{"id":"cb10aeee-c63d-4344-8e08-a0d3516d88ec","arxiv_id":"2505.07278","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A three-level hierarchical multi-armed bandit scheduler for coordinated spatial reuse in IEEE 802.11bn Wi-Fi raises aggregate throughput by about 80% over legacy random access in simulations, with a smaller real testbed validation.","lead":"This paper designs a scheduler that lets multiple Wi-Fi access points transmit at the same time on the same channel without ruining each other's signals. It pairs a mathematical model that finds the best possible schedule with a lightweight learning algorithm, multi-armed bandits, that approaches that ideal in practice.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"H-MAB's independent power selection is the soft spot; joint-action coupling is untested and the 80% claim may not generalize.","rationale":"Read in good faith: the paper's contribution is a probing H-MAB scheduler plus an upper-bound MILP; the empirical evaluation is substantial and the code is open. The MILP column-generation scheme appears standard. The load-bearing issue is the H-MAB's credit assignment for transmit power. Footnote 4 explicitly disclaims precision for this level, and the simulation/testbed scenarios do not isolate the regime where power choices are strongly coupled. The reader identified the same assumption, and I agree. I would not change the verdict: CONDITIONAL remains appropriate because the concern is about generality, not a demonstrated internal inconsistency. A targeted experiment with a joint-action baseline would settle whether the independent-learners approximation is benign or the reason the reported gains may not transfer to denser or more coupled deployments.","tokens_in":24545,"tokens_out":23529,"duration_ms":247485,"concrete_test":"Use a purpose-built two-AP/two-station simulator (Fig. 1-like) with path losses chosen so aggregate throughput as a function of (p_A,p_B) has a unique complementary optimum (high A/low B or low A/high B) and a local maximum at the symmetric independent solution. Run the paper's H-MAB (independent UCB third level) against a joint-action MAB whose arms are full (p_A,p_B) vectors, using identical exploration budgets, TXOP durations, and reward definitions. Repeat for 10 seeds and 50,000 TXOPs. If the joint-action agent reaches the complementary optimum and H-MAB plateaus at the symmetric solution with a statistically significant throughput gap, the independent power-selection assumption is the cause. If H-MAB reaches the same optimum, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V, footnote 4 concedes that the third-level power-selection agents are independent learners even though transmit powers jointly determine the observed reward. The same aggregate throughput r is back-propagated to all selected agents, so each power agent faces a moving target as other APs' power policies evolve. The central claim—80% average throughput gain over legacy 802.11 and approach to the MILP upper bound—requires that this independent scheme converges to a near-optimal power vector in general deployments. The evidence only covers 2–5 APs in open-space simulations, three APs with 3 discrete power levels in the testbed, and multi-room topologies where wall isolation weakens inter-AP power coupling. In a strongly coupled topology the independent attribution can fail without contradicting any tested result. The paper's own footnote therefore marks the load-bearing assumption, but no experiment is designed to stress it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses coordinated spatial reuse (C-SR) scheduling for IEEE 802.11bn multi-AP coordination. It contributes a mixed-integer linear programming model based on column generation that yields an upper bound on achievable throughput or fairness, and a practical probing-based scheduler using flat and hierarchical multi-armed bandits (MAB/H-MAB). The H-MAB selects transmitting APs, recipient stations, and transmit power levels in a three-level hierarchy and is updated with the aggregate throughput reward. The proposal is evaluated in simulations across multi-room, open-space, clustered, and legacy-device scenarios, and in an openwifi testbed with three APs. The headline claim is that H-MAB-based C-SR improves aggregate throughput over legacy 802.11 by on average 80% in random open-space scenarios without reducing per-station transmission opportunities, while approaching the MILP-derived upper bound.","tokens_in":24691,"tokens_out":2747,"duration_ms":29620,"significance":"If the claims hold, the paper provides a lightweight, implementable learning-based alternative to measurement-heavy C-SR scheduling, which is directly relevant to the ongoing 802.11bn standardization. The MILP upper-bound model is a solid methodological contribution and the open-source release of code is a strength that supports reproducibility. The testbed validation with openwifi demonstrates real-world feasibility, which is rare in this literature. The central risk is that the 80% gain and the closeness to the upper bound rest on an independent-learning assumption for transmit power selection that the authors themselves concede is not generally precise; the evidence presented does not stress the strongly coupled regimes where that assumption could fail.","major_comments":[{"comment":"The third-level agents select transmit powers independently (Algorithm 1, lines 12-15) while the same aggregate reward r is back-propagated to all selected agents (lines 18-19). As the authors concede in footnote 4, this independent cooperative learning framework 'may not be precise' for power selection, which formally requires a joint action. This assumption is load-bearing for the paper's central claim that H-MABs approach the MILP upper bound, because transmit powers jointly determine the SINR and hence the reward. The current evaluation does not include a strongly coupled topology in which power choices interact adversarially; the multi-room walls and the open-space 2-5 AP deployments weaken or randomize this coupling. Please add an experiment designed to stress this point, e.g., two or three APs with symmetric cross-interference and a small discrete power set, comparing H-MAB against T-Optimal and against a joint-action learner, or alternatively restrict the general claim to topologies where power coupling is weak.","section":"Section V, footnote 4 and Algorithm 1"},{"comment":"The abstract's 80% average throughput improvement is a simulation result from the random open-space scenarios of Section VI-C, not from the testbed. The testbed evaluation in Section VII-C reports the number of concurrent transmissions and the frequency of selected actions, but it does not measure aggregate throughput relative to DCF or SR. The testbed therefore validates feasibility and learning behavior, but it does not corroborate the 80% gain or the no-TXOP-reduction claim in a real system. Please report a testbed throughput comparison against a legacy baseline, or rephrase the abstract and conclusion to state explicitly that the 80% figure is simulation-only.","section":"Section VI-C and Section VII-C"},{"comment":"The convergence results in Table II show that H-MABs do not converge within 200,000 steps in the 4x4 multi-room topology (inf), and the 3x4 case requires 120,000 steps. The clustered approach of Section VI-D restores convergence, but the paper's broad statements that H-MABs 'approach the T-Optimal value' are qualified by these scalability limits. Please state the applicability envelope more precisely: the near-optimal behavior is demonstrated for small and medium topologies (up to about 12 APs with clustering), and clusterless H-MABs are not a general solution for large deployments without the clusterization mechanism.","section":"Section VI-B and Table II"}],"minor_comments":[{"comment":"The sentence 'APs can transmit data at any power lever from the set P' contains a typo; 'lever' should be 'level'.","section":"Section V"},{"comment":"The text says 'constraints (4h) are the SINR constrains' — 'constrains' should be 'constraints'.","section":"Section IV, constraints (4h)"},{"comment":"The caption of Figure 10 labels the left panel '(a) 3x3 topology', but the text and the clustering scheme refer to the 3x4 topology; please correct the label.","section":"Figure 10"},{"comment":"Table III uses the header 'Large TXOP' while the text and Table II use 'Long TXOP'; please use consistent terminology.","section":"Table III"},{"comment":"Reference [27] lists the author as 'J. C. et al.' with no full name; please provide the complete author list.","section":"References"},{"comment":"The constant M(e,m) is introduced only after the constraint is presented; defining it before the constraint would improve readability.","section":"Section IV, Eq. (4h)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits JSAC's scope and the open-source artifacts are a plus. The main risk is the untested joint-action coupling in the H-MAB power-selection level; I would ask for the additional stress experiment or a clearly scoped claim before acceptance. The novelty relative to the authors' earlier two-level H-MAB letter is incremental but sufficient for a journal publication if the power-selection issue is addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this is a solid engineering paper from a group that does this well. The most durable piece is the MILP/column-generation upper bound for C-SR scheduling, a standard optimization methodology applied to a problem that lacked a benchmark. The three-level H-MAB is a real extension of their two-level version, and they evaluate it seriously: multi-room and open-space simulations, plus an openwifi testbed, with code released. That is real evidence and they deserve credit for it.\n\nWhat I like most is that they are unusually honest about limitations. Footnote 4 in Section V concedes that independent power-selection agents are not precise because transmit powers jointly determine the reward. That is the load-bearing weakness, and they put it in writing. They also admit flat MABs do not scale past six APs and that large multi-room topologies need clustering. This is not a paper that hides its warts.\n\nThe soft spots are real but proportionate. The 80% average improvement over DCF is simulation-only, measured in random open-space topologies; the testbed demonstrates feasibility, convergence, and TXOP preservation, but not end-to-end throughput gain over legacy. Hyperparameters were tuned with Optuna on the same family of random scenarios used for evaluation, so the aggregate numbers carry some selection optimism. And the joint-action concern is not hypothetical: the same aggregate reward is back-propagated to all selected agents, so each power agent faces a moving target as other APs' policies change. The tested topologies do not stress this—2-5 APs in open space, three APs with three power levels in the testbed, and multi-room walls weaken inter-AP coupling. None of this sinks the paper. It just means the headline number should be read as \"in scenarios like these,\" not as a universal property.\n\nWho is this for? Anyone working on 802.11bn MAPC or C-SR scheduling. The upper bound model gives the community a proper benchmark, and the MAB framework is lightweight and implementable. It is also a good example of honest experimental RL for wireless. If this were a fresh submission, I would send it to peer review without hesitation. It is not a paradigm shift, but it is exactly the kind of well-grounded engineering science a serious venue should publish. I would only ask the authors to recalibrate the abstract so the 80% claim matches the evidence.","headline":"A credible, openly-coded C-SR scheduler with a genuinely useful MILP upper bound; the 80% headline is simulation-bounded, and the paper's own footnote marks its biggest soft spot.","tokens_in":25309,"tokens_out":2202,"would_cite":true,"duration_ms":22804,"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 hierarchical multi-armed bandits can schedule coordinated spatial reuse in IEEE 802.11bn networks, lifting aggregate throughput by about 80% over legacy Wi-Fi while keeping per-station transmission opportunities…","keywords":["IEEE 802.11bn","coordinated spatial reuse","multi-armed bandits","hierarchical MAB","MILP upper bound","Wi-Fi scheduling","transmit power control","MAPC"],"falsifier":"Run the H-MAB in a two-access-point topology where successful simultaneous transmission requires a specific transmit-power balance, and compare the learned power pair against the power assignment found by the MILP upper bound; if the learned pair consistently misses the joint configuration and throughput stays below the bound, the independent-power-learning assumption is falsified.","tokens_in":24365,"feed_emoji":"📶","tokens_out":6599,"duration_ms":65902,"temperature":0.7,"pith_summary":"The paper tries to establish that coordinated spatial reuse (C-SR) in the upcoming IEEE 802.11bn amendment can be scheduled by a lightweight online learner that probes transmission configurations instead of relying on channel measurements. It first builds a mixed-integer linear programming (MILP) model that computes an upper bound on the best possible schedule, then proposes a hierarchical multi-armed bandit (H-MAB) that learns which access points should transmit together, to which stations, and at what power. In simulations of random open-space topologies, the H-MAB raises aggregate throughput over legacy IEEE 802.11 by 80% on average and does not reduce the number of transmission opportunities per station. A small real testbed confirms that the learned scheduler can find useful concurrent transmission configurations. If the claim holds, dense Wi-Fi deployments could gain large throughput improvements from a scheduler simple enough to run on existing hardware.","feed_headline":"Hierarchical bandits lift Wi-Fi throughput 80% over legacy 802.11","feed_subtitle":"A probing scheduler for coordinated spatial reuse approaches the optimal schedule and preserves per-station fairness.","key_machinery":"The central object is a three-level hierarchical multi-armed bandit: a first-level agent selects the set of simultaneously transmitting access points, second-level agents select each access point's recipient station, and third-level agents select transmit power levels, with the throughput reward propagating back through all three levels. The hierarchy shrinks the effective action space from all joint configurations to a sequence of smaller choices, which is what lets the learner converge quickly. The upper-bound model is a MILP built on column generation: a main linear program allocates time shares to transmission sets, and a pricing problem iteratively generates new transmission sets with positive reduced cost until no improving set remains.","core_discovery":"The paper's central claim is that C-SR scheduling does not need RSS or CSI measurements: a probing hierarchy of multi-armed bandit agents, rewarded by the aggregate throughput of each attempted transmission configuration, can approach the performance of an optimal schedule computed offline by MILP. In random open-space simulations, the H-MAB improves aggregate throughput over legacy 802.11 by 80% on average, without cutting any station's transmission opportunities; in the testbed, it learns to run three concurrent downlink transmissions and adapts to uneven radio conditions. The flat MAB works in small topologies but its action space explodes as the network grows, while the hierarchical decomposition plus clustering restores scalability for denser deployments.","pith_inferences":["Editorial inference: the same probing hierarchy could also learn modulation-and-coding-scheme selection jointly with power and concurrency, an extension the paper leaves mostly idealized.","Editorial inference: the 80% average is measured under full-buffer downlink-only, single-channel conditions; the gain may be smaller under uplink traffic or bursty arrivals, which the paper does not test.","Editorial inference: the clustering results suggest dynamic cluster creation as the natural deployment path for very dense networks, a step the paper explicitly leaves for future work."],"forward_implications":["A central controller can run H-MABs with only wired signaling, avoiding the inter-BSS over-the-air sounding that measurement-based C-SR solutions require.","Dense random Wi-Fi deployments can expect roughly 80% higher aggregate throughput than legacy DCF while no station receives fewer transmission opportunities than it would under DCF.","When the topology changes abruptly, the learner re-converges without resetting its state, so nomadic environments do not require manual reconfiguration.","For larger multi-room deployments, splitting the network into independent clusters restores fast convergence and near-optimal throughput, pointing to a practical deployment recipe.","The MILP model gives the field a quantitative upper bound against which future C-SR schedulers can be measured, for either throughput or fairness objectives."],"supporting_citations":[{"why":"Provides an alternative C-SR scheduling algorithm based on RSSI and SINR estimation, serving as the measurement-based contrast to the probing approach.","marker":"[1]"},{"why":"Supplies the multi-armed bandit formalism and exploration-exploitation background on which both the flat MAB and H-MAB are built.","marker":"[6]"},{"why":"Earlier two-level H-MAB application to C-SR that this paper extends with a third level for transmit power selection.","marker":"[7]"},{"why":"Describes the centralized C-SR testbed operation and experimental setup used to validate the proposed scheduler in real devices.","marker":"[14]"},{"why":"Supplies the column-generation methodology for decoupling non-interfering transmission sets from time allocation, adapted for the C-SR upper-bound model.","marker":"[19, 20]"},{"why":"Justifies the independent learning framework used at the third level for transmit power selection, where joint action would formally be required.","marker":"[23]"},{"why":"Defines the TGax enterprise path-loss and simulation scenarios used in the performance evaluation.","marker":"[25]"}],"fun_headline_variants":["Hierarchical bandits boost Wi-Fi spatial reuse throughput 80%","Wi-Fi speedup: learning-based spatial reuse adds 80% throughput","Coordinated Wi-Fi: hierarchical MABs yield 80% throughput gain","Probing-based scheduler lifts 802.11 throughput 80% via bandits","Reinforcement learning for Wi-Fi C-SR: 80% faster aggregate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument hinges on assuming that learning each transmitter's power level independently from the same aggregate reward is accurate enough, even though the power choices of all simultaneous transmitters jointly decide whether a transmission succeeds.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical bandits boost Wi-Fi spatial reuse throughput 80%","Wi-Fi speedup: learning-based spatial reuse adds 80% throughput","Coordinated Wi-Fi: hierarchical MABs yield 80% throughput gain","Probing-based scheduler lifts 802.11 throughput 80% via bandits","Reinforcement learning for Wi-Fi C-SR: 80% faster aggregate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000384,"raw_usage":{"total_tokens":2014,"prompt_tokens":907,"completion_tokens":1107,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":1007}},"tokens_in":523,"tokens_out":1107,"duration_ms":9247,"temperature":1.0,"reasoning_tokens":1007,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:20:47.258585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the H-MAB in a two-access-point topology where successful simultaneous transmission requires a specific transmit-power balance, and compare the learned power pair against the power assignment found by the MILP upper bound; if the learned pair consistently misses the joint configuration and throughput stays below the bound, the independent-power-learning assumption is falsified.","supporting_citations":[{"cited_title":"Beyond Wi-Fi 7: Spatial reuse through multi-AP coordination,","cited_arxiv_id":null,"evidence_quote":"Provides an alternative C-SR scheduling algorithm based on RSSI and SINR estimation, serving as the measurement-based contrast to the probing approach."},{"cited_title":"Introduction to multi-armed bandits,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-armed bandit formalism and exploration-exploitation background on which both the flat MAB and H-MAB are built."},{"cited_title":"IEEE 802.11bn Multi-AP Coordinated Spatial Reuse with Hierarchical Multi-Armed Bandits,","cited_arxiv_id":null,"evidence_quote":"Earlier two-level H-MAB application to C-SR that this paper extends with a third level for transmit power selection."},{"cited_title":"Coordinated Spatial Reuse for WiFi Networks: A Centralized Approach,","cited_arxiv_id":null,"evidence_quote":"Describes the centralized C-SR testbed operation and experimental setup used to validate the proposed scheduler in real devices."},{"cited_title":"Bench- marking multi-agent deep reinforcement learning algorithms in cooper- ative tasks,","cited_arxiv_id":null,"evidence_quote":"Justifies the independent learning framework used at the third level for transmit power selection, where joint action would formally be required."},{"cited_title":"Merlin et al., TGax Simulation Scenarios , doc.: IEEE 802.11- 14/0980r16, Nov","cited_arxiv_id":null,"evidence_quote":"Defines the TGax enterprise path-loss and simulation scenarios used in the performance evaluation."}],"review_version":1}