{"id":"f2a75f58-c2d9-4eef-8fff-6bb0329be8db","arxiv_id":"2506.09268","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A multi-armed bandit controller that tunes bandwidth allocation, user association, and macro-cell shutdown improves throughput and reduces energy in a simulated satellite-augmented network.","lead":"This paper proposes an online optimization framework for integrated terrestrial and non-terrestrial networks, using a multi-armed bandit algorithm to pick bandwidth splits, user association, and base station shutdown settings in real time. The authors report up to 19% throughput gains and 5% energy savings in low-traffic periods compared with standard 3GPP configurations, based on a 24-hour system-level simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed online gains rest on an evaluation that appears to train on all 168k snapshots and then sample one action per hour, so the reported 19%/5% improvements may reflect offline batch training with test data rather than online optimisation.","rationale":"The reader's weakest assumption correctly identifies the evaluation protocol as the load-bearing point. My analysis sharpens this concern: the paper's own wording in Section V indicates a policy 'learned' from all 168,000 snapshots and then used to sample one action per hour, which is not the online loop of Algorithm 1. With n = 875 arms and T = 24 hourly decisions, the claimed online learning cannot be validated by the reported experiment unless each snapshot is treated as a time step, in which case training and evaluation overlap on the same hours. The central claim is that the framework adaptively optimises in real time; for that claim to hold, the evaluation must be a genuine chronological online experiment. A strict chronological rerun is the decisive check. Because the reader already assigned CONDITIONAL and my concern does not move the verdict, no change is needed; the conditional status appropriately reflects that the empirical support is incomplete until the online protocol is clarified and tested.","tokens_in":7915,"tokens_out":5772,"duration_ms":65787,"concrete_test":"Run a strict chronological online evaluation: initialise x_1 uniformly, and for each hour h = 0..23 draw action a_h ~ x_h using only snapshots from hours < h (or, at most, hour h-1), observe the cost and constraint on hour h, then apply the Algorithm 1 update. Report cumulative regret R_T (Eq. 11), cumulative violation V_T (Eq. 12), and the daily throughput/energy curves under this protocol. If these reproduce the 19%/5% figures, the concern is resolved; if the gains require training on all 168k snapshots before evaluation, the paper's central online-optimisation claim is not demonstrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V states: 'we collected 7·10^3 snapshots of the network for each hour of the day, yielding a total of 168·10^3 samples. Then, we used the learned policy to sample an action for each hour of the day to evaluate the resulting performance.' This does not describe the online loop of Algorithm 1, where x_t is updated after each bandit observation and no separate 'learned policy' exists before the horizon ends. If the policy was first trained on all snapshots, including the hours against which it is evaluated, then the unsatisfied-UE curves and the 19% throughput / 5% energy figures are contaminated by test data. The structural mismatch is stark: the action space has n = 875 arms, while the evaluation period has only T = 24 hourly decisions. A genuine online BCOMD run would receive only 24 bandit-feedback samples, which is far too few to learn among 875 arms or to demonstrate the regret and constraint-violation guarantees borrowed from unpublished reference [7]. If, instead, each of the 168k snapshots is treated as a time step, then the policy is trained over the entire 24-hour period and later evaluated on those same hours; either way, the online claim is not supported by the reported protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an online optimization framework for integrated terrestrial and non-terrestrial networks (TN-NTN), formulated as a multi-armed bandit problem and solved with the BCOMD algorithm. The action space consists of 875 configurations of four parameters: bandwidth split epsilon, load threshold tau_nu, RSRP threshold tau_RSRP, and association weight alpha. The cost function trades off sum log-throughput against terrestrial energy consumption, while the constraint penalizes unsatisfied UEs. A 24-hour system-level simulation is described, comparing the proposed framework against 3GPP-TN and 3GPP-NTN baselines, and reporting lower unsatisfied-UE fractions, up to 19% throughput gains, and 5% energy savings in low-traffic periods. The manuscript does not provide code, data, or confidence intervals, and the evaluation protocol is ambiguous regarding whether the algorithm is run online or trained on snapshots that include the evaluation hours.","tokens_in":8143,"tokens_out":7379,"duration_ms":74618,"significance":"If substantiated, the framework is relevant to green TN-NTN management because it couples load-aware UE association, bandwidth split, and MBS shutdown in a single online decision process, going beyond fixed max-RSRP baselines. The problem is well motivated and the system model follows 3GPP channel models. However, the central empirical claims are currently not supported by the reported evaluation: the protocol does not clearly implement an online loop, the main terrestrial baseline is handicapped by using only 10 MHz of the 40 MHz total bandwidth, and the optimality narrative relies on an unpublished algorithm reference. The paper also lacks statistical uncertainty quantification and reproducible implementation details. These issues are load-bearing for the paper's central claim, so the manuscript requires substantial revision before the empirical conclusions can be accepted.","major_comments":[{"comment":"The evaluation protocol described in Section V ('we collected 7·10^3 snapshots ... Then, we used the learned policy to sample an action for each hour of the day') does not implement the online loop of Algorithm 1, in which x_t is updated after each bandit observation and no separate learned policy exists before the horizon ends. If the policy was trained on all 168,000 snapshots, including the hours used for evaluation, the reported unsatisfied-UE curves and the 19% throughput / 5% energy figures are contaminated by test data. If instead the 24 hourly evaluations are the only online steps, then BCOMD receives only T=24 bandit feedback samples for an action space of n=875 arms, which is far too few to support the regret and constraint-violation guarantees claimed via reference [7]. The authors should run a genuine online experiment with time-separated train/test snapshots, or report per-time-step regret and cumulative constraint violation over a horizon long enough to exercise the algorithm.","section":"Section V, first paragraph; Algorithm 1"},{"comment":"The 3GPP-TN benchmark allocates only 10 MHz to the terrestrial tier even though no satellite tier is present, while Table I states total bandwidth W=40 MHz. A terrestrial-only deployment would reasonably use the full 40 MHz, so the TN baseline is handicapped in both throughput and the number of unsatisfied UEs. The reported average throughput gains of 19% (low-traffic) and 1% (high-traffic) over 3GPP-TN are therefore not attributable to the proposed framework alone. The comparison should be rerun with a 40 MHz terrestrial-only baseline, or the authors should cite a 3GPP specification that fixes the TN tier at 10 MHz even without a satellite tier.","section":"Section V-B, benchmarks; Table I"},{"comment":"The optimality narrative for the framework relies on BCOMD's sublinear regret and constraint-violation guarantees, which are stated to come from the unpublished reference [7]. Because no theorem or proof is reproduced, and because the setting here is non-stationary with a large action space, a reader cannot verify that those guarantees apply to this problem. In addition, the BCOMD hyperparameters eta, mu, gamma, and Omega, as well as the exact functional form and values of zeta used in Eq. (7), are not reported. Without these, the simulation results cannot be reproduced. Please include the relevant conditions from [7] in an appendix and report all hyperparameters.","section":"Section IV-B, Algorithm 1, and reference [7]"},{"comment":"The constraint is defined in Eq. (8) as the fraction of UEs with R_i < rho_i, but the text immediately after says 'if a UE perceives a RSRP lower than a set threshold RSRPmin, it is considered unsatisfied.' These two criteria are not equivalent, and the figures in Section V-A are described in terms of unsatisfied UEs without stating which definition is plotted. The discrepancy affects the meaning of the constraint violation and should be clarified.","section":"Section III, Eq. (8), and Section V-A"},{"comment":"All quantitative claims are based on a single 24-hour simulation run, with no confidence intervals, random seeds, or statistical tests. Since the reported gains of 19% and 5% are differences between single trajectories, it is impossible to assess whether they are significant or within simulation noise. Multiple independent runs with uncertainty quantification are needed to support the empirical conclusions.","section":"Section V, Figs. 1-3"}],"minor_comments":[{"comment":"The indicator function is typeset as `/BD{p_j>0}`; it should be a standard indicator and the energy units should be stated explicitly.","section":"Eq. (6)"},{"comment":"The oracle policy in (9) is defined over x_t although f_t(a_t,x_t) depends on the realized action a_t; the expectation over a_t should be made explicit, as in (11).","section":"Eqs. (9)-(11)"},{"comment":"The symbol f_t is used both for the cost vector and for the scalar cost f_t(a_t,x_t), which makes the estimator in (16) harder to follow; please distinguish vector and scalar.","section":"Eqs. (14)-(16)"},{"comment":"The caption says 'satisfied UE proportion' while the y-axis label and text refer to 'unsatisfied UE'; please align the terminology.","section":"Fig. 1"},{"comment":"The demand parameter lambda_U, the hourly UE counts, and the BCOMD hyperparameters (eta, mu, gamma, Omega, zeta) are not listed, so the simulation cannot be reproduced from the paper alone.","section":"Table I"},{"comment":"Reference [7] is marked 'Under Submission'; either provide a citable version or reproduce the required regret and constraint-violation guarantees in an appendix.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the issues in the evaluation are central and should gate acceptance. I think the paper can be revised rather than rejected: the framework is coherent and the problem is relevant, but the authors need to provide a genuine online evaluation with a fair terrestrial baseline and either release code/data or report all hyperparameters and snapshot separation. I would not rely on the unpublished [7] as a substitute for presenting the required guarantees."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper is a reasonable engineering attempt: it casts joint control of bandwidth split, user association, and macro-cell shutdown in an integrated TN-NTN as a bandit problem and uses an online constrained mirror-descent algorithm to choose among 875 parameter configurations. Second, the evaluation as written does not actually demonstrate online optimisation. Section V says the authors collected 168,000 snapshots, \"used the learned policy to sample an action for each hour,\" which sounds like offline batch training on the same data they later evaluate on. That is a mismatch with Algorithm 1’s online loop, where no final “learned policy” exists before the horizon ends. A genuine online run would have only 24 bandit feedback samples, far too few to learn among 875 arms. The reported 19% throughput and 5% energy savings are therefore not backed by the protocol as described.\n\nWhat is genuinely new: the specific action grid and the 24-hour TN-NTN evaluation combining pricing-based association, bandwidth split, and MBS shutdown. The simulation follows standard 3GPP channel and traffic models, and the parameter space is well explained. That is a legitimate contribution to system-level management, though not a fundamental one.\n\nThe soft spots are serious. Besides the evaluation protocol, the 3GPP-TN baseline is handicapped: it allocates only 10 MHz of 40 MHz to the terrestrial network, even though no satellite uses the other 30 MHz. So part of the apparent gain is just giving the proposed scheme more spectrum. Also, no code or data are provided, no confidence intervals are reported, and the optimality language leans on an unpublished manuscript [7], which cannot be independently checked. The claimed gains are modest—1-4% average throughput gain in high-traffic hours—so the lack of statistical support matters.\n\nMy bottom line: this is a plausible framework with a flawed evaluation section. The paper deserves a serious referee because the problem is relevant and the approach is not clearly wrong, but it needs a major revision to clarify the train/test split, describe the actual online procedure, and fix the baseline allocations. I would not cite the empirical numbers until that is resolved.","headline":"A plausible bandit-based controller for green TN-NTN, but the reported gains are undermined by an evaluation that appears to train on the same snapshots it tests on and by an unfair baseline that wastes 30 MHz of spectrum.","tokens_in":8715,"tokens_out":2930,"would_cite":false,"duration_ms":29109,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a bandit controller choosing user-cell associations, bandwidth splits, and macro-cell shutdowns hour by hour can keep users satisfied while cutting terrestrial energy consumption.","keywords":["multi-armed bandits","online optimisation","non-terrestrial networks","energy efficiency","user association","load balancing","BCOMD","5G new radio"],"falsifier":"A decisive check is to split the 24-hour data chronologically: train the bandit on the first part of the day and test on a held-out day or held-out hours, then compare its realised cost and constraint violations against the same time-of-day baseline. If the reported 19% throughput gain and 5% energy saving disappear or become negative under a strict train/test split, the online-optimisation claim is not supported.","tokens_in":7675,"feed_emoji":"🛰️","tokens_out":7486,"duration_ms":69348,"temperature":0.7,"pith_summary":"This paper tries to establish that an integrated terrestrial and non-terrestrial network can be steered online by a bandit controller that, hour by hour, chooses how to split bandwidth, how to associate users, and which macro cells to switch off. The claim is that a constrained multi-armed bandit algorithm (BCOMD) can adapt to changing traffic without knowing the traffic distribution in advance, reducing the share of unsatisfied users in peak hours and saving energy when demand is low. In 24-hour system-level simulations, the proposed policy reports up to 19% higher sum throughput and about 5% lower terrestrial energy consumption in low-traffic periods compared with static 3GPP configurations. A sympathetic reader would care because energy use and load imbalance in dense cellular deployments are practical operating costs, and this is a concrete proposal for putting the satellite tier to work on both.","feed_headline":"Bandit policy lifts throughput 19% and cuts energy 5%","feed_subtitle":"Hourly choices of user-cell association and bandwidth split can switch off idle macro cells without starving users.","key_machinery":"The machinery is the BCOMD algorithm paired with a four-parameter action space. BCOMD—bandit-feedback constrained online mirror descent—maintains a distribution over arm configurations, receives the cost and constraint violation of only the sampled arm, forms unbiased gradient estimates (loss, constraint, and a bias term), and updates the distribution in the dual space via a mirror map before projecting back to the simplex. The action space is defined by the parameter vector θ = [ε, τν, τRSRP, α], where ε is the fraction of bandwidth given to the satellite tier, τν and τRSRP are shutdown thresholds on load and received signal strength, and α weighs cell load in the pricing association rule Pi(j)=RSRP_ij − αν_j. This mechanism lets the controller balance the Lagrangian cost (sum-throughput minus regularised energy) against the long-run constraint on unsatisfied UEs, with the dual variable λt updated from accumulated violations.","core_discovery":"The paper's central claim is that jointly optimising three control levers—UE association, bandwidth split between terrestrial and satellite tiers, and shutdown of terrestrial macro base stations—amounts to an online decision problem that can be solved with bandit feedback alone. Each lever setting is an arm; the action is selected from a distribution updated by BCOMD, a constrained online mirror descent algorithm that estimates cost and constraint-violation gradients from the observed cost of the single chosen arm. The authors report that the learned policy keeps the proportion of unsatisfied UEs near zero in low traffic, improves throughput over 3GPP-TN by up to 19% and over 3GPP-NTN by up to 10% in low-traffic hours, and reduces terrestrial energy consumption by about 5%, while in high-traffic periods it trades a slight energy increase for better load balancing and user satisfaction.","pith_inferences":["A likely implicit consequence is that the approach can be combined with existing 5G sleep-mode mechanisms, since MBS shutdown is already one of the control levers; the bandit would then decide which cells sleep and when.","The fixed grid of 875 arms could be replaced by continuous or hierarchical action search, but the bandit's regret bound would need re-examination because the current guarantees come from an unpublished analysis and assume constraints that may not hold in a non-stationary 24-hour traffic cycle.","Because traffic has a strong daily periodicity, a simple time-of-day look-up table trained on the same snapshots might match much of the reported gain; the distinctive test is whether the bandit stays competitive when traffic is shifted, e.g., by a special event or outage."],"forward_implications":["If the framework works as reported, a network operator can reduce terrestrial energy use by letting satellites absorb traffic from lightly loaded macro cells during off-peak hours.","The same bandit policy can improve user satisfaction in peak hours by shifting users away from overloaded cells through the load-aware pricing association rule.","Because the controller needs only the realised cost and constraint violation of the action it took, it can be applied without an explicit traffic model or full network state.","The reported 19% throughput and 5% energy gains are upper-end figures for low-traffic periods; during high-traffic periods the benefit shifts from energy to load balancing and satisfaction.","The framework's performance depends on the chosen grid of 875 arm configurations; finer or differently chosen grids would change the achievable trade-off."],"supporting_citations":[{"why":"Supplies the BCOMD algorithm and the bandit-feedback constraint-satisfaction framework the whole policy rests on.","marker":"[7]"},{"why":"Defines the traffic-aware TN-NTN optimisation model and objective that this paper extends to online settings.","marker":"[6]"},{"why":"Provides the pricing-based user-association rule that the framework adapts with load weight α.","marker":"[4]"},{"why":"Gives the terrestrial channel model, path-loss tables, and line-of-sight probability used in simulations.","marker":"[8]"},{"why":"Gives the satellite channel model including clutter, scintillation, and building-entry losses.","marker":"[9]"},{"why":"Provides the three-component MBS energy consumption model used to compute terrestrial energy.","marker":"[10]"},{"why":"Supplies satellite altitude, transmit power, antenna gain, and the 3GPP-NTN benchmark configuration.","marker":"[11]"}],"fun_headline_variants":["Bandit algorithm optimizes TN-NTN for 19% throughput gain, 5% energy saving","MAB framework finds 19% throughput boost and 5% energy cut in green networks","Smart bandit policy shuts off idle cells, saves 5% energy, boosts throughput 19%","Online bandit control lifts throughput 19% and trims energy 5% in hybrid nets","Bandit feedback enables green TN-NTN with 19% throughput gain, 5% saving"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a policy trained and selected on 168,000 pre-collected network snapshots and then used to sample one action per hour faithfully represents online optimisation, and that BCOMD's constraint-satisfaction guarantees, taken from an unpublished analysis, carry over to this non-stationary 24-hour deployment.","fun_headline_variants_meta":{"raw":{"variants":["Bandit algorithm optimizes TN-NTN for 19% throughput gain, 5% energy saving","MAB framework finds 19% throughput boost and 5% energy cut in green networks","Smart bandit policy shuts off idle cells, saves 5% energy, boosts throughput 19%","Online bandit control lifts throughput 19% and trims energy 5% in hybrid nets","Bandit feedback enables green TN-NTN with 19% throughput gain, 5% saving"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000835,"raw_usage":{"total_tokens":3644,"prompt_tokens":949,"completion_tokens":2695,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2572}},"tokens_in":565,"tokens_out":2695,"duration_ms":17849,"temperature":1.0,"reasoning_tokens":2572,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:52:45.148918+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is to split the 24-hour data chronologically: train the bandit on the first part of the day and test on a held-out day or held-out hours, then compare its realised cost and constraint violations against the same time-of-day baseline. If the reported 19% throughput gain and 5% energy saving disappear or become negative under a strict train/test split, the online-optimisation claim is not supported.","supporting_citations":[{"cited_title":"Adversarial multi-armed bandits with constraints in d ynamic environ- ments,","cited_arxiv_id":null,"evidence_quote":"Supplies the BCOMD algorithm and the bandit-feedback constraint-satisfaction framework the whole policy rests on."},{"cited_title":"Optimizing Integrated Terrestrial and Non-Terrestrial Networks Performance with Traffic-Aware Resource Management","cited_arxiv_id":"2410.06700","evidence_quote":"Defines the traffic-aware TN-NTN optimisation model and objective that this paper extends to online settings."},{"cited_title":"Distributed pricing-based user association for downlin k heterogeneous cellular networks,","cited_arxiv_id":null,"evidence_quote":"Provides the pricing-based user-association rule that the framework adapts with load weight α."},{"cited_title":"TR 38.901, Study on channel model for frequ encies from 0.5 to 100 GHz ,","cited_arxiv_id":null,"evidence_quote":"Gives the terrestrial channel model, path-loss tables, and line-of-sight probability used in simulations."},{"cited_title":"TR 38.811, Study on New Radio (NR) to support non-ter restrial networks,","cited_arxiv_id":null,"evidence_quote":"Gives the satellite channel model including clutter, scintillation, and building-entry losses."},{"cited_title":"Machine learning and analytical power consumption models for 5g base stations,","cited_arxiv_id":null,"evidence_quote":"Provides the three-component MBS energy consumption model used to compute terrestrial energy."},{"cited_title":"TR 38.821, Solutions for NR to support non - terrestrial networks (NTN),","cited_arxiv_id":null,"evidence_quote":"Supplies satellite altitude, transmit power, antenna gain, and the 3GPP-NTN benchmark configuration."}],"review_version":1}