{"id":"cc3e6d06-d892-462b-a3fe-ab1d635e0d8f","arxiv_id":"2507.19377","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A PPO-trained scheduler for coordinated Wi-Fi access points reduces simulated 99th-percentile delays by 15 to 30 percent compared with heuristic baselines, though overloaded scenarios are excluded.","lead":"This paper trains a deep reinforcement learning agent to pick which access points in a Wi-Fi 8 network should transmit together, aiming to cut the worst-case delay. In simulations, it reports 15 to 30 percent lower 99th-percentile delays than heuristic schedulers, but only after discarding overloaded network scenarios.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Overload filtering in Section V-A (41% of high-load deployments discarded) supports the central claim only on a selected subset; full-data evaluation is needed.","rationale":"The reader's conditional verdict is well aligned with the evidence. The simulation study is competently described, the training curves are honest, and the reported figures do show a consistent advantage for ML-G over TAT in the retained medium/high-load and scalability scenarios. I found no mathematical error in the MDP formulation, reward function, or group-feasibility condition. The load-bearing weakness is the evaluation protocol: the Section V-A exclusion of overloaded deployments is a selection on the dependent variable. Because the inclusion rule is at least one scheduler below 100 ms and the ML scheduler is one of the candidates, the set of retained deployments is enriched for conditions in which a scheduler succeeds; at minimum, the paper reports nothing about the 41% of high-load deployments that are discarded, so the generalization claim is not supported in that regime. The low-load result in Fig. 8a further contradicts the abstract's consistently and should be acknowledged in the conclusions. These are correctable by re-running the analysis on the unfiltered set or by explicitly re-scoping the claim; they do not require rejecting the paper. Thus the appropriate verdict remains conditional, pending the full-data check or a revised claim.","tokens_in":14307,"tokens_out":5458,"duration_ms":59204,"concrete_test":"Re-run the random-deployment evaluation for Fig. 9 and Fig. 8c without the Section V-A inclusion filter, keeping all 100 deployments per load regime, and report each scheduler's 99th-percentile delay on the unfiltered set together with the pairwise difference ML-G minus TAT and the fraction of deployments in which each scheduler achieves the lowest p99. If ML-G remains superior on the full set, the overload-filtering concern is resolved; if the advantage shrinks, disappears, or reverses, the central claim must be re-scoped to deployments where at least one scheduler meets the 100 ms target.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, that the trained ML schedulers consistently achieve lower 99th-percentile delays than MNP, OP, and TAT, is not established on the full distribution of random deployments because Section V-A removes every deployment in which no scheduler reaches a 99th-percentile delay below 100 ms. The filter is applied to the same metric used for comparison, and it is not blind to the ML policies: since ML-G is one of the schedulers whose sub-100 ms performance can trigger inclusion, the retained data are conditioned on the outcome being evaluated. The scale of the removal is large: Fig. 8c discards 41% of high-load deployments, Fig. 9 discards 15%, and Fig. 10 discards 4-24% depending on N. No results are reported for the discarded deployments, so the claim that ML-G consistently outperforms heuristics across a wide range of loads is unverified precisely in the regime where the schedulers are stressed. The low-load panel (Fig. 8a) independently undercuts the word consistently: there OP is the most effective scheduler and ML-G does not outperform it. The finding is therefore conditional on the non-overloaded subset and on the 100 ms threshold; it cannot support the unrestricted conclusion as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep reinforcement learning (DRL) scheduler for multi-access point coordination (MAPC) in IEEE 802.11bn Wi-Fi networks. The channel access follows a contention-based MAPC scheme in which the sharing AP collects buffer-status reports and selects a spatially reusable group of AP-STA pairs. The authors formulate group selection as a sequential decision problem with a PPO agent, a normalized observation vector of queue sizes, head-of-line delays, and channel coefficients, action masking for feasible groups, and a shaped reward that combines immediate delay reduction with a long-term worst-case-delay term. The evaluation compares two trained agents (ML-G, a general agent, and ML-E, a deployment-specialized agent) against the heuristics MNP, OP, and TAT in a simulated 4-AP enterprise scenario, across random deployments and user counts. The headline results are a 30% worst-case delay reduction over TAT in the sample deployment and a 15% reduction over TAT at 99th-percentile across random deployments, together with qualitative claims that ML-G is more stable and scales better as the number of stations grows.","tokens_in":14509,"tokens_out":2280,"duration_ms":24861,"significance":"If the central claim is validated, the paper would make a useful contribution to the emerging IEEE 802.11bn MAPC literature: it extends the authors' earlier heuristic scheduling work [5] with a modern DRL pipeline, and it provides a detailed and openly available simulation environment, concrete training hyperparameters, and an explicit comparison against three baselines. The reward-shaping ablation and the action-masking scheme are also presented clearly. However, the current evidence is conditional rather than conclusive. The reported improvements are computed on a filtered subset of deployments from which all overloaded cases have been removed using the same delay metric that is being compared, no statistical significance tests are provided, and the low-load regime explicitly contradicts the abstract's 'consistently' claim. These issues are fixable with additional analysis, but they currently prevent the paper's main conclusion from being established at the strength stated.","major_comments":[{"comment":"The evaluation excludes every deployment in which no scheduler achieves a 99th-percentile delay below 100 ms, and the retained data are therefore selected on the outcome variable of interest. The scale of the removal is substantial: 41% of high-load deployments are discarded in Fig. 8c, 15% in Fig. 9, and 4-24% in Fig. 10. Since no results are reported for the discarded deployments, the claim that ML-G 'consistently' outperforms the heuristics is unverified in exactly the stressed regime where schedulers matter most. A load-bearing re-analysis is needed: include all deployments, report the discarded set separately, or at minimum report an appropriate conservative bound that shows the relative ordering is unchanged.","section":"Section V-A and Figs. 8-10"},{"comment":"The abstract and conclusions state that the ML models 'consistently achieve lower 99th-percentile delays' across a wide range of loads, but the low-load panel of Fig. 8a shows that OP is the most effective scheduler and that ML-G does not outperform it, as the text itself acknowledges. The claim should be qualified to medium and high loads, or the low-load regime should be shown to be consistent with the stated conclusion through a more appropriate aggregate measure.","section":"Fig. 8a and abstract"},{"comment":"The quantitative improvements (30% in the sample deployment, 15% across random deployments) are reported as point estimates with no confidence intervals, error bars, or statistical significance tests. Since each data point is an independent random deployment or traffic realization, the differences between ML-G and TAT could be evaluated with paired bootstrap intervals or a nonparametric test. Without such tests, the reader cannot assess whether the observed margins are distinguishable from simulation noise, especially in Fig. 9 where the mean delay of ML-G (16.87 ms) is actually higher than that of MNP (15.55 ms).","section":"Section VI-B, Figs. 6, 8, 9, 10"}],"minor_comments":[{"comment":"The feasibility condition in Eq. (2) is taken directly from the authors' prior work [5] and is self-cited, and TAT is also introduced as originally proposed in [5]. Since both the group-selection mechanism and one of the three baselines come from the same source, the authors should state more explicitly what is new in this paper beyond replacing the scheduler with a DRL agent, and should consider whether an independent reimplementation or a third-party baseline would strengthen the comparison.","section":"Eq. (2) and Section III-B"},{"comment":"The reward shaping parameters beta and nu are fixed at 1e-3 and 1e-6 with no sensitivity analysis. A brief study of the sensitivity of the learned policy to these values would increase confidence that the reported gains are not an artifact of a particularly tuned shaping schedule.","section":"Section IV-C, reward function"},{"comment":"The text states that retransmissions are not subject to any maximum attempt limit, which may cause unbounded queue build-up in overloaded conditions and interacts with the overload-exclusion rule. The authors should either justify this assumption or discuss its effect on the tail-delay metric.","section":"Section V-C, retransmissions"},{"comment":"The caption of Fig. 6 mentions that 10% of traffic realizations are discarded, but the main text does not explain why a different discard fraction applies here than in Fig. 9. This should be clarified, or the text and caption should be made consistent.","section":"Fig. 6 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is heavily self-referential, with the system model, feasibility condition, and one of the three baselines all coming from the authors' earlier article [5]. That is not disqualifying per se, but the editor may want to ensure that the incremental novelty over [5] is substantial enough for this venue. The overload-exclusion issue is the main risk: if the full-deployment evaluation reverses the ordering, the central claim collapses, so this referee would request the additional analysis before accepting the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: this is a careful simulation study that applies PPO to 802.11bn MAPC spatial-reuse group selection, and it gets directionally positive results, but the abstract's \"consistently\" is not supported by the paper's own data. The main reason is the overload exclusion rule: deployments where no scheduler gets 99th-percentile delay below 100 ms are dropped, and 41% of high-load deployments vanish that way. That is conditioning on the outcome being measured, and it makes the 15-30% gains provisional rather than established.\n\nWhat is genuinely new here is the problem framing: treating SR group selection as a sequential decision problem and training a PPO agent in a custom Gymnasium wrapper around a detailed 802.11 simulator. The reward design, observation space, and action masking are all clearly described. The authors also honestly report that under low load OP still beats ML-G, which undercuts the blanket claim but shows they're not hiding the weakness. The comparison against MNP, OP, and TAT is straightforward, and the training details are complete enough to reproduce.\n\nThe overload filter is the biggest issue. The criterion is stated in Section V-A, and the discard rates are reported, but the paper never reports what happens in the discarded cases. Since ML-G is one of the schedulers that can trigger inclusion, the retained set is biased. A simple fix is to also show results with a higher threshold or on all data, even if overloaded cases are separated out. Second, the reward shaping term rsh = min{ε} − min{ε′} is confusing: if ε are HoL delays, then serving the oldest packet (which is what you'd want for worst-case delay) would give a negative reward, not the positive one the text claims. Likely a sign or min/max typo, but as written it's an inconsistency in the core RL machinery. Third, there are no significance tests; the boxplots overlap in some regimes, so the \"consistently\" claim is stronger than the statistics. Finally, the custom Gymnasium environment is not released, only a base simulator link, so the exact training setup isn't reproducible.\n\nNone of this invalidates the work; the simulation methodology is detailed, and the core comparison is honest within its narrow scope. But the central claim as stated in the abstract is not established on the full distribution of deployments. This is a solid conference paper if the authors tighten the claims and supply the missing robustness checks.\n\nRecommendation: yes, send it to peer review. It deserves a serious referee—the problem is relevant, the method is reasonable, and the soft spots are fixable with revision. I'd ask the authors to (1) report results on all deployments or explicitly frame the contribution as operating under a non-overload assumption, (2) fix the reward shaping equation, (3) add basic confidence intervals, and (4) release the RL environment code.","headline":"A solid DRL-for-Wi-Fi scheduling study whose headline claim is undercut by overload filtering; deserves a serious referee but needs a full-data evaluation.","tokens_in":15138,"tokens_out":6197,"would_cite":false,"duration_ms":56252,"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":"Learned Wi-Fi scheduler cuts worst-case delay by up to 30%","keywords":["coordinated spatial reuse","IEEE 802.11bn","machine learning","multi-access point coordination","reinforcement learning","scheduling","Wi-Fi 8"],"falsifier":"Re-run the evaluation on all 100 deployments from Fig. 9 without excluding the 15% currently discarded as overloaded, and compute each scheduler's 99th-percentile delay; if ML-G's delay is not below TAT's on that full set, the claim that the learned scheduler outperforms the best baseline across loads fails on the excluded cases.","tokens_in":14073,"feed_emoji":"📶","tokens_out":6105,"duration_ms":58188,"temperature":0.7,"pith_summary":"This paper tries to establish that a deep reinforcement learning scheduler can choose, at each transmission opportunity, which spatially reusable access-point–station groups to serve, and that this learned choice reduces worst-case latency in dense overlapping Wi-Fi deployments below what three fixed heuristics achieve. The agent is trained with proximal policy optimization on observations of per-station queue sizes, head-of-line delays, and channel coefficients, selecting among precomputed spatial-reuse groups. In the paper's simulations, the learned schedulers reach 99th-percentile delays as low as 72.98 ms in a fixed four-AP deployment, about 30% below the best heuristic, and 64.54 ms across 100 random deployments, about 15% below the best heuristic. If the result holds, it suggests that Wi-Fi 8's multi-AP coordination can be steered by learned policies rather than hand-tuned rules, with direct consequences for latency-sensitive applications.","feed_headline":"Learned Wi-Fi scheduler cuts worst-case delay by up to 30%","feed_subtitle":"A PPO-trained agent beats three heuristic schedulers on 99th-percentile latency in dense multi-AP Wi-Fi deployments.","key_machinery":"The load-bearing machinery is the sequential decision-making formulation of MAPC scheduling as a Markov decision process, solved with PPO in an actor–critic architecture over a discrete action space of feasible spatial-reuse groups. Feasibility is precomputed by admitting a group only if, for every station in the group, the group throughput expression $|M_n|R^{\\mathrm{CoSR}}_{i,n}/R^{\\mathrm{ST}}_i$ is at least 1, so the agent never wastes decisions on groups that cannot beat single-user transmission. At each step the observation $s=[\\bar{\\delta},\\bar{\\varrho},\\bar{h}]$ collects normalized per-STA head-of-line delays, queue sizes, and channel coefficients; a binary mask removes groups whose intended receivers have empty queues; and the reward $r=r_{sh}+r_{lg}$ shapes learning by rewarding dispatch of the oldest head-of-line packet while also penalizing growth of the maximum queueing delay.","core_discovery":"The paper's central claim is that a PPO-trained policy, choosing a valid coordinated-spatial-reuse group at every transmission opportunity from a masked action space, minimizes network-wide worst-case delay and generalizes across deployments and traffic types. The supporting evidence is a set of IEEE 802.11bn simulations: in a sample 4-AP, 16-STA deployment the deployment-specialized agent (ML-E) achieves a 99th-percentile delay of 72.98 ms versus 103.85 ms for the best heuristic TAT, while a generalist agent (ML-G) evaluated on 100 random deployments achieves 64.54 ms versus 75.92 ms for TAT under per-STA loads of 10–90 Mb/s. The paper argues that the learned policy's advantage comes from balancing delay reduction with transmission efficiency instead of always serving the oldest packet or the fullest queues, and that this advantage grows as the number of STAs increases from 8 to 20.","pith_inferences":["If the high-load deployments discarded in the paper (41% of cases in Fig. 8c and 15% in Fig. 9) were included in the comparison, the headline improvement over TAT could shrink or disappear; testing the trained agent on those discarded deployments is the cleanest extension of the reported results.","Because the observation is already per-STA and the action masking is structural, the same recipe could likely be extended to jointly select transmit power or MCS along with the SR group, capabilities the paper lists as future work.","In the low-load regime OP matches or beats ML-G, suggesting the realistic value of the learned scheduler is concentrated in medium-to-high load and bursty traffic, where queue and interference interactions are too complex for hand-coded rules.","Since a separate ML-G model is trained for each network size, a broader claim about scaling to arbitrary WLANs would require a size-invariant observation and action space; the paper itself identifies this as an open direction."],"forward_implications":["Trained in simulation, the learned scheduler can run inside the sharing AP at each TXOP and issue the same MAPC trigger frames, so the 802.11bn control-plane exchange does not need to change for the scheme to be deployed.","In random unseen deployments, ML-G keeps its 99th-percentile delay below roughly 100 ms including outliers under high load, while MNP and OP frequently exceed 100 ms, meaning the approach is most useful precisely where fixed heuristics become unreliable.","As station count grows from 8 to 20 at fixed average network load, the gap between ML-G and the best heuristic widens, indicating learned scheduling becomes more valuable in denser scenarios.","The trained policy deliberately serves a mix of delay-priority levels rather than always the highest-delay group, which is why it avoids the starvation patterns seen with the oldest-packet heuristic.","Reward shaping makes training converge faster and reach slightly better final 99th-percentile delays than the same PPO setup without shaping."],"supporting_citations":[{"why":"Supplies the three heuristic schedulers (MNP, OP, TAT) used as baselines and the SR-group feasibility method used to precompute the action space.","marker":"[5]"},{"why":"Defines the unified MAPC framework whose control-frame exchange the system model mimics.","marker":"[17]"},{"why":"Provides the standardized reinforcement-learning interface that couples the simulator to the learning agent.","marker":"[18]"},{"why":"Supplies the proximal policy optimization algorithm used to train the actor and critic networks.","marker":"[20]"},{"why":"Sets the enterprise deployment, path-loss, and wall-loss parameters used in the simulations.","marker":"[22]"},{"why":"Provides the packet-error-rate versus SNR curves used to map estimated SINR to MCS values.","marker":"[23]"},{"why":"Supplies the PPO implementation with action masking used for training the agents.","marker":"[24]"}],"fun_headline_variants":["Wi-Fi scheduler trained with RL slashes worst-case latency by 30%","PPO-trained agent cuts Wi-Fi worst-case delay 30% in dense networks","Reinforcement learning finds Wi-Fi coordination to cut latency by 30%","DRL-based Wi-Fi scheduler cuts worst-case latencies by up to 30%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison keeps only deployments in which at least one scheduler already achieves a 99th-percentile delay below 100 ms, discarding up to 41% of high-load deployments, so the claimed advantage is established only on that non-overloaded subset.","fun_headline_variants_meta":{"raw":{"variants":["Wi-Fi scheduler trained with RL slashes worst-case latency by 30%","PPO-trained agent cuts Wi-Fi worst-case delay 30% in dense networks","Reinforcement learning finds Wi-Fi coordination to cut latency by 30%","DRL-based Wi-Fi scheduler cuts worst-case latencies by up to 30%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000747,"raw_usage":{"total_tokens":3342,"prompt_tokens":975,"completion_tokens":2367,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":2281}},"tokens_in":591,"tokens_out":2367,"duration_ms":16603,"temperature":1.0,"reasoning_tokens":2281,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:53:30.462022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the evaluation on all 100 deployments from Fig. 9 without excluding the 15% currently discarded as overloaded, and compute each scheduler's 99th-percentile delay; if ML-G's delay is not below TAT's on that full set, the claim that the learned scheduler outperforms the best baseline across loads fails on the excluded cases.","supporting_citations":[{"cited_title":"Enabling Reliable Latency in Wi-Fi 8 Through Multi-AP Joint Scheduling,","cited_arxiv_id":null,"evidence_quote":"Supplies the three heuristic schedulers (MNP, OP, TAT) used as baselines and the SR-group feasibility method used to precompute the action space."},{"cited_title":"IEEE 802.11-25/0502r0: Details on the unified MAPC framework,","cited_arxiv_id":null,"evidence_quote":"Defines the unified MAPC framework whose control-frame exchange the system model mimics."},{"cited_title":"TGax Simulation Scenarios,","cited_arxiv_id":null,"evidence_quote":"Sets the enterprise deployment, path-loss, and wall-loss parameters used in the simulations."},{"cited_title":"802.11be Packet Error Rate Simulation for an EHT MU Single-User Packet Format,","cited_arxiv_id":null,"evidence_quote":"Provides the packet-error-rate versus SNR curves used to map estimated SINR to MCS values."},{"cited_title":"Stable-Baselines3: Reliable Reinforcement Learning Implementations,","cited_arxiv_id":null,"evidence_quote":"Supplies the PPO implementation with action masking used for training the agents."}],"review_version":2}