ Dt , let Iu,n ∈ I denote the information available to an LA policy when making its decision for UE u in slot n. I is the space of possible information, including Key Performance Indicators (KPIs) describing the recent conditions (e.g., throughput, CQI, Signal-to-Noise-Ratio (SNR), BLER, previous MCS) and any internal state maintained by the policy (e.g., an OLLA offset). An LA policy π is a mapping π : I −→ M with M = {mmin , mmin + 1, . . . , mmax } such that the MCS selected for UE u in slot n is mu,n = π(Iu,n ) ∈ M. This MCS decision determines the rate-reliability tradeoff of the transmission. For a given resource allocation, a higher MCS carries a larger TB, but also increases the probability of retransmissions. A lower MCS reduces this risk at the cost of carrying fewer information bits over the same resources. Consider the set Bu,t of new TBs whose initial transmissions are scheduled to UE u during frame t. As with downlink slots, we index TBs so that each i ∈ Bu,t belongs to a unique frame t, hence the pair (u, i) identifies a TB uniquely. For each i ∈ Bu,t , let Bu,i > 0 denote the payload size of the TB, fixed at its initial transmission and unchanged throughout its HARQ chain. Let R be the maximum number of HARQ retransmissions configured for the cell, and let Au,i ∈ {1, . . . , R + 1} denote the realized number of transmission attempts for TB i, including its initial transmission. We index these attempts by r = 0, . . . , Au,i − 1, where r = 0 denotes the initial transmission and r ≥ 1 a retransmission. After the HARQ chain terminates, let Su,i ∈ {0, 1} indicate whether the TB was (r) successfully decoded, and let Cu,i > 0 denote the number of Physical Resource Blocks (PRBs) allocated to attempt r. this rate-reliability tradeoff over time, let BT = S To capture S {(u, i) : i ∈ Bu,t } collect the TBs initiated during u∈U t<T the first T frames. We characterize the resource efficiency of an LA policy by the long-run ratio of successfully delivered information bits to the radio resources consumed across all transmission attempts: nP o Eπ (u,i)∈BT Bu,i Su,i nP J(π) = lim inf PAu,i −1 (r) o , T →∞ Eπ Cu,i r=0 (u,i)∈BT
(1)
where expectation is taken over stochastic elements (e.g., channel, traffic, decoding outcomes) under policy π. Thus, J(π) measures the long-run useful information delivered per allocated PRB. Ideally, π would track the instantaneous channel quality to continuously operate at the Pareto frontier of this trade-off. This, however, requires the ability to track changes of Iu,n
Distributed unit NOSTRAdAMUS overlay dApp π ′ Core network
Fu,j,b
Feature engineering ρ
ρ
ρ
ρ
xu,t Gradient
pu,t+1
boosting
ρ
?
Confidence Gate pu,t+1 ≥ 1 − τ au,n = −δ − pu,t+1 ≤ τ au,n = +δ + au,n = 0 otherwise
N2/N3
au,n Iu,n
CU
UE 1
Host LA policy π0 OLLA/SALAD/…
mu,n = π0 (Iu,n ) m′u,n = π0 (Iu,n ) + au,n
ACK/NACK + CQI
RU
+
UE 2 DCI + data
FH
F1
RLC
HARQ SCHEDULING
ACK/NACK + CQI
Perform transmission DCI + data
PHY
. . . UE k
Fig. 2. NOSTRAdAMUS architecture. The overlay dApp observes the host LA policy π0 , predicts first-retransmission activity for the next frame from the recent ratios ρ, and applies the confidence-gated adjustment au,n of (5) to the selected MCS.
instantaneously despite its noisy nature. Typical algorithms struggle with this, which we address in the following section. III. NOSTRA DAMUS Figure 2 shows the NOSTRAdAMUS architecture. Starting from a reference LA policy π0 : I → M, NOSTRAdAMUS constructs a new policy π ′ that adjusts the MCS decision of π0 using ML-based predictions derived from recent HARQ activity. The adjusted MCS decision is m′u,n = π ′ (π0 (Iu,n )) = π0 (Iu,n ) + au,n ∈ M.
du,j W W X X X 1 du,j × Fu,j,b . j=1
(4)
Inference runs once per UE per frame, at the end of frame t, and pu,t+1 is then held for every MCS decision in frame t + 1. Each prediction is performed independently for each UE using that UE’s own retransmission history. If fewer than H valid observations are available, or inference misses its deadline, NOSTRAdAMUS abstains (au,n = 0) and m′u,n = π0 (Iu,n ). Given the predicted probability pu,t+1 in (4), the adjustment applied to a MCS in slot n ∈ Dt+1 is
(2)
where au,n is the MCS adjustment described below. Consider a TB that successfully decodes after Au,i attempts. NOSTRAdAMUS only considers the outcome of the first transmission for learning purposes (i.e., it discards the outcomes of transmissions 2 to Au,i ), as successive attempts do not capture the channel state as accurately, since they also factor in HARQ chase combining, in which UEs use previously failed transmissions to increase their coding gain. Let Fu,j,b ∈ {0, 1} denote the HARQ outcome of the first transmission of the b-th new TB in frame j, with 0 and 1 respectively meaning ACK and NACK. For a trailing window of W ≥ 1 frames, in which the count of new downlink transmissions for UE u is du,1 , . . . , du,W the first-transmission BLER is: (W ) ρu,t =
pu,t+1 = fϕ (xu,t ) ≈ Pr(yu,t+1 = 1 | xu,t ) .
(3)
j=1 b=1
Note that windows without initial transmissions (du,j = 0 for every j in the window) are treated as missing observations. In the remainder of this paper, BLER refers to the first transmission BLER. NOSTRAdAMUS predicts whether UE u’s new TBs will likely require retransmission in the next frame. n For a frame o (1) satisfying du,t+1 > 0, the label is yu,t+1 = 1 ρu,t+1 > 0 . At the end of frame t, let xu,t ∈ RH collect the H most (W ) recent values of ρu,k , k ≤ t, and let fϕ : RH → [0, 1] denote the learned predictor (see Sec. IV). Its output is
− − −δ , pu,t+1 ≥ 1 − τ, π0 (Iu,n ) − δ ≥ mmin , au,n = +δ + , pu,t+1 ≤ τ, π0 (Iu,n ) + δ + ≤ mmax , 0, otherwise. (5) where τ ∈ 0, 21 is a confidence threshold and δ − , δ + ∈ Z>0 are configurable downward/upward MCS corrections. IV. T RAINING AND D EPLOYMENT To train NOSTRAdAMUS we use real world data collected on X5G [14] using the AutoRAN [15] automation framework. The Radio Access Network (RAN) runs on a Gigabyte E251 server equipped with an Intel Xeon 6240R Central Processing Unit (CPU) and an NVIDIA A100 GPU. We use a Foxconn RPQN configured for 2×2 MIMO operation over a 40 MHz bandwidth, with a 6DS3U TDD pattern and numerology 1. The core network is Open5GS. An OAI gNB with the NVIDIA Aerial PHY layer serves saturated DL UDP iPerf traffic to a Samsung S25 UE, carried by a user walking indoors (inside our lab). The walk exercises a wide channel condition range: Reference Signal Received Power (RSRP) spans -107 to -53 dBm between its 5th and 95th percentiles and the scheduler visits every MCS from 0 to 27. Resulting 2.4 million slot-level samples are aggregated to frame granularity, yielding 172,324 frame observations. The dataset is chronologically split into 70% training, 15% validation, and 15% testing. The held-out test set contains 25,805 samples, of which 39.3% are positive. We benchmark candidate models spanning probabilistic (Naive Bayes), distance-based (k-Nearest Neighbors), treeensemble (Random Forest, Extra Trees, Gradient Boosting),
F1
Recall
Offline
Precision
100 80 60 40 20 0
Score (%)
Score (%)
Accuracy
Accuracy
Extra Trees
Gradient Random Boosting Forest
KNN
MLP
Naive Bayes
Fig. 3. ML model benchmarking. KNN Extra Trees Random Forest Naive Bayes MLP Gradient Boosting 10 µs
1 ms
100 µs
10 ms
100 ms
Latency
Fig. 4. Per-UE inference latency of the six ML models in ONNX, 10,000 runs, amber bar marks the 99th percentile.
and neural (Multi-Layer Perceptron) families on the same feature vector. Figure 3 summarizes accuracy, precision, recall, and F1 across models. Tree ensembles dominate the benchmark, with Gradient Boosting best overall. Because NOSTRAdAMUS operates in the scheduling loop as a dApp, inference must complete in time as discussed in Sec. III. We export the trained models to ONNX to support lowoverhead real-time inference and evaluate their 99th-percentile latency over 10,000 runs (Fig. 4). Gradient Boosting achieves the lowest inference latency at 5.5 µs and, together with its predictive performance, is therefore selected for deployment. Figure 5 summarizes the selected Gradient Boosting model, which achieves 82.9% accuracy and 75.1% recall on the full test set. For all experiments, we set W = 10, H = 5, δ − = δ + = 1, and select τ = 0.1 on the validation set as a balance between confidence and intervention coverage (motivated by Fig. 5(c)). On the test set, our confidence gate intervenes on 54.1% of samples, achieving 94.2% precision. Figure 6 compares prediction accuracy for offline and OTA regimes. Overall accuracy drops by 8.9 and by 12.2 points on the high-confidence subset, while precision increases in both cases (+11.9 and +1.8 points) as recall decreases, and higher precision reduces conservative steps on clean frames. V. E XPERIMENTAL E VALUATION In this section, we evaluate whether (i) the learned overlay generalizes beyond the channel conditions seen during training; (ii) it adapts to user mobility; and (iii) we validate the complete High-Confidence
12.1% =0 87.9% 13,774 1,897
97.7% 2.3%
75.1% >0 24.9% 2,520 7,614
14.0% 86.0%
=0
>0
Predicted (a)
9,548
221
588
3,610
=0
>0
Predicted (b)
AUC (val.) = 0.901 True positive rate
Overall
True
Offline High-Confidence
1 p ≤ 0.1 p ≥ 0.5
0.5
0
p ≥ 0.9
0
0.5
1
False positive rate (c)
Fig. 5. Confusion matrices, (a) full test set and (b) high-confidence subset, (c) ROC on the validation set.
OTA
OTA High-Confidence
100 80 60 40 20 0 Precision
Recall
F1
Fig. 6. Offline vs. deployment metrics, overall and high-confidence.
closed-loop system OTA on the X5G testbed with commercial radio hardware. In all three cases, we use the same Gradient Boosting model trained on the OTA traces of Sec. IV, without retraining. Figure 7 summarizes goodput, BLER, and BLERspike (we count a spike for each frame where BLER exceeds 0.5) behavior across all scenarios. OpenAirLink: TDL-A/B/C. For controlled, repeatable single-antenna fading tests we use OpenAirLink (OAL), an open Universal Software Radio Peripheral (USRP)-based programmable channel emulator [16]. An OAI gNB drives a Foxconn RPQN RU inside a shielded enclosure, with the RF path between the RU and a Sierra Wireless EM9293 Commercial Off-the-Shelf (COTS) UE passing through OAL running on a USRP X410. OAL applies controlled attenuation and 3GPP fading profiles to the bidirectional radio link. To show NOSTRAdAMUS’s benefits even in static conditions, we start our evaluation with a single-tap flat channel. Figure 8 shows that NOSTRAdAMUS improves mean goodput by 27.1% over OLLA at a nearly unchanged BLER (+4.6%) and by 37.2% over SALAD while decreasing the BLER by 71.8% (Fig. 7). This means that even if the MCS that maximizes goodput should be fixed, iterative processes like OLLA and SALAD cannot converge to it stably. To understand if these gains are generalizable, we deploy NOSTRAdAMUS, without retraining, on 3GPP TDL-A, TDLB, and TDL-C channel models [17] at a fixed delay spread of 30 ns. Although NOSTRAdAMUS operates in completely unseen channel conditions with different delay profiles, Fig. 9 shows mean goodput improvements in all three scenarios: by 5.5–21.3% over OLLA and 5.1–20.4% over SALAD. Thus, NOSTRAdAMUS generalizes and improves both hosts on three unseen delay profiles, with no per-channel tuning. We continue our cross-channel generalization study and evaluate what happens when the channel degrades the feedback the base policy depends on (e.g., CQI) by fixing TDL-A and sweeping the delay spread from 30 to 300 ns (Fig. 10). A longer spread narrows the coherence bandwidth, so a single wideband CQI becomes less representative of the increasingly uneven per-PRB channel. Both base policies collapse: mean goodput falls by 78.7% for OLLA and 65.8% for SALAD between DS 30 and DS 300. NOSTRAdAMUS does not remove this loss, but its goodput gains persist, from 5.5% to 53.7% over OLLA and from 5.1% to 21.6% over SALAD. As we can observe in Fig. 10, the base policies concentrate more on lower MCS values as the delay spread grows, while the horizontal gap between each base and its overlay does not shrink. Thus, predictions of NOSTRAdAMUS remain useful when frequency selectivity degrades the feedback LA depends on.
120 60 0 −60 80
60 40 20 0
−80
60 40 20 0
0 −30 −60
150 100 50 0
0
150 100 50 0 Flat
A 30
B 30
A A 100 300
B B 100 300
C 30
C C OTAUMa UMi 100 300
0
∆G (%)
40
120 60 0 −60 80 0 −80 0 −30 −60 Flat
A 30
B 30
A A 100 300
B B 100 300
C 30
∆ρ (%)
0
80
300 200 100 0 60 40 20 0
∆L (ms) ∆P (%)
40
BLER (%)
80
300 200 100 0 60 40 20 0
Goodput (Mbps)
(b) πSALAD vs π ′ SALAD (hatched)
Spike len. P (spike) (ms) (%)
Spike len. P (spike) (ms) (%)
BLER (%)
Goodput (Mbps)
(a) πOLLA vs π ′ OLLA (hatched)
C C OTA 100 300
1 0.8 0.6 0.4 0.2 0
πOLLA
0
25
50
π ′ OLLA
πSALAD
CDF
πSALAD
TDL-A
1
π ′ OLLA
π ′ SALAD
0 100 50 0
Flat
TDL-C
TDL-B
A 30
A 100
A 300
B B B 30 100 300 (b) πSALAD
C 30
C 100
C 300
0.5 0 0
150
300
0
150
300
0
150
300
Fig. 11. First-attempt MCS distribution over all OAL experiments, base solid and NOSTRAdAMUS hatched: (a) on the OLLA base, (b) on the SALAD base.
Goodput (Mbps)
πOLLA
Fig. 9. Goodput CDFs at a fixed 30 ns delay spread.
We also repeat the delay spread sweep for TDL-B and TDL-C. Fig. 7 shows that NOSTRAdAMUS generalizes across different channel profiles and improves goodput in all ten scenarios. Over OLLA, the median gain is 30.2% (range 5.5– 71.5%) and 17.9% (range 5.1–60.9%) over SALAD. Overall, NOSTRAdAMUS benefits the two base policies differently. With OLLA, the BLER remains unchanged (−1.2%), while the first-attempt transmissions below MCS 5 fall by 73.4% (Fig. 11(a)). Thus, the overlay prevents delayed feedback from keeping OLLA at overly conservative rates. SALAD already operates at higher MCS values, and the overlay changes its high-MCS share by only 3.1% at the median. Instead, it reduces the BLER by 12.0–71.8% in every scenario and lowers the BLER spike probability by 41.7% and its duration by 38.4% (Fig. 12). In short, NOSTRAdAMUS accelerates rate recovery
πOLLA
πSALAD
DS 30 ns
1 CDF
50
(a) πOLLA MCS (%)
Goodput (Mbps)
Fig. 8. Goodput CDF on the single-tap flat channel.
π′
MCS ≥ 20
100
75 100 125 150 175 200 225 250 275 300 325
πOLLA
mid
MCS < 5
π ′ SALAD
MCS (%)
CDF
Fig. 7. Summary across all scenarios for base policy π0 and NOSTRAdAMUS overlay π ′ , with change ∆ (right axes) and 95% bootstrap intervals.
π ′ OLLA
π ′ SALAD
DS 300 ns
DS 100 ns
0.5 0
0
150
300 0
150
300 0
150
Goodput (Mbps)
Fig. 10. Goodput CDFs for TDL-A as the delay spread grows.
300
π ′ OLLA
πSALAD
60
150
40
100
20
50
0
OLLA
SALAD
(a) P (spike) (%)
0
OLLA
π ′ SALAD
SALAD
(b) spike duration (ms)
Fig. 12. BLER spikes’ probability (a) and its mean duration (b).
for OLLA, and reduces retransmissions for SALAD. PROPSIM: CDL-C UMi/UMa. We now add mobility with a MIMO evaluation under the 3GPP 38.901 CDL-C model on the Keysight PROPSIM FS64. We use UMi at 1.5 m/s (pedestrian) and UMa at 30 km/h (vehicular), directly stressing the failure mode introduced in Sec. I, i.e., the channel varies while a reactive policy waits for delayed HARQ outcomes. In both scenarios, NOSTRAdAMUS holds a consistent advantage. In OLLA, deep fades cause MCS to drop fast with a slow recovery rate. This causes MCS 0 to be selected 15.2% and 8.2% of the time in UMa and UMi, respectively. NOSTRAdAMUS avoids MCS 0 almost entirely, shrinking the sub-MCS-5 region by 82.9% on UMa, and by 75.5% on UMi, and raising the MCS ≥ 20 share by 108.3% and 27.0% (Fig. 13). The time series in Fig. 14 better explain this gain. NOSTRAdAMUS restores the high-rate regime faster and holds it for longer continuous intervals, whereas OLLA over-corrects
Usage (%)
UMi CDL-C
UMa CDL-C 100 10 1 0.1
10 1 0.1 0
6
12
18
24
0
6
12
MCS
18
24
MCS
Fig. 13. MCS usage under CDL-C mobility. Log scaled y-axis.
Goodput (Mbps) BLER
0.5 0.4 0.3 0.2 0.1 0
MCS
into prolonged low-MCS episodes. In short, our CDL-C experiments show that under mobility, NOSTRAdAMUS contributes by reducing the recovery time, and its value grows with the rate at which the channel forces recoveries. 50 40 30 20 10 0
25 20 15 10 5 0
20
40
20
40
20
40 Time (s)
50 40 30 20 10 0 60 0.7 0 0.6 0.5 0.4 0.3 0.2 0.1 0 60 0 25 20 15 10 5 0 60 0
100 75 50 25 0
39
30 65
26 35
85
84
16 20
(a) MCS (%)
+3.1%
18.3
20 10
π ′ SALAD
πSALAD −27.7% 25.3
11.2 11.5
0
+45.2%
200
172
+14.4% 208
182
118
100
(b) first-TX BLER (%)
0
(c) goodput (Mbps)
Fig. 15. OTA results: (a) first-attempt MCS, (b) BLER, (c) goodput.
Future work will evaluate NOSTRAdAMUS in larger multiUE deployments, extend the framework to uplink LA, and investigate its ability to augment additional base policies, including RL-based approaches. R EFERENCES
UMi CDL-C
UMa CDL-C
π ′ OLLA
πOLLA
π ′ OLLA
πOLLA
50
100
150
200
250
50
100
150
200
250
50
100
150
200
250
Time (s)
Fig. 14. Time series under CDL-C mobility. OLLA (dashed), overlay (solid).
X5G: Over the Air. The OTA evaluation uses the same testbed and configuration as the training campaign (Sec. IV), with a bandwidth of 100 MHz. Overall, our OTA experiments reproduce the trends observed under emulation. With OLLA, the gain again comes from spending less time at lower MCS regions, lifting goodput by 45.2% with similar BLER (+3.1%). On SALAD the MCS distribution is almost identical, with goodput gains (+14.4%) resulting from reducing the BLER (−27.7%) (Fig. 15). VI. C ONCLUSIONS & F UTURE W ORK We proposed NOSTRAdAMUS, a novel plug-and-play predictive overlay that adds foresight to existing LA policies without replacing or redesigning them. Using only short-term HARQ history, a lightweight Gradient Boosting model predicts upcoming retransmission activity and proactively adjusts the MCS selected by the underlying policy. Across 3GPP TDL and CDL channels, SISO and MIMO configurations, pedestrian and vehicular mobility, and OTA deployment, the same predictor generalizes and improves both OLLA and SALAD. Despite their different adaptation mechanisms, NOSTRAdAMUS complements both policies by accelerating link recovery when OLLA becomes overly conservative and reduces unnecessary retransmissions over SALAD. Overall, the overlay increases average goodput by up to 71.5% and reduces retransmissions by up to 71.8%, demonstrating that predictive adaptation can be introduced as a lightweight layer on top of existing LA algorithms.
[1] K. I. Pedersen, G. Monghal et al., “Frequency domain scheduling for OFDMA with limited and noisy channel feedback,” in Proc. IEEE 66th Vehicular Technology Conference (VTC-2007 Fall), 2007, pp. 1792– 1796. [2] R. Wang, L. Zhang et al., “LOLLA: Deep reinforcement learning for closed-loop link adaptation towards a GPU-accelerated AI-RAN,” arXiv preprint arXiv:2606.23110, 2026. [3] X. Ye, Y. Yu, and L. Fu, “Deep reinforcement learning based link adaptation technique for LTE/NR systems,” IEEE Transactions on Vehicular Technology, vol. 72, no. 6, pp. 7364–7379, 2023. [4] F. Blánquez-Casado, G. Gómez et al., “eOLLA: an enhanced outer loop link adaptation for cellular networks,” EURASIP Journal on Wireless Communications and Networking, vol. 2016, no. 1, p. 20, 2016. [5] L. Zhu, C. Bockelmann et al., “NOLLA: Non-linear outer loop link adaptation for enhancing wireless link transmission,” in 2023 IEEE 34th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), 2023, pp. 1–6. [6] V. Saxena, H. M. Tullberg, and J. Jaldén, “Reinforcement learning for efficient and tuning-free link adaptation,” IEEE Transactions on Wireless Communications, vol. 21, no. 2, pp. 768–780, 2022. [7] L. Tsipi, M. Karavolos et al., “Machine learning-based methods for MCS prediction in 5G networks,” Telecommunication Systems, vol. 86, no. 4, pp. 705–728, 2024. [8] L. You, N. Zhou et al., “From simulation to reality: Practical deep reinforcement learning-based link adaptation for cellular networks,” arXiv preprint arXiv:2603.00689, 2026. [9] R. Wiesmayr, L. Maggi et al., “SALAD: Self-adaptive link adaptation,” arXiv preprint arXiv:2510.05784, 2025. [10] H. Yin, X. Guo et al., “Predicting channel quality indicators for 5G downlink scheduling in a deep learning approach,” arXiv preprint arXiv:2008.01000, 2020. [11] F. Dı́az-Ruiz, F. J. Martı́n-Vega et al., “CSI prediction frameworks for enhanced 5G link adaptation: Performance-complexity trade-offs,” arXiv preprint arXiv:2511.20160, 2025. [12] J. H. Friedman, “Greedy function approximation: A gradient boosting machine,” The Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, 2001. [13] A. Lacava, L. Bonati et al., “dApps: Enabling Real-Time AI-Based Open RAN Control,” Computer Networks, vol. 269, p. 111342, 2025. [14] D. Villa, I. Khan et al., “X5G: An open, programmable, multi-vendor, end-to-end, private 5G O-RAN testbed with NVIDIA ARC and OpenAirInterface,” IEEE Transactions on Mobile Computing, vol. 24, no. 11, pp. 11 305–11 322, November 2025. [15] S. Maxenti, R. Shirkhani et al., “AutoRAN: Automated and Zero-Touch Open RAN Systems,” 2025. [Online]. Available: https://arxiv.org/abs/2504.11233 [16] Y. Deshpande, X. Wang, and W. Kellerer, “OpenAirLink: Reproducible wireless channel emulation using software defined radios,” in Proc. IFIP Networking Conference, 2024, pp. 678–683. [17] 3GPP, “TR 38.901 5G; Study on channel model for frequencies from 0.5 to 100 GHz (3gpp tr 38.901 version 19.4.0 release 19),” 3rd Generation Partnership Project, Tech. Rep., July 2026. [Online]. Available: https://www.etsi.org