ML-OLLA: Link Adaptation for Real-Time Traffic in 5G/6G NTN Systems

I. INTRODUCTION

• NTN offers a way to extend cellular coverage beyond the urban and suburban areas commonly served by terrestrial networks.Release 17 standardized NTN using satellites or high-altitude platforms; satellite constellations can provide broadband access globally.Source: §I · p. 1
• Real-time services, including voice and video, AR/VR/XR, gaming, and mission-critical applications, typically require end-to-end packet latency on the order of tens of milliseconds.Such latency requirements are attainable only with LEO satellites or high-altitude platforms, according to the passage.Source: §I · p. 1
• Link adaptation selects the MCS for each transmission attempt and determines how many attempts fit the application's delay budget.In NTN, long propagation delays consume budget, mobility drives rapid channel variation, and delayed CQI and ACK/NACK feedback complicate adaptation.Source: §I · pp. 1–2
• ML-OLLA increases system capacity and, at 99.9% reliability, reduces the minimum supported latency budget from 60 ms to 40 ms versus the baselines.Its data-driven parameter-selection method learns per-attempt target error rates from observed MCS indexes.Source: §I · p. 2

II. RELATED WORKS

• Prior NTN studies address throughput-oriented traffic or retransmission performance, but not link adaptation under application-specific QoS constraints.The identified gap is especially relevant when delayed CQI and HARQ feedback leave few feasible attempts within a packet's latency budget.Source: §II · p. 2
• Terrestrial adaptive-BLER and learning-based methods optimize terrestrial throughput or long-term BLER, not per-attempt reliability targets for satellite links with long delays.The passage also notes that inference or iterative training may be infeasible for satellite gNBs with limited onboard processing resources.Source: §II · p. 2
• Enhanced, non-linear, and exponential-decay OLLA improve convergence, stability, or feedback use, but do not select target BLERs separately for initial transmissions and retransmissions.Low-BLER analysis finds that stringent targets make NACK feedback sparse and less informative.Source: §II · p. 3
• Unequal target BLERs can reduce resource consumption because the first transmission is always sent, whereas retransmission resources are used only after failure.Prior multicast URLLC work – is limited to two attempts and fixed targets; this work extends the approach to several retransmissions with adaptive targets.Source: §II · p. 3

III. SCENARIO AND PROBLEM STATEMENT

• The modeled LEO satellite serves mobile UEs in a single spot beam, with the gNB placed either on the ground or on the satellite.For satellite-based gNB placement, the feeder link is assumed not to constrain system performance.Source: §III · p. 3
• The study considers downlink flows because the gNB controls scheduling and MCS selection; uplink grant acquisition and UE buffer constraints are left for future work.Each flow must meet a packet-delivery deadline and a limit on packets lost or delivered late.Source: §III · p. 3
• System capacity Nmax(α) is the maximum number of simultaneously served UEs for which at least an α fraction meet both QoS constraints.Those constraints are the latency budget DQoS and packet-loss-rate target PLRQoS.Source: §III · p. 3
• The design goal is to minimize each UE's channel resource consumption while satisfying the QoS constraints, thereby supporting more real-time flows on shared resources.The beam's UEs share the same time-frequency resources.Source: §III · p. 4

IV. PROPOSED LA ALGORITHM

• The gNB estimates the maximum number of attempts as N = ⌊DQoS/THARQ⌋ and selects an MCS for each attempt, with the available attempts typically few under LEO real-time delay budgets.For the considered ground-gNB scenario, a 30 ms budget allows two attempts.Source: §IV · p. 4
• The algorithm combines ILLA, which selects MCSs from recent channel measurements, with ML-OLLA, which corrects MCSs using per-attempt ACK/NACK feedback.A third component learns per-MCS error probabilities from a sliding window of observed MCS indexes to determine target BLERs and step sizes.Source: §IV · p. 4

A. Inner-Loop Link Adaptation

• ILLA uses each UE’s latest subband CQI report, whose integer values represent quantized SINR for individual frequency bands.Reports arrive with period TCQI.Source: A. Inner-Loop Link Adaptation · p. 4
• For allocated resource blocks, the gNB maps CQIs to MCS indexes using the 3GPP TS 38.214 tables and closest spectral efficiencies.Source: A. Inner-Loop Link Adaptation · p. 4
• ILLA averages the mapped MCS indexes over the allocated resource blocks and rounds the result to the nearest integer index.Source: A. Inner-Loop Link Adaptation · p. 4

B. ML-OLLA

• ML-OLLA assigns different MCSs and reliability targets to transmission attempts, with retransmission target BLERs no higher than the initial target.The targets satisfy p1 ≥ p2 ≥ ··· ≥ pN, with their product constrained by the QoS packet-loss target.Source: B. ML-OLLA · p. 4
• Unlike conventional OLLA, which selects the initial MCS and reuses it for retransmissions, ML-OLLA runs a separate OLLA process for each attempt.Source: B. ML-OLLA · p. 4
• Each attempt’s MCS applies its learned correction to the ILLA selection and is limited to the valid interval from 0 to MCSmax.Source: B. ML-OLLA · p. 4
• Each process updates its correction using ACK/NACK feedback for that attempt, maintaining its own target BLER despite ILLA inaccuracies from delay and CQI quantization.Source: B. ML-OLLA · p. 5
• The ML-OLLA processes inherit the baseline OLLA convergence properties established in.Source: B. ML-OLLA · p. 5

C. Parameter Selection

• For each MCS, the method estimates error probability from a sliding window of recent CQI reports, using the cumulative frequency of supported MCSs below it.Each CQI report contributes a virtual resource-block allocation and an ILLA-selected MCS sample.Source: C. Parameter Selection · p. 5
• The optimization selects MCS indices for transmission attempts to minimize average resource consumption while satisfying the reliability constraint.An attempt consumes resources only if all preceding attempts fail.Source: C. Parameter Selection · p. 5
• The model assumes independent decoding errors across attempts and omits HARQ-CC gain, conservatively overestimating packet-loss probability.Accounting for HARQ-CC is left for future work.Source: C. Parameter Selection · p. 5
• Exhaustive search considers finite MCS combinations with retransmission MCSs no higher than earlier ones, while per-UE storage remains limited to the window, empirical CDF, and OLLA state.Source: C. Parameter Selection · p. 5

A. Simulation Setup

• The ns-3 evaluation uses 5G-LENA, NTN propagation delays, and the 3GPP NTN channel model; Table II lists the main simulation parameters.Source: A. Simulation Setup · p. 6, Table II · p. 6
• The scheduler uses only CQI reports and ACK/NACKs received by the gNB, explicitly accounting for propagation-delay feedback.Source: A. Simulation Setup · p. 6
• The ITS traffic model sends each UE 256-byte packets every 20 ms, with the gNB placed either on the ground or on the satellite.Source: A. Simulation Setup · p. 6
• Baselines include manually targeted OLLA, SINR-adaptive OLLA using, and selection of the most robust MCS.Source: A. Simulation Setup · p. 6
• ML-OLLA uses a 2,000-sample window, sets the attempt count from the delay budget, and targets the QoS reliability level.The window trades estimation precision against adaptation speed.Source: A. Simulation Setup · p. 6
• Evaluation measures the fraction of UEs meeting QoS and capacity, defined as the maximum UE count with at least 95% meeting requirements.Results are averaged over 15 independent simulation runs.Source: A. Simulation Setup · p. 6

• At DQoS = 30 ms, ML-OLLA alone stays above the 95% capacity threshold across a wide load range, while OLLA with Ptarget = PLRQoS satisfies less than 80% of UEs.For some UEs, SINR degradation between attempts exceeds 3 dB with probability higher than PLRQoS, while HARQ-CC compensates only about 3 dB; ML-OLLA uses more conservative retransmission target BLERs to tolerate such drops.Source: B. Analysis of Results · pp. 6–7, Figure 2 · p. 7
• Setting OLLA’s Ptarget = 5 · 10−3 increases the fraction of satisfied UEs but consumes more channel resources and degrades more sharply at high loads.Source: B. Analysis of Results · p. 7
• Up to fourfold capacity gain: ML-OLLA supports about 80 UEs in [40, 50] ms and offers notable capacity down to about 25 ms, versus about 20 UEs for the cited baseline in [40, 50] ms.Source: Figure 4 · p. 7, B. Analysis of Results · p. 7
• At PLRQoS = 10−3, ML-OLLA provides capacity down to roughly 40 ms versus above 60 ms for baselines, with nearly a 30% capacity gain at DQoS = 70 ms.Source: Figure 5 · p. 7, B. Analysis of Results · pp. 7–8
• With a satellite-based gNB, minimum supported delay falls from about 25 to 20 ms for PLRQoS = 10−2 and from about 40 to 30 ms for PLRQoS = 10−3, with a 5–10% capacity gain.HARQ retransmissions are twice as fast, but feeder-link transmission delay means the first attempt takes about the same time in either deployment.Source: Figure 6 · p. 8, B. Analysis of Results · p. 8

VI. CONCLUSION

• ML-OLLA gives each transmission attempt a dedicated target BLER and independent OLLA process, enabling more aggressive MCS selection early and more robust MCSs for retransmissions.Its data-driven parameter-selection method derives per-attempt targets from the empirical distribution of recent CQI reports.Source: §VI · p. 8
• Up to four times higher system capacity than the baselines, with minimum supported latency reduced from 40 to 25 ms at PLRQoS = 10^-2 and from 60 to 40 ms at PLRQoS = 10^-3.Placing the gNB on the satellite further reduces minimum supported delay and adds 5–10% capacity in the evaluated scenario.Source: §VI · p. 8
• The parameter-selection model will be extended to account explicitly for HARQ-CC gain and for OLLA-caused MCS variations and their effect on channel resource consumption.Source: §VI · p. 8