{"id":"80272de8-cde6-4593-9a75-559ace33f942","arxiv_id":"2506.08263","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Trained neural networks approximate greedy user scheduling in mmWave MIMO-OFDM hybrid beamforming, trading modest proportional-fairness loss for large run-time savings.","lead":"This paper compares greedy, sorting, random, and neural-network-based user scheduling for millimeter-wave MIMO-OFDM base stations with hybrid beamforming. It shows that a trained neural network can cut scheduling run time sharply while losing only moderately in proportional fairness, helping designers pick a scheduler based on their latency or performance needs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DNN input features average effective channels across all PRBs, discarding frequency-selective information; the claimed OFDM scheduling surrogate may be fitting to a summary statistic rather than the actual per-PRB scheduling problem.","rationale":"The reader's weakest assumption is about distribution shift: the DNN is trained and evaluated on the same synthetic channel generator, with no robustness test. I agree that is a limitation, but I see a more fundamental, in-distribution concern that the reader did not identify. The input feature construction in Section VI-A1 collapses the K per-PRB effective channels to a single complex average. The objective being imitated, however, is a sum over PRBs of per-PRB log-rates under ZF precoding. Averaging complex channel coefficients is not a sufficient statistic for this objective in general, especially in frequency-selective mmWave channels. Therefore the DNN's ceiling on imitation accuracy is limited before any distribution shift is considered. This concern is concrete: it is stated in the paper itself, it is testable, and it bears directly on whether the learning-based approach is actually learning the OFDM scheduling problem or merely fitting a channel-statistics shortcut. I do not think this invalidates the paper's empirical trade-off for the specific tested scenario; the reported PF and runtime may still hold there. But it makes the central claim conditional in a way the paper does not acknowledge. I recommend no change to the reader's CONDITIONAL verdict, since that verdict already requests additional validation; the feature-sufficiency test should be one of the conditions.","tokens_in":9918,"tokens_out":6115,"duration_ms":86508,"concrete_test":"Train the identical DNN architecture on two feature sets: the current averaged features from Section VI-A1 (size 2I^2 + I(N_TX + 1)) and a per-PRB feature set that includes all K effective channels (size 2I^2K + I(N_TX + 1)), using the same training episodes, labels, optimizer, and epochs. Compare held-out greedy-label accuracy and closed-loop PF. If the per-PRB variant yields materially higher accuracy or PF (e.g., more than 5% relative PF gain), the current feature averaging is a load-bearing information bottleneck. As a second, more direct check, generate pairs of channel realizations that share the exact same PRB-averaged effective channels but differ in per-PRB phase structure, and verify whether the greedy incremental scheduler changes its decision while the DNN output is forced to be identical.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a supervised DNN can serve as a fast surrogate for greedy user scheduling in mmWave MIMO-OFDM. That claim requires the DNN to have enough information to imitate the greedy decisions. However, Section VI-A1 builds the first two input subvectors from \"the effective channel averaged over all PRBs,\" i.e., (1/K) * sum_k u_{k,ij}, for each i,j, while the greedy labels are computed from per-PRB rates r_i(t) = sum_k B_k log2(1 + SINR_{k,i}) where each SINR_{k,i} depends on the full per-PRB effective channel matrices U_k(M(t); t). In a frequency-selective channel with 20 subpaths per cluster and K=12 PRBs, per-PRB effective channels can vary considerably; two channel realizations can have identical complex averages across PRBs yet lead to different greedy schedules. The DNN therefore cannot, even in principle, distinguish such realizations. The paper reports only the final PF and runtime, never the imitation accuracy of the DNN, and never compares against a variant with per-PRB features. If the favorable PF result is an artifact of this particular channel generator making the PRB-average sufficiently informative, the claim that the learning-based approach solves the OFDM scheduling problem would not transfer to stronger frequency selectivity, different bandwidths, or other deployments. This is not a question of distribution shift; it is an in-distribution representational gap directly tied to the frequency-selective setting the paper claims to extend.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies downlink multiuser scheduling in a mmWave MIMO-OFDM system with hybrid beamforming. It adopts a two-timescale protocol in which analog beams are assigned on a long timescale, while user scheduling and ZF digital precoding are performed on a short timescale. The scheduling objective is formulated as a weighted-sum-rate proxy for proportional fairness. The authors propose greedy incremental and decremental user-selection algorithms, a sorting heuristic, random selection, and a supervised DNN scheduler trained to imitate the greedy incremental decisions. Numerical simulations for I=20 users, NRF=8 RF chains, and K=12 PRBs at 28 GHz show that the incremental and decremental algorithms achieve higher PF than the learning-based scheduler but require 2.3x and 16.1x longer run time, respectively; the paper concludes that the choice among approaches depends on whether PF or run time is the primary criterion.","tokens_in":10221,"tokens_out":7228,"duration_ms":85637,"significance":"The paper offers a clear system model and an empirical comparison of several scheduling algorithms plus a DNN surrogate in a realistic mmWave OFDM setup. The PF decomposition and the two-timescale protocol are standard and clearly stated, and the evaluation metric is computed externally from actual scheduling outcomes, which is appropriate. If it were firmly established that the DNN can imitate the greedy policy with a modest PF loss at much lower run time, the result would be practically useful for fast scheduling in hybrid beamforming systems. However, the central empirical claim is supported only by point estimates without uncertainty quantification, the nearest frequency-selective baseline [9] is not evaluated, and the DNN input is a PRB-averaged summary that may be insufficient for the frequency-selective scheduling task. The result is plausible but not yet established.","major_comments":[{"comment":"The first two input subvectors of the DNN are built from the effective channel averaged over all PRBs, namely (1/K) * sum_k u_{k,ij}, whereas the greedy labels are computed from per-PRB rates r_i(t) = sum_k B_k log2(1 + SINR_{k,i}), where each SINR_{k,i} depends on the full per-PRB matrices U_k(M(t);t). Since a sum of logarithms across PRBs is not determined by the average effective channel, two channel realizations with identical PRB averages can lead to different greedy schedules. The paper reports neither the label accuracy of the DNN nor an ablation with per-PRB features, so the learning-based scheduler may be fitting a summary statistic rather than solving the OFDM scheduling problem. The authors should either include per-PRB features, demonstrate that the PRB average is sufficient for the adopted channel generator, or add a controlled experiment with stronger frequency selectivity.","section":"VI-A1 and Eqs. (6)-(7), (13)-(15)"},{"comment":"The DNN is trained on samples generated by the same greedy incremental algorithm against which it is evaluated. Consequently, the comparison in Fig. 3(a) measures how well the DNN approximates that greedy policy, not how close the learning-based scheduler is to the optimal PF. The paper also does not compare with the frequency-selective baseline [9], despite citing its limitations. Please add an independent reference point (exhaustive search for a reduced-size problem, an upper bound, or the algorithm of [9]) and report the DNN's selection accuracy over the test set.","section":"VI-B and VII-C"},{"comment":"All numerical experiments set NSB=1, so the analog beamformer is recomputed in every short-time block. This means the two-timescale structure presented in Section IV-A is never actually exercised in the evaluation, and the claimed benefit of separating long- and short-timescale decisions is not demonstrated. Experiments with NSB>1, or at least a sensitivity sweep over NSB, are needed to support the two-timescale protocol as a contribution.","section":"Table I and Figs. 3-4"},{"comment":"The trade-off ratios quoted in Section VII-C (incremental provides 1.76x higher PF at 2.3x runtime; decremental provides 1.47x at 16.1x) are point estimates from a single set of simulations. No error bars, confidence intervals, or multiple training seeds are reported, and there is no sensitivity to the number of PRBs, number of subpaths per cluster, user mobility, or codebook size. Without this information, the headline conclusion that the learning-based approach strikes a favorable balance is not statistically grounded.","section":"Section VII and Table I"},{"comment":"The training setup is not fully specified: the optimizer and learning rate are omitted, no regularization or initialization is described, the threshold converting continuous DNN outputs into the binary selection variables alpha_i is not given, and no random seeds or code/data are provided. Since the paper's main evidence is empirical, these omissions prevent replication and verification of the reported run-time and PF figures.","section":"Section VI-B and Table I"}],"minor_comments":[{"comment":"The phrase 'weighted sum-rate user' should read 'user that offers the highest weighted sum-rate'; the decremental procedure in Section V-A2 is described only in prose and would benefit from being presented as a numbered algorithm for symmetry.","section":"Algorithm 1, line 3"},{"comment":"The dimension should be typeset as C^{M(t) x M(t)}; the current 'CM (t) x M (t)' is ambiguous.","section":"Section II, Eq. (2)"},{"comment":"The PRB-average should be written (1/K) * sum_{k=1}^K u_{k,ij}; as printed, 'sum_{k=1}^K u_{k,ij}/K' is ambiguous. The normalization applied to the four subvectors is also not specified.","section":"Section VI-A1"},{"comment":"The rounding rule for obtaining alpha_i from the DNN output is not defined; please specify the threshold (e.g., 0.5) and how ties are broken.","section":"Section VI-A2"},{"comment":"The sorting algorithm does not state how many users are selected; clarify whether it always selects Imax users.","section":"Section V-B"},{"comment":"Since Table I fixes NSB=1, the text in Section IV-A that describes NSB-1 repeated short-time blocks should explicitly note that the simulations do not exercise this long-timescale reuse.","section":"Section VII-A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope for a communications journal and the topic is timely. The main substantive risk is the PRB-averaged feature representation, which may prevent the DNN from imitating the frequency-selective greedy labels; this should be addressed before acceptance. The paper is positioned against [9] but never evaluates it, which reviewers will likely notice. I see no signs of misconduct; the issues are about completeness of the empirical validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper is a competent extension of the authors' prior flat-fading scheduler to frequency-selective MIMO-OFDM with hybrid beamforming, and it gives system designers a concrete menu of schedulers with a clear performance/runtime trade-off. Second, the central claim that the DNN is a fast surrogate for greedy scheduling in OFDM is weaker than it looks, because the DNN's input features discard all per-PRB information.\n\nThe system model and ZF precoding are standard and correctly stated. The two-timescale protocol is sensible, and the greedy incremental/decremental, sorting, and random baselines are fairly implemented. The numerical study is reasonably extensive, and Fig. 3(c)'s trade-off plot is a clean way to present the results. The concrete numbers—incremental gives 1.76x higher PF at 2.3x runtime, decremental 1.47x at 16.1x compared to the learning-based approach—are useful if they hold up.\n\nNow the soft spots. The most serious is the PRB-averaged DNN input. As the stress-test note says, the first two input subvectors use (1/K)*sum_k u_{k,ij}, while the greedy labels are computed from per-PRB rates with per-PRB SINRs. Frequency-selective channels with 20 subpaths per cluster and K=12 PRBs can produce realizations with identical PRB averages but different greedy schedules; the DNN cannot distinguish them. The paper never reports imitation accuracy or compares against a variant with per-PRB features. This is an in-distribution representational gap, not a distribution-shift concern. The claim that the learning-based approach solves the OFDM scheduling problem is not supported by the evidence. The PF numbers may still be good for the simulated scenario, but that would be an artifact of the channel generator making the averages sufficiently informative.\n\nSecond, Table I sets NSB=1, meaning each long-time block equals one short-time block. That collapses the two-timescale protocol and means the long-timescale analog beam assignment is re-derived every slot. The paper never tries another value, so the two-timescale motivation is untested.\n\nThird, there are missing experimental details: no error bars, no optimizer or learning rate for the DNN, no EMA smoothing factor eta_i, and no comparison with the closest baselines—the authors' own prior work [8] or the OFDM baseline [9]. These are minor-to-moderate gaps, but they make the quantitative trade-off claims hard to assess.\n\nOverall, the paper is a solid engineering study with a plausible speed-fairness trade-off, but the PRB-averaging issue is a genuine flaw that needs addressing. It deserves peer review—an editor should send it out—but the reviewers should ask for per-PRB feature comparison, robustness tests across frequency selectivity, and the missing hyperparameters. I'd bring it to reading group if the goal is to discuss what makes a learned surrogate credible, but not as a finished result.","headline":"Useful engineering extension of flat-fading scheduling to OFDM, but the DNN's PRB-averaged features leave a real representational gap that the paper never checks.","tokens_in":10777,"tokens_out":2045,"would_cite":false,"duration_ms":26223,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A supervised neural network can act as a fast surrogate for greedy user scheduling in mmWave MIMO-OFDM systems with hybrid beamforming, trading a modest loss in proportional fairness for much shorter run times.","keywords":["multiuser scheduling","hybrid beamforming","MIMO-OFDM","millimeter wave","proportional fairness","supervised learning","deep neural network","greedy algorithm"],"falsifier":"Take the trained DNN and evaluate it on channel realizations generated by a different, independently calibrated mmWave channel model, or on measured channels from a testbed, and compare the proportional fairness and the selected-user set against the greedy incremental scheduler. If the fairness gap grows substantially or the selected-user agreement collapses while the run-time advantage remains, the paper's claim that the learned scheduler is a dependable fast surrogate is falsified.","tokens_in":9701,"feed_emoji":"📶","tokens_out":9396,"duration_ms":83418,"temperature":0.7,"pith_summary":"Multiuser scheduling in a millimeter-wave MIMO-OFDM downlink is hard because the number of RF chains caps how many users can be served at once, and choosing the wrong subset wastes the limited multiplexing gain. The paper's proposal is a two-timescale protocol that fixes each user's analog beam on a slow timescale and then, on a fast timescale, selects the user subset and designs a zero-forcing digital precoder. Against greedy incremental, greedy decremental, and sorting schedulers, the authors train a supervised deep neural network to imitate the greedy incremental scheduler's decisions. Their numerical results show the greedy incremental (decremental) scheduler achieves 1.76 (1.47) times higher proportional fairness at the cost of 2.3 (16.1) times longer run time than the learning-based one. The paper concludes that the right scheduler depends on whether the operator prioritizes fairness or speed, with the learning-based approach offering a favorable balance.","feed_headline":"Neural scheduler cuts scheduling time up to 16x in mmWave MIMO","feed_subtitle":"In tests it cuts run time 2.3-16x with only modest fairness loss.","key_machinery":"The argument rests on two coupled mechanisms. First, the two-timescale protocol exploits the mmWave property that path angles change more slowly than path gains: analog beams are assigned from a codebook on a long timescale, and on each short timescale the base station forms an effective channel matrix, applies zero-forcing digital precoding, and schedules users. Second, the learning component is a supervised DNN whose normalized input features are the average effective-channel amplitudes and angles across PRBs, the angles of the assigned analog beams, and the user weights; its outputs are per-user selection decisions, and a cascade filter keeps the selected set within the RF-chain limit. The DNN is trained on labels from the greedy incremental scheduler, so the greedy algorithm acts as the teacher that defines what the network is trying to learn.","core_discovery":"Under the paper's synthetic mmWave channel model, supervised learning can reproduce the decisions of greedy incremental user scheduling closely enough to be practically useful. A fully connected DNN with three hidden layers, trained with cross-entropy loss on labels produced by the greedy incremental algorithm, predicts which users to serve; a selection filter enforces the RF-chain limit. In the evaluated scenario with 20 users and 8 RF chains, the incremental, decremental, and learning-based schedulers select 18.8%, 29.49%, and 16.62% of users on average, respectively, showing that all three avoid serving everyone to keep interference low. The paper reports that the incremental (decremental) method provides 1.76 (1.47) times higher PF at the cost of 2.3 (16.1) times longer run time compared to the learning-based method, and interprets this as evidence that learning-based scheduling is a fast surrogate for greedy scheduling in hybrid beamforming MIMO-OFDM systems with modest fairness loss.","pith_inferences":["The paper does not test transfer, but because the DNN's input features are generic effective-channel quantities, the same supervised pipeline could in principle be retrained for other mmWave deployments, codebooks, or antenna geometries without architectural changes.","A natural extension the authors do not run is to train the network with a loss that directly maximizes proportional fairness instead of imitating greedy labels; this could close the fairness gap while preserving the speed advantage.","The paper assumes one user per resource grid; extending to OFDMA, where multiple users share a PRB, would require a different output structure and is explicitly left for future work."],"forward_implications":["In a deployment where scheduling decisions must be made within a slot, the learning-based approach keeps run time nearly constant as the maximum number of served users grows, unlike the greedy methods whose search cost increases with Imax.","Operators with a strict fairness target should use greedy incremental scheduling, while operators with a tight computation budget should use the learning-based scheduler; the decremental method's 16.1 times longer run time gives it little practical advantage in the tested regime.","Because the greedy and learning-based schedulers deliberately serve far fewer than the maximum number of users (around 17–29% on average), interference-aware subset selection matters as much as the RF-chain limit in these systems.","The learning-based scheduler's performance ceiling is set by the greedy incremental teacher that produced its labels, so it cannot be expected to exceed the fairness of the incremental algorithm in this setup."],"supporting_citations":[{"why":"Provides the prior learning-based user selection scheme in mmWave hybrid beamforming that this paper extends from single-carrier to MIMO-OFDM.","marker":"[8]"},{"why":"Supplies the mmWave channel model (Eq. (9)) and large-scale parameters used to generate all training and evaluation samples.","marker":"[13]"},{"why":"Gives the primal-dual decomposition that turns long-term proportional-fairness maximization into per-slot weighted sum-rate maximization.","marker":"[12]"},{"why":"Establishes the hybrid beamforming architecture, the RF-chain constraint on simultaneous users, and the quantized analog codebook.","marker":"[2]"},{"why":"Justifies the use of zero-forcing digital precoding as near-optimal in mmWave multiuser broadcast scheduling.","marker":"[16]"},{"why":"Provides the closed-form user-search counts for incremental and decremental greedy selection used in the complexity comparison.","marker":"[17]"}],"fun_headline_variants":["AI scheduler matches greedy performance at 16x speed in mmWave MIMO","Deep learning schedules MIMO users 16x faster with minuscule fairness loss","Neural net mimics greedy scheduling for 16x faster MIMO-OFDM","Learning-based scheduling speeds up mmWave MIMO by up to 16x","DNN surrogate for greedy scheduling cuts run time 16x in MIMO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned scheduler is trained and evaluated on samples drawn from the same synthetic millimeter-wave channel generator, so the claimed speed-fairness trade-off is only shown for that channel distribution; if real deployment channels differ in user distribution, mobility, or scattering, the DNN's decisions may no longer track the greedy algorithm.","fun_headline_variants_meta":{"raw":{"variants":["AI scheduler matches greedy performance at 16x speed in mmWave MIMO","Deep learning schedules MIMO users 16x faster with minuscule fairness loss","Neural net mimics greedy scheduling for 16x faster MIMO-OFDM","Learning-based scheduling speeds up mmWave MIMO by up to 16x","DNN surrogate for greedy scheduling cuts run time 16x in MIMO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2691,"prompt_tokens":970,"completion_tokens":1721,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":1619}},"tokens_in":586,"tokens_out":1721,"duration_ms":14204,"temperature":1.0,"reasoning_tokens":1619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:15:00.622657+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained DNN and evaluate it on channel realizations generated by a different, independently calibrated mmWave channel model, or on measured channels from a testbed, and compare the proportional fairness and the selected-user set against the greedy incremental scheduler. If the fairness gap grows substantially or the selected-user agreement collapses while the run-time advantage remains, the paper's claim that the learned scheduler is a dependable fast surrogate is falsified.","supporting_citations":[{"cited_title":"Learning-based adaptive user selection in millimeter wave hybrid beamforming systems,","cited_arxiv_id":null,"evidence_quote":"Provides the prior learning-based user selection scheme in mmWave hybrid beamforming that this paper extends from single-carrier to MIMO-OFDM."},{"cited_title":"Millimeter wave channel modeling and cellular capacity evaluation,","cited_arxiv_id":null,"evidence_quote":"Supplies the mmWave channel model (Eq. (9)) and large-scale parameters used to generate all training and evaluation samples."},{"cited_title":"Maximizing queueing network utility subject to stability: Greedy primal-dual algorithm,","cited_arxiv_id":null,"evidence_quote":"Gives the primal-dual decomposition that turns long-term proportional-fairness maximization into per-slot weighted sum-rate maximization."},{"cited_title":"Limited feedback hybrid precoding for multi-user millimeter wave systems,","cited_arxiv_id":null,"evidence_quote":"Establishes the hybrid beamforming architecture, the RF-chain constraint on simultaneous users, and the quantized analog codebook."},{"cited_title":"On the optimality of multiantenna broadcast scheduling using zero-forcing beamforming,","cited_arxiv_id":null,"evidence_quote":"Justifies the use of zero-forcing digital precoding as near-optimal in mmWave multiuser broadcast scheduling."},{"cited_title":"MMSE-based user selection algorithms for multiuser linear precoding,","cited_arxiv_id":null,"evidence_quote":"Provides the closed-form user-search counts for incremental and decremental greedy selection used in the complexity comparison."}],"review_version":1}