{"id":"4b239322-e181-4882-a883-eb6e6045d46c","arxiv_id":"2506.07980","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DesRUTGe combines PPO-based reinforcement learning with decentralized federated learning to generate 24-hour traffic profiles that match real detector counts better than SUMO's RouteSampler and a centralized RUTGe baseline.","lead":"The paper introduces DesRUTGe, a tool that uses reinforcement learning and decentralized federated learning to generate 24-hour urban traffic patterns for the SUMO simulator, trained on Barcelona detector data. It reports lower error than SUMO's RouteSampler and a centralized predecessor, while sharing only model parameters between neighboring zones to preserve privacy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's stopping rule (threshold 1e-3 on |T-O|) cannot be satisfied for non-integer hourly-count targets, so the described generation loop cannot terminate; the reported MAE comparison (Table III) depends on an unstated implementation detail.","rationale":"The reader's weakest assumption concerns the realism definition. I agree that count-matching is an incomplete realism proxy, but the most load-bearing weakness sits one step earlier: the exact stopping rule of Algorithm 1 is not executable for integer detector counts paired with fractional mean targets. This is an internal inconsistency, not a matter of external standards, and it directly threatens the quantitative basis of the central claim. If the threshold applies to normalized values, a one-line correction plus re-verification of Table III handles it; if not, the reported MAE values were produced by a different algorithm. The provided GitLab code allows a definitive check, which is why I retain the reader's conditional verdict rather than escalating to rejection. I do not allege any bad faith; the issue is that the manuscript's algorithmic description and its reported results are not currently reconcilable.","tokens_in":22382,"tokens_out":8803,"duration_ms":112022,"concrete_test":"Clone the public DesRUTGe repository (https://gitlab.com/compromise3/desrutge), locate the 1h-TGA loop corresponding to Algorithm 1, and execute it with a non-integer target (e.g., detector 4010's mean hourly count) under the stated hyperparameters. If it fails to terminate within a bounded budget, Algorithm 1 as published cannot produce Table III. If it terminates because observed and target intensities are normalized to [0,1] before the threshold comparison, then Eq. (4) and Algorithm 1 must be corrected to state that normalization, and Table III should be re-verified under that documented criterion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the headline accuracy claim to hold, Algorithm 1 must terminate and produce the errors in Table III. As written, it cannot. Section IV-B1 defines the reward in Eq. (4) with a goal threshold of 10^-3 on |T-O|, and Algorithm 1 loops while |observed traffic intensity - target[h]| > 0.001. But Section V-A states targets are mean hourly vehicle counts per detector, and Section IV-A says detectors output integer hourly vehicle counts. A mean target such as 102.3 vehicles/h is never within 0.001 of an integer count (102 or 103), so the loop never exits. The text also says 'mean squared error' before Eq. (4), but the equation uses absolute error. Either the threshold applies to normalized observations (normalization is described only for the state s in Section IV-B1), or there is an unstated maximum-iteration/early-stop rule. Without knowing which, Table III's mean MAE 25.88 vs 69.32 cannot be reproduced from the paper, and the central 'more accurate' claim rests on underdocumented behavior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DesRUTGe, a decentralized federated learning framework for generating 24-hour urban traffic patterns in the SUMO simulator. Each traffic detector and its Voronoi-based zone run a local PPO agent that repeatedly adjusts the number of injected vehicles in order to match hourly detector-count targets. Local policies are exchanged with selected peers (geographic neighbors or affinity-based clusters) and aggregated with FedAvg, with no central coordinator. The method is evaluated on ten Barcelona detectors using real-world hourly counts, and the authors compare it against SUMO's routeSampler.py and their own earlier centralized RUTGe system. The headline quantitative result is a mean MAE of 25.88 for DesRUTGe versus 69.32 for routeSampler.py over ten executions.","tokens_in":22684,"tokens_out":6696,"duration_ms":87268,"significance":"If the implementation details are clarified, the paper makes a useful contribution: it provides an open-source DFL+DRL pipeline for SUMO, uses only open data (OpenStreetMap and public Barcelona counts), releases code, and reports a quantitative comparison with a standard SUMO tool. The decentralized training architecture is a plausible step toward privacy-preserving and scalable traffic generation, and the comparison against routeSampler.py is a relevant baseline that is often missing in this literature. However, the main accuracy claim currently rests on an underspecified and internally inconsistent stopping rule in Algorithm 1, and the evaluation defines 'realistic traffic' solely by count-matching at ten detectors. The contribution is therefore promising but needs a major revision before the central claims can be accepted.","major_comments":[{"comment":"The stopping criterion in Algorithm 1 is inconsistent with the data and with the reward definition. The targets are mean hourly vehicle counts (Section V-A), which are generally non-integer values such as 102.3 vehicles/h, while the observed traffic intensity from SUMO induction loops is an integer vehicle count (Section IV-A). The loop condition |observed − target| > 0.001 can therefore never be satisfied for such non-integer targets, and Algorithm 1 contains no maximum-iteration or early-stop rule. The text before Eq. (4) also says the goal is a 'mean squared error' smaller than 0.001, but Eq. (4) uses the absolute error |T−O|. As written, the generation loop cannot terminate, and Table III cannot be reproduced from the paper. Please specify whether the threshold applies to a normalized error, and state the actual termination rule (or early-stop condition) used to produce Table III.","section":"§IV-B1, Eq. (4), Algorithm 1"},{"comment":"The residual-subtraction mechanism assumes that the residual traffic vector computed from an isolated one-hour simulation is additive in the final 24-hour route file. The paper does not explain how residual[] is measured in Algorithm 1 (line 9), nor whether the simulation for hour i includes previously generated routes. Since the final route file concatenates all hourly route sets, congestion and route interactions can change the residual that actually appears in hour j, so subtracting the isolated residual from target[j] may systematically bias the generated counts. Please clarify the implementation and provide a validation of the full 24-hour simulation against the target profile, rather than only reporting errors of the concatenated hourly generation procedure.","section":"§IV-C2, Algorithm 2, Eq. (5)"},{"comment":"The evaluation defines realism and accuracy solely as matching hourly vehicle counts at the ten detector locations, and Table III reports detector-level MAE. The title and abstract claim generation of 'realistic traffic patterns,' but no route-level, network-level, or sub-hourly temporal metric is reported. Two route sets can produce identical detector counts while having very different trip-length distributions, path choices, or congestion patterns away from detectors. If the claim is limited to count calibration, the paper should state that scope explicitly; if the claim is about realism, additional validation is needed.","section":"§V-E and §VI-D, Table III"},{"comment":"The statement that DesRUTGe 'consistently achieves lower deviations' is contradicted by the row for detector 4063, where routeSampler.py achieves a mean MAE of 33.49±8.83 versus 54.71±0.01 for DesRUTGe. The average improvement (25.88 vs 69.32) is still favorable to DesRUTGe, but the claims of consistent superiority and of better performance 'particularly during peak congestion periods' are not supported for this high-intensity detector. Please either qualify the claim or provide an explanation for this exception.","section":"Table III and §VI-D"}],"minor_comments":[{"comment":"The state is first described as a natural number and then as normalized to [0,1], while the action is described as continuous in [−0.1,+0.3]. The PPO implementation for a continuous action space (policy distribution, squashing, and re-normalization after the multiplicative update) is not specified. Please clarify.","section":"§IV-B1"},{"comment":"The metric named 'Relative Error' is defined as RE = T − D, which is a signed absolute difference, not a relative error (there is no division by T). A different name or a proper relative-error definition would avoid confusion.","section":"§V-E, Eq. (6)"},{"comment":"Several rows report standard deviations of 0.01 for DesRUTGe (e.g., detectors 4026 and 8009). Please state whether the ten executions are independent training runs or repeated evaluations of the same trained policy, since near-zero standard deviations are surprising for stochastic RL training with SUMO.","section":"Table III"},{"comment":"The hyperparameter discussion says the configuration is 'consistent with configurations known to perform well in discrete action spaces,' but the method uses a continuous action range. This sentence should be corrected or reconciled with the action-space description.","section":"§V-C"},{"comment":"There is a typo in the figure caption: 'yelow lines' should be 'yellow lines.'","section":"§VI-C"},{"comment":"The sentence 'This setup ensured parity in simulation effort and exposure between both approaches' is misleading, because routeSampler.py is not a learning method and does not have training episodes; parity of effort is not a meaningful fairness criterion for this baseline.","section":"§VI-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of an applied machine-learning or simulation journal, and the authors have provided a code link. I recommend asking the authors to confirm, with reference to the released code, the exact stopping rule used for Algorithm 1 and to describe how residual[] is computed in practice. The editor may also wish to verify that the routeSampler.py baseline was given a fair configuration (same network, detector positions, and OD sampling), since the comparison is central to the paper's headline claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll give you the short version first: this paper describes a genuinely new artifact—neighbor-only decentralized federated RL (PPO) for SUMO traffic demand generation—with public code and a real Barcelona evaluation. The residual-vector chaining that stitches hourly simulations into a 24-hour profile is a thoughtful engineering contribution. That part is solid and worth reading.\n\nThe soft spots are real, and one is load-bearing. Algorithm 1's termination condition (|observed − target| > 0.001) cannot be satisfied for non-integer hourly-count targets against integer detector counts. A mean count like 102.3 is never within 0.001 of an integer. Unless the implementation uses a normalized error, a max-iteration cap, or a different threshold, the loop never exits. Table III's MAE values must have come from some undocumented stopping rule, which makes the central \"more accurate than RouteSampler\" claim non-reproducible. That's the first thing to fix.\n\nSecond, the \"realistic traffic patterns\" claim overreaches. The reward and all evaluation metrics only measure how well the generated flows match average hourly counts at ten detector locations. That's a useful calibration target, but it does not establish realism in any broader sense (route distributions, congestion away from detectors, within-hour dynamics). The paper should either narrow the claim or add evaluation.\n\nThird, \"privacy-preserving\" is asserted but never analyzed. Sharing model parameters is a common FL privacy posture, but the paper itself notes inversion and membership-inference risks; without an attack model or any empirical privacy evaluation, that label is unsupported.\n\nMinor issues: the text says \"mean squared error\" just before Eq. (4), which is absolute error; the two affinity-based sharing strategies are said to produce the same partition, so their comparative discussion is vacuous; and some standard deviations (0.01) look artificially small, suggesting near-deterministic runs that the text doesn't explain.\n\nWho it's for: the ITS and traffic-simulation community, especially people who need to generate demand inputs for SUMO. RL researchers won't find new methods, but the system itself is useful.\n\nMy recommendation: send it to peer review, but with major revision. The framework is valuable and the code is public; the accuracy claim can likely be fixed by documenting the actual stopping rule and rerunning or clarifying the evaluation. Right now, I wouldn't cite the numerical results until that gap is closed.","headline":"A genuinely new decentralized RL traffic generator with public code, but the headline accuracy numbers are not reproducible from the paper as written because Algorithm 1's stopping threshold cannot be satisfied on integer detector counts.","tokens_in":23216,"tokens_out":4253,"would_cite":true,"duration_ms":51103,"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":"This paper claims that per-detector reinforcement-learning agents, trained in a decentralized federation, can generate realistic 24-hour traffic for SUMO and reduce mean absolute count error to 25.88 vehicles/hour, versus 69.32 for SUMO's…","keywords":["traffic generation","SUMO","decentralized federated learning","deep reinforcement learning","proximal policy optimization","induction loop detectors","urban traffic simulation","24-hour traffic profiles"],"falsifier":"Take the best route file DesRUTGe produces for a Barcelona weekday and measure vehicle counts at induction-loop locations held out from training, or re-aggregate the same run at 15-minute resolution; if the error at unseen detector sites or within-hour scales is no better than routeSampler.py's, the central realism claim would be refuted.","tokens_in":22250,"feed_emoji":"🚦","tokens_out":9940,"duration_ms":113338,"temperature":0.7,"pith_summary":"This paper claims that a city's 24-hour traffic demand can be synthesized for the SUMO simulator by assigning each induction-loop detector its own deep-reinforcement-learning agent, training that agent to adjust vehicle injections until simulated hourly counts match the detector's historical averages, and then letting neighboring agents exchange only their model weights. The case for caring is that the input data are minimal, just the hourly average counts that monitoring stations already record, and the scheme never routes raw sensor data to a central server. Evaluated on real weekday profiles from Barcelona, the paper reports a mean absolute error of 25.88 vehicles/hour across ten detectors, versus 69.32 for SUMO's standard routeSampler.py, and also shows the decentralized per-zone agents tracking low-, medium-, and high-traffic detectors more closely than the previous centralized single-agent version. The paper concludes that realistic, privacy-preserving traffic synthesis is achievable with local learning plus neighbor-to-neighbor model sharing.","feed_headline":"Decentralized agents cut SUMO traffic error 2.7x","feed_subtitle":"Per-zone learning recreates a Barcelona weekday from hourly loop counts, sharing weights but no raw data.","key_machinery":"The mechanism that carries the argument is the closed loop between a PPO agent and the SUMO simulator, wrapped in a decentralized federation graph. The state is a normalized count of vehicles scheduled for injection; the action is a multiplicative adjustment factor in the range [-0.1, +0.3]; the observation is the average intensity returned by virtual induction-loop detectors after a one-hour run; and the reward is the negative absolute difference to the target, plus a small per-step penalty and a large bonus once the error drops below $10^{-3}$. To chain hours, each run also returns a 23-hour residual vector of vehicles still on the network, and the next hour's target is reduced by that residual. Decentralization is organized as an undirected graph whose vertices are detectors; after local training each node averages its PPO weights with its neighbors' weights via FedAvg, with three possible neighbor choices: zones sharing a Voronoi boundary, zones with similar traffic volume, or zones with similar standardized 24-hour shapes.","core_discovery":"DesRUTGe is presented as an automatic way to go from one 24-dimensional vector of average hourly vehicle counts per detector to a SUMO route file that reproduces those counts. The discovery claim is that the accuracy of the generated traffic does not require a central coordinator: each detector's Voronoi zone trains its own PPO policy inside the SUMO loop, and after every round of local training the agents average their policy weights only with chosen neighbors, using coordinate-wise averaging. The reported result is a mean MAE of 25.88 vehicles/hour across 10 executions and 10 Barcelona detectors, compared with 69.32 for routeSampler.py, with the largest gains at high-traffic detectors, and a comparison against the prior centralized approach showing that per-zone agents follow each detector's individual profile instead of collapsing toward a city-wide average.","pith_inferences":["Editorial inference: the reward loop is agnostic about what the detector counts, so the same controller could be pointed at other demand signals, such as occupancy, emissions, or floating-car speeds, if per-zone targets for those signals were available.","Editorial inference: since realism is measured only at the ten detectors that define the targets, a natural stress test is to hold out some detectors or measure link flows elsewhere; the reported accuracy does not by itself guarantee that off-detector congestion is realistic.","Editorial inference: geographic adjacency won in this ten-node Barcelona network, but one could test whether affinity sharing pays off in cities with spatially separated zones of identical land use, where similar traffic patterns are not adjacent.","Editorial inference: wall-clock training grew from 1.4 hours centralized to 3.34 hours decentralized on the same machine; a testable question is whether gossip-style partial aggregation cuts that gap without erasing the accuracy margin."],"forward_implications":["A complete one-day route file for SUMO can be produced from hourly average counts at loop detectors, without origin-destination matrices or individual vehicle traces.","Raw mobility measurements stay at the local zone: the only objects crossing the network are neural-network weights.","Geographic neighbor sharing is the topology the paper recommends, as it converges to lower hourly error faster than volume-based or pattern-based clustering.","The per-zone design fixes a known failure of the centralized prototype, whose city-wide average target cannot track a low-, medium-, or high-intensity detector individually.","Because hourly targets are adjusted for residual traffic from earlier hours, vehicles that overflow an hour do not cause double-counting in the next hour's generation."],"supporting_citations":[{"why":"Supplies the SUMO microscopic simulator and the routeSampler.py count-calibration baseline used for comparison.","marker":"[1]"},{"why":"Defines PPO, the clipped-objective policy-gradient algorithm that every local agent uses to adjust vehicle injections.","marker":"[56]"},{"why":"Defines FedAvg, the coordinate-wise averaging operation used when each node aggregates the PPO weights it receives from neighbors.","marker":"[17]"},{"why":"Describes the prior centralized single-agent prototype whose city-wide average targets the per-zone results are contrasted with.","marker":"[45]"},{"why":"Provides the earlier SUMO-based route-generation pipeline that this framework adapts for vehicle injection and detector feedback.","marker":"[3]"},{"why":"Introduces Voronoi diagrams, the geometric partitioning used to assign each road segment to exactly one traffic detector.","marker":"[58]"},{"why":"Establishes the neighbor-only decentralized federated learning pattern for traffic tasks that motivates the DFL architecture.","marker":"[47]"}],"fun_headline_variants":["Decentralized learning cuts SUMO error 2.7x","Per-zone agents beat RouteSampler on Barcelona traffic","Privacy-first traffic gen for SUMO outdoes past tools","Federated learning gives SUMO real Barcelona traffic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that matching the average number of vehicles per hour at ten fixed detector locations defines 'realistic traffic patterns'; if a route file can match those counts while remaining unrealistic in route distribution, off-detector congestion, or sub-hourly timing, the reported accuracy figures do not establish the realism claim.","fun_headline_variants_meta":{"raw":{"variants":["Decentralized learning cuts SUMO error 2.7x","Per-zone agents beat RouteSampler on Barcelona traffic","Privacy-first traffic gen for SUMO outdoes past tools","Federated learning gives SUMO real Barcelona traffic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1347,"prompt_tokens":926,"completion_tokens":421,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":353}},"tokens_in":542,"tokens_out":421,"duration_ms":5734,"temperature":1.0,"reasoning_tokens":353,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:21:17.652082+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the best route file DesRUTGe produces for a Barcelona weekday and measure vehicle counts at induction-loop locations held out from training, or re-aggregate the same run at 15-minute resolution; if the error at unseen detector sites or within-hour scales is no better than routeSampler.py's, the central realism claim would be refuted.","supporting_citations":[{"cited_title":"Microscopic Traffic Simulation using SUMO,","cited_arxiv_id":null,"evidence_quote":"Supplies the SUMO microscopic simulator and the routeSampler.py count-calibration baseline used for comparison."},{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Defines FedAvg, the coordinate-wise averaging operation used when each node aggregates the PPO weights it receives from neighbors."},{"cited_title":"Rutge: Realistic urban traffic generator for urban environments using deep reinforcement learning and sumo simulator,","cited_arxiv_id":null,"evidence_quote":"Describes the prior centralized single-agent prototype whose city-wide average targets the per-zone results are contrasted with."},{"cited_title":"STGT: SUMO-based traffic mobility generation tool for evaluation of vehicular networks,","cited_arxiv_id":null,"evidence_quote":"Provides the earlier SUMO-based route-generation pipeline that this framework adapts for vehicle injection and detector feedback."},{"cited_title":"V oronoi diagrams—a survey of a fundamental geometric data structure,","cited_arxiv_id":null,"evidence_quote":"Introduces Voronoi diagrams, the geometric partitioning used to assign each road segment to exactly one traffic detector."},{"cited_title":"A decentralized federated learning-based spatial–temporal model for freight traffic speed forecasting,","cited_arxiv_id":null,"evidence_quote":"Establishes the neighbor-only decentralized federated learning pattern for traffic tasks that motivates the DFL architecture."}],"review_version":1}