acceptodds
Under review as a conference paper at ICLR 2027

SOLA: Second-Order Link Adaptation from ACK/NACK Feedback

Abstract

Link adaptation is a central control loop in cellular networks: a transmitter must choose a spectral-efficiency target from limited radio-state information and receives only binary ACK/NACK feedback after transmission. We study a contextual threshold-feedback model in which the latent decodable capacity is a linear function of the observed context plus calibrated log-concave fading noise. Under this stationary, model-specified setting, we propose Second-Order Link Adaptation (SOLA), an Online-Newton-style method that updates the channel parameter through the ACK/NACK negative log-likelihood and then selects the expected-throughput-maximizing rate. For a fixed unknown channel parameter, known fading CDF, bounded continuous rate set, and adversarially chosen contexts, we prove expected effective-throughput regret, including boundary rate optima. Experiments compare SOLA with OLLA, GLM-UCB-style methods, and SALAD, test CDF misspecification, and evaluate discounted D-SOLA under abrupt changes and across 77 recorded 5G context trajectories with synthetic stochastic feedback. These controlled studies support the empirical value of second-order ACK/NACK learning; the theorem does not cover discounting, finite MCS, misspecified CDFs, or safety margins.

Primary area:
learning on time series and dynamical systems
Keywords:
Online convex optimization, link adaptation, wireless

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_Vh74OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies contextual link adaptation with a linear latent capacity and a known calibrated noise CDF. SOLA applies Online Newton Step to the ACK/NACK negative log-likelihood and greedily selects the throughput-maximizing rate under its current estimate. The derived bounds throughput suboptimality by squared prediction error and then reduces it to exp-concave likelihood regret, obtaining logarithmic expected regret.

Critical strengths
  1. A strong connection between the observation and the theoretical guarantee.
  2. The paper distinguishes standard deviations from confidence intervals. This makes the practical case more credible.
Critical weaknesses
  1. The technical novelty is relatively limited. The estimator is standard ONS, and the objective has the same structure as binary-feedbackr evenue maximization with a known demand model.
  2. The empirical contribution of the proposed algorithm is not isolated. The conducted experimental results support the practical scenarios, but do not establish the advantage of the analyzed second-order update over a simpler likelihood update.
Questions for authors
  1. Could you compare directly with the closest known-CDF contextual-pricing or likelihood-learning methods?

  2. In the same setup, how do a first-order likelihood update and the practical SOLA implementation compare?

  3. Which contextual and calibration information is available to each method in the finite-MCS SALAD comparison?

Reviewer_deNFOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies contextual link adaptation with binary ACK/NACK feedback. It models the latent capacity as a linear function of the context plus noise with a known CDF. SOLA selects the rate that maximizes predicted goodput and updates the model through an Online Newton Step on the feedback likelihood. The authors prove O(d log T) expected regret for the stationary, continuous-rate setting with bounded non-anticipating contexts. They also evaluate a discounted implementation under abrupt changes and on recorded 5G contexts with simulated feedback.

Critical strengths

This paper addresses a practically relevant link adaptation problem and presents a clear formulation of rate selection under censored ACK/NACK feedback. The likelihood-based analysis provides a theoretically grounded interpretation of the learning process under calibrated channel models. The authors also explicitly distinguish the stationary regret guarantee from the discounted and finite-MCS implementations evaluated experimentally, which helps clarify the scope of the theoretical results.

Critical weaknesses
  1. My main concern is the overlap with Xu and Wang (2021) [R1]. After identifying rates with prices and ACK/NACK with purchase decisions, SOLA's likelihood and expected-goodput objective match Equations (1) and (3) of [R1]. Its greedy decision and ONS update also follow Algorithm 2 of that paper. Theorem 10 of [R1] already gives O(d log T) regret with adversarial features and a known noise distribution. SOLA discusses the older RMLP result in Section 2 and Appendix E, but does not address this closer work. The bounded rate set and the treatment of upper-boundary optima in Appendix A are differences that deserve discussion. However, the paper needs to explain what they add beyond the existing algorithm and analysis. As written, I do not see a sufficiently distinct algorithmic or theoretical contribution.

[R1] J. Xu and Y.-X. Wang, "Logarithmic Regret in Feature-based Dynamic Pricing," Advances in Neural Information Processing Systems (NeurIPS), 2021.

  1. The 77-session study uses recorded 5G contexts, but constructs the latent capacity through a fitted linear model and generates ACK/NACK outcomes with matched logistic noise (Section 6 and Appendix F). The feedback therefore follows the assumed model. This experiment adds variation in the contexts, but provides limited evidence about performance under actual cellular decoding conditions. Normalization and fitting use the entire session. The authors explicitly describe this as offline environment construction, so I view it as a limitation of the evaluation's scope. Validation with measured ACK/NACK feedback or a link-level simulator that does not impose the assumed threshold model would strengthen the case for practical use.

  2. The implemented OLLA baseline uses a scalar rate controller without a comparable context-dependent inner loop, while SOLA directly exploits contextual features. The rate intervals also differ in the stationary experiments. Additionally, SALAD targets a specified BLER, whereas SOLA directly maximizes expected goodput. These differences make it difficult to attribute the observed improvements specifically to the proposed second-order learning mechanism. A comparison against a first-order likelihood-based learner using identical contextual information, rate constraints, and objectives, as well as the closely related ONSP method, is needed to establish the additional value of SOLA.

Questions for authors
  1. How does SOLA differ technically from ONSP in [R1]? In particular, which results in Section 5 are genuinely new beyond the existing logarithmic-regret analysis for contextual threshold feedback? Can the authors clarify the role and significance of the bounded-rate and boundary-optimum treatment?

  2. How much of SOLA's performance improvement can be attributed specifically to the second-order update? Would the advantage persist against a first-order likelihood-based learner and context-aware link adaptation baselines under identical rate sets, feedback assumptions, and optimization objectives? Is there any validation using measured ACK/NACK feedback?

Reviewer_nd4qOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers a transmission rate selection problem with binary feedback determined by whether the transmission rate exceeds the instantaneous decodable capacity. The objective is to maximize expected throughput. Specifically, the instantaneous decodable capacity is modeled as a linear function of the observed context and an unknown fixed parameter, plus i.i.d. random noise.

Assuming that the noise CDF is known and satisfies a uniform compact-curvature condition, together with boundedness assumptions, the authors establish that the negative log-likelihood of the binary feedback is exp-concave. This allows them to apply the Online Newton Step (ONS) method to estimate the unknown parameter and obtain a logarithmic cumulative likelihood-regret bound, which in turn controls the cumulative squared prediction error along the observed contexts. The authors further show that throughput regret can be upper-bounded by the squared prediction error, allowing them to translate the logarithmic likelihood-regret bound into a logarithmic throughput-regret bound.

Simulation results show that the proposed algorithm accumulates regret more slowly than the considered baselines.

Critical strengths
  • Strong application motivation: This work addresses a clearly motivated problem in wireless link adaptation.

  • Generality of context assumptions: This work accommodates arbitrary bounded contexts, without requiring i.i.d. contexts or a positive-definite context covariance matrix. It appears to replace assumptions on the context distribution with stronger assumptions on the likelihood model.

  • Connections across related fields: The main novelty appears to lie in connecting online convex optimization, generalized linear bandits, and contextual dynamic pricing, particularly by showing how likelihood regret can be translated into throughput regret.

Critical weaknesses
  • Unclear distinction between established and novel theoretical results: I encourage the authors to clarify which key theoretical results follow from existing literature and which are novel contributions of this work (see my questions below).

  • Limited intuition for the theoretical assumptions and regret bounds: The theoretical results are not easy to interpret. I would encourage the authors to provide more concrete explanations and examples, particularly:

    • Immediately after Assumption 1, illustrate how changes in the noise distribution and the relevant interval affect the compact-curvature parameters.
    • Immediately after Theorem 2, illustrate how changes in the compact-curvature parameters affect the regret bound.
Questions for authors
  • Novelty of the theoretical results: If I understand correctly, a key contribution of this work is connecting ideas from several related fields. It is understandable that some theoretical ingredients may already be well established in their respective fields, such as logarithmic likelihood regret in online convex optimization. My understanding is that translating likelihood regret into throughput regret is a major contribution of this work. However, it remains unclear to me whether the result that throughput regret can be upper-bounded by squared prediction error is already established in the contextual dynamic pricing literature. Could the authors clarify this distinction and identify which parts of the theoretical analysis are novel?

  • Strength of the noise-distribution assumptions: Could the authors comment on how restrictive the assumption of a known noise CDF satisfying uniform compact curvature is in practical applications? I would also appreciate their perspective on what might happen if the true noise distribution does not satisfy this condition.

Reviewer_DecAOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SOLA, an online learning method for wireless link adaptation using binary ACK/NACK feedback. It models the channel capacity as a linear function of context with known noise distribution. SOLA uses an online newton step (ONS) update to estimate the channel parameter and greedily selects the rate that maximizes expected throughput. The paper proves an (O(d\log T)) expected regret bound under stationary assumptions and evaluates SOLA and its discounted variant using synthetic simulations and 5G trace-based experiments.

Critical strengths
  • The paper is clearly written, and the problem formulation and algorithm are easy to follow. The authors also clearly explain the assumptions and limitations of their theoretical results.
  • The experiments cover several noise distributions, CDF misspecification, and comparisons with wireless baselines, including SALAD. These results provide some useful evidence for the method in controlled settings.
Critical weaknesses
  1. The methodological novelty seems quite limited. The use of negative log-likelihood and second-order online updates for learning from binary threshold feedback is already well studied. SOLA appears to combine these existing techniques with greedy rate selection, and the regret analysis seems to follow standard exp-concavity and ONS arguments. It is unclear what fundamentally new learning technique or theoretical insight is introduced. While the wireless application is meaningful, its contribution to the broader ML community seems limited. In particular, the formulation and algorithm appear closely related to Xu and Wang (2021), but this work is not cited or discussed.

Relevant prior work: Xu and Wang, Logarithmic Regret in Feature-based Dynamic Pricing, NeurIPS 2021.

  1. The experimental evidence is not sufficiently convincing. Most experiments rely on synthetic feedback generated under assumptions favorable to the proposed model. Even the 5G trace evaluation uses fitted linear thresholds and synthetic ACK/NACK outcomes rather than actual decoding feedback. Moreover, the discounted variant and safety margins used in practical experiments are not theoretically analyzed, leaving a gap between the theoretical results and the claimed practical benefits.
Questions for authors
  1. The proposed method relies on a standard likelihood-based ONS update followed by greedy rate selection. What is the main technical novelty compared with existing binary-feedback online learning methods? In particular, how does SOLA differ from Xu and Wang (2021), which seems to use a very similar formulation and learning approach but is not discussed in the paper?

  2. The (O(d\log T)) regret bound seems to follow from the exp-concavity of the likelihood and the connection between prediction error and throughput regret. Which part of this analysis is specific to the proposed method, and what new theoretical insight does it provide beyond existing ONS-based results?

  3. The method assumes a known and calibrated noise CDF, which seems quite restrictive in practical wireless environments. How sensitive is SOLA to context-dependent or time-varying noise distributions, where the model may not be well-specified?

  4. The 5G trace experiments use synthetic ACK/NACK feedback generated from a fitted linear model. How can these results demonstrate the practical effectiveness of the method under actual wireless decoding feedback? Is there any validation using measured ACK/NACK outcomes?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.