{"id":"837abb1f-46d0-47ec-8478-dd054b0da9b8","arxiv_id":"1909.00395","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An open-source SUMO framework for adaptive traffic signal control is introduced, and experiments on a two-intersection network show Max-pressure outperforms deep reinforcement learning controllers.","lead":"This paper presents an open-source software framework for testing different adaptive traffic signal controllers in a traffic simulation, including traditional and machine-learning based methods. The authors use the framework to compare the controllers on a small two-intersection network and report that a simple pressure-based controller performs best, while the deep reinforcement learning controllers are more sensitive to hyperparameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Same three-hour demand scenario is used for both hyperparameter tuning (Section IV-A) and final comparison (Section IV-B); the Max-pressure ranking may reflect in-sample selection rather than a fair comparison, especially for high-variance learning controllers.","rationale":"The paper makes a useful, modest contribution: an open-source SUMO framework with multiple traffic signal controllers, a hyperparameter sensitivity study, and a small comparative experiment. The code is public and the authors explicitly limit their conclusions in Section V, stating that the results are not claimed to hold in all environments and circumstances. The reader's weakest assumption focuses on the single synthetic two-intersection network; that is a real limitation, but it is not the most load-bearing issue because the authors already disclaim universal generalization. The sharper concern is that the same network and demand scenario are used both to tune hyperparameters (Section IV-A) and to evaluate final performance (Section IV-B). This creates an in-sample selection effect that works against high-dimensional, stochastic learning controllers and in favor of low-dimensional heuristics like Max-pressure. The framework contribution itself is not invalidated by this concern, and the public code is genuine supporting evidence, though the lack of a pinned commit and detailed RL hyperparameters weakens reproducibility. The conditional verdict remains appropriate: the framework is a solid starting point, but the comparative ranking should be re-tested on a held-out demand scenario before being reported as a reliable property of the controllers themselves.","tokens_in":15593,"tokens_out":9968,"duration_ms":101209,"concrete_test":"Use the repository to run a three-way split on the same two-intersection network: tune each controller on demand profile A (8 seeds), select the best hyperparameters on a second, held-out profile B, and report final mean/median travel time plus standard deviation over 32 seeds on profile C (a third profile). If Max-pressure still has the lowest mean/median and the DQN/DDPG gap remains, the ranking is robust. If any learning controller closes the gap or beats Max-pressure on the held-out profile, the Section IV-B conclusion is an artifact of same-scenario tuning and should be reworded. As a minimum, the authors should report the full hyperparameter grid and RL training settings so the audit can be reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is not that the two-intersection network is unrepresentative, because Section V explicitly disclaims generalizing to all environments. It is that Section IV-A selects each controller's hyperparameters on the same network and the same three-hour dynamic demand scenario that Section IV-B then uses to judge performance. Hyperparameter choices are therefore optimized for the evaluation scenario. This matters asymmetrically: DQN and DDPG have many hyperparameters and stochastic training, so the 8-seed grid search can pick configurations that happen to fit this particular demand pattern; Max-pressure has one or two tunable parameters and has far less room to overfit. The result presented as 'controllers with optimal hyperparameters' is thus an in-sample comparison, and the claim that Max-pressure achieves the lowest mean, median and standard deviation in Section IV-B is not as secure as it appears. The concern is compounded by the paper not reporting the grid ranges or the RL training hyperparameters (learning rate, batch size, target update interval, exploration schedule, action repeat), so a reader cannot tell how thoroughly the learning controllers were tuned. A held-out evaluation is required before 'best' is attributed to the algorithm rather than to the tuning protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an open-source framework for adaptive traffic signal control in the SUMO microsimulator, with implementations of Uniform, Webster's, Max-pressure, SOTL, DQN, and DDPG controllers. The framework is demonstrated on a two-intersection network under a simulated three-hour dynamic demand scenario: Section IV-A performs a grid search over each controller's hyperparameters using eight random seeds, and Section IV-B compares the selected 'optimal' configurations over 32 additional simulations using travel time, queue, and delay measures. The authors report that Max-pressure achieves the lowest mean, median, and standard deviation of travel time, while learning-based controllers show higher hyperparameter sensitivity and larger variance. The manuscript emphasizes the framework's value as a reusable, parallelized starting point for adaptive signal control research.","tokens_in":15924,"tokens_out":3530,"duration_ms":35219,"significance":"The framework itself is a useful community resource: the code is open-source, integrates with SUMO, and provides a distributed actor-learner architecture that makes the included RL controllers tractable on modest hardware. The authors are explicit (Section V) that their results are not intended to establish the superiority of Max-pressure in all environments, which moderates the scope of the empirical claim. If the comparison is made fair and reproducible, the finding that a hyperparameter-sparse heuristic can outperform heavily-tuned deep RL controllers on a small network is a plausible and useful counterpoint to the prevailing emphasis on learning-based methods. The reported multi-seed boxplots and confidence intervals support internal comparisons, although formal significance testing is absent.","major_comments":[{"comment":"The same two-intersection network and the same three-hour dynamic demand scenario are used for both hyperparameter selection (Section IV-A) and the final performance evaluation (Section IV-B). The controllers are therefore compared in-sample: the hyperparameter configurations are selected on the exact scenario used to rank the controllers, which asymmetrically favors low-variance methods such as Max-pressure over the more hyperparameter-rich DQN and DDPG. This undermines the Section IV-B claim that Max-pressure achieves the best performance, because the ranking may reflect the tuning protocol rather than the algorithms. A held-out evaluation on a different demand profile or network, or an explicit cross-validation scheme, is needed before the ranking is attributed to the controllers.","section":"Section IV-A and IV-B"},{"comment":"The grid-search details are not reported: the manuscript does not state the number of hyperparameter configurations per controller, the ranges of the grid, or the criterion used to select the 'best' configuration (the Figure 3 caption mentions ordering by sum of mean and standard deviation, but the formal selection rule is not defined). For DQN and DDPG, the training hyperparameters (learning rate, batch size, target update interval, exploration schedule, action repeat, replay buffer size, and network size) are omitted entirely, although these are listed as free parameters in the framework. Without these details, a reader cannot determine whether the learning controllers were tuned as thoroughly as the non-learning controllers, which is load-bearing for the comparative conclusion.","section":"Section IV-A"},{"comment":"The performance differences among controllers are described qualitatively (e.g., 'approximately equal performance' for DQN, DDPG, Uniform, and Webster's) with boxplots and confidence intervals, but no significance tests or effect-size measures are reported. The claim that Max-pressure has the lowest mean, median, and standard deviation would be strengthened by a statistical test of paired or unpaired differences across the 32 seeds, and the 'approximately equal' grouping should be supported by a test showing that the observed differences are not statistically significant.","section":"Section IV-B"},{"comment":"The conclusion states that 'the Max-pressure controller was found to achieve the best performance' and that 'learning-based controllers can be further developed to offer improved performance.' While the authors disclaim universal generalization, the conclusion presents the in-sample comparison as an established finding. The wording should be tempered to reflect that the result is specific to the tested network and demand scenario, and that the comparison was not performed on a held-out scenario.","section":"Section V"}],"minor_comments":[{"comment":"The caption acknowledges that differing axis scales make direct visual comparison biased. Consider using a common normalized scale or subplots with identical axis limits so that hyperparameter sensitivity is visually comparable across controllers.","section":"Figure 3"},{"comment":"The manuscript does not state how many hyperparameter configurations were tested for each controller, nor does it list the grid ranges. Adding a table of grid ranges and the number of configurations would improve reproducibility and would help readers assess whether the search was exhaustive enough.","section":"Section IV-A"},{"comment":"The Max-pressure pressure definition uses the difference of total vehicle counts on incoming and outgoing lanes, whereas the original max-pressure controller uses turning-movement-specific pressures. The authors should state that this is a simplified version and justify its adequacy for the tested network.","section":"Section III-A.3, Eq. (1)"},{"comment":"The contribution statement says the framework 'was designed to scale to develop adaptive controllers for any SUMO network,' but the experiments only cover a two-intersection network. The claim would be better phrased as a design goal rather than an achieved property.","section":"Section II-C"},{"comment":"SOTL is omitted from the intersection-level queue and delay plots to improve readability. The text explains this, but it would be clearer to state that SOTL values are outside the ordinate range in the caption as well.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is on the boundary between a software demonstration and an empirical comparison paper. The strongest contribution is the open-source framework, and the empirical finding is secondary. The in-sample tuning issue is fixable by rerunning the evaluation on a held-out demand scenario or adding a cross-validation-style protocol, and by reporting the full hyperparameter grids and RL training settings. I would also suggest asking the authors to de-emphasize the 'best controller' framing in the abstract and conclusion, since the evidence base is a single small network. The self-citation of [32], [38], [52] is appropriate given the prior work, but the novelty relative to those works should be stated more explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: this is a useful open-source engineering contribution, not a breakthrough empirical result. The headline claim, that Max-pressure beats deep RL on a two-intersection network, is plausible but not well-supported, because the hyperparameters were tuned on the same demand pattern used for the final comparison.\n\nWhat is actually new: the framework. Five controllers—Webster, Max-pressure, SOTL, DQN, DDPG—integrated in one SUMO-based codebase with a parallel actor-learner architecture. The hyperparameter sensitivity study, with multiple seeds and boxplots, is a solid second contribution. The controllers themselves are prior work, and the literature review covers the field thoroughly.\n\nThe good parts: the code is public, the experiments use multiple random seeds, and the figures show confidence intervals. That puts it above many papers in this area. The discussion of DQN's poor low-demand behavior is honest and appropriately tentative, and the conclusion explicitly disclaims generalization beyond this setting.\n\nThe soft spots: first, the in-sample tuning issue. Section IV-A grid-searches hyperparameters on the same three-hour demand scenario that Section IV-B uses to rank the controllers. For DQN and DDPG, with many hyperparameters and stochastic training, that can overfit the scenario; Max-pressure has very few parameters, so the comparison is biased in its favor. Calling Max-pressure 'best' needs a held-out demand scenario or multiple scenarios. Second, the RL training details—learning rates, batch size, target update interval, exploration schedule—are not reported, which weakens reproducibility. Third, it's one small synthetic network; the authors say so, but that means the relative ranking is a demonstration, not a general result. Minor: no significance tests and the code isn't pinned to a commit hash.\n\nDoes this sink it? No. The framework stands as a useful common starting point. Read the comparative ranking as an illustration of the framework's use, not as a robust finding. I would cite the framework if I needed open-source baselines, but not the 'Max-pressure beats RL' conclusion.\n\nRecommendation: send it to review. The engineering contribution deserves referee time, but the comparative section needs a stronger protocol—at minimum a held-out demand pattern—before the performance claims can be trusted. The authors are not hiding their caveats.","headline":"A genuinely useful open-source framework for adaptive signal control, wrapped around a comparative experiment whose 'best controller' claim is undermined by in-sample hyperparameter tuning.","tokens_in":16391,"tokens_out":2625,"would_cite":true,"duration_ms":22662,"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":"An open-source framework for adaptive traffic signal control reports that the Max-pressure algorithm, once hyperparameters are optimized, outperforms deep Q-network and deep deterministic policy gradient controllers on a two-intersection…","keywords":["adaptive traffic signal control","Max-pressure","deep Q-network","deep deterministic policy gradient","hyperparameter optimization","SUMO","self-organizing traffic lights","travel time"],"falsifier":"Run the same framework with the optimized hyperparameters on a different network, such as a city grid or a corridor with oversaturated demand, using multiple random seeds; if a learning-based controller or Webster's produces a lower mean and median travel time than Max-pressure, the paper's central comparative claim fails.","tokens_in":15395,"feed_emoji":"🚦","tokens_out":4396,"duration_ms":37775,"temperature":0.7,"pith_summary":"The paper contributes an open-source framework that implements five adaptive traffic signal controllers (Webster's, Max-pressure, Self-Organizing Traffic Lights, deep Q-network, and deep deterministic policy gradient) for the SUMO traffic microsimulator, and demonstrates the framework by tuning and comparing the controllers on a two-intersection network. The central empirical claim is that, once every controller's hyperparameters are optimized, the Max-pressure heuristic achieves the lowest travel time, the lowest queue and delay, and the lowest variance of all tested methods, including the deep reinforcement learning controllers. The paper also claims that controllers with more hyperparameters, especially the learning-based ones, show much larger performance variance across hyperparameter settings, and that a poor hyperparameter choice can lead to erroneous conclusions. The broader contribution is a reproducible starting point intended to let researchers generate adaptive signal control results faster and with less effort.","feed_headline":"Max-pressure outperforms deep RL on traffic signal test","feed_subtitle":"An open-source comparison of five controllers finds the simplest heuristic had the lowest travel time and variance.","key_machinery":"The framework combines the SUMO traffic microsimulator with a distributed acting, centralized learning architecture for the reinforcement learning controllers, and grid-search hyperparameter optimization for all controllers. The load-bearing identity is the Max-pressure rule, which for a candidate green phase $p$ computes the pressure as the sum of queue lengths on incoming lanes minus the sum on outgoing lanes, $\\mathrm{Pressure}(p)=\\sum_{l\\in L_{p,\\mathrm{inc}}}|V_l|-\\sum_{l\\in L_{p,\\mathrm{out}}}|V_l|$, and switches to the phase with the largest pressure after a minimum green time. The other controllers include Webster's fixed-cycle timing computed from recent flow, SOTL's vehicle-time integral threshold with platoon protection, and deep neural network policies for DQN and DDPG. The comparison is carried out by optimizing each controller's hyperparameters with a grid search and then estimating performance from 32 seeded simulations.","core_discovery":"On the authors' terms, the paper establishes that in a simulated two-intersection network under a three-hour dynamic demand scenario, the Max-pressure controller is the best of the five adaptive controllers: it has the lowest mean and median travel time and the lowest standard deviation across 32 randomized simulations. DQN and DDPG perform about as well as Uniform and Webster's on average, but DQN exhibits significant outliers. The paper also establishes that hyperparameter choice dramatically changes controller performance, with learning-based controllers showing high variance, and that the DQN controller performs poorly at low demand but competitively at peak demand, a pattern the authors attribute to possible overfitting to high-reward periods.","pith_inferences":["If the Max-pressure advantage holds across more networks, cities could adopt simpler, interpretable controllers rather than deep reinforcement learning, which would ease certification, maintenance, and explanation to the public.","Because pressure-based control has theoretical throughput guarantees in certain network models, combining this framework's benchmark methodology with that theory could sharpen predictions about which network structures favour Max-pressure.","A testable extension is to retrain the DQN with a reward that does not normalize by the peak-demand magnitude, to see whether the observed low-demand underperformance disappears.","The open-source nature of the framework allows other researchers to add new controllers to the same benchmark, turning the comparison into a living leaderboard rather than a one-off result."],"forward_implications":["Researchers using the framework can obtain reproducible baseline comparisons without reimplementing controllers from scratch.","A fair comparison between traffic signal controllers requires optimizing each controller's hyperparameters rather than relying on default settings, because poor choices can drastically alter rankings.","The Max-pressure controller's low variance and strong average performance on the tested network make it a compelling candidate for deployment on small networks similar to the experiment.","Learning-based controllers, at least in this setup, do not yet outperform simple heuristics, so future work on function approximators or training algorithms is needed to close the gap.","The DQN's pattern of poor low-demand performance and strong peak-demand performance suggests a concrete direction for studying overfitting in reward-driven signal control."],"supporting_citations":[{"why":"Supplies the SUMO traffic microsimulator that evaluates all controllers.","marker":"[5]"},{"why":"Defines the Max-pressure algorithm that is the best-performing controller in the comparison.","marker":"[13]"},{"why":"Provides the self-organizing traffic lights algorithm used as one of the non-learning controllers.","marker":"[17]"},{"why":"Gives Webster's method, the basis for the adaptive Webster's controller.","marker":"[53]"},{"why":"Defines the deep Q-network algorithm used for one of the learning controllers.","marker":"[65]"},{"why":"Defines the deep deterministic policy gradient algorithm used for the other learning controller.","marker":"[69]"}],"fun_headline_variants":["Max-pressure beats deep RL in traffic signal test","Simple controller tops deep RL for traffic signals","Max-pressure wins in traffic signal comparison","Open-source benchmark: Max-pressure best on signals","Traffic signal study: heuristics outperform learning models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ranking of controllers rests on a single synthetic two-intersection network with one three-hour demand scenario; if that network is not representative of other networks or real demand patterns, the comparative conclusions may not generalize.","fun_headline_variants_meta":{"raw":{"variants":["Max-pressure beats deep RL in traffic signal test","Simple controller tops deep RL for traffic signals","Max-pressure wins in traffic signal comparison","Open-source benchmark: Max-pressure best on signals","Traffic signal study: heuristics outperform learning models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2531,"prompt_tokens":904,"completion_tokens":1627,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":1558}},"tokens_in":520,"tokens_out":1627,"duration_ms":12140,"temperature":1.0,"reasoning_tokens":1558,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:53:21.363758+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same framework with the optimized hyperparameters on a different network, such as a city grid or a corridor with oversaturated demand, using multiple random seeds; if a learning-based controller or Webster's produces a lower mean and median travel time than Max-pressure, the paper's central comparative claim fails.","supporting_citations":[{"cited_title":"Recent devel- opment and applications of SUMO - Simulation of Urban MObility,","cited_arxiv_id":null,"evidence_quote":"Supplies the SUMO traffic microsimulator that evaluates all controllers."},{"cited_title":"The max-pressure controller for arbitrary networks of signalized intersections,","cited_arxiv_id":null,"evidence_quote":"Defines the Max-pressure algorithm that is the best-performing controller in the comparison."},{"cited_title":"Traffic signal settings, road research technical paper no. 39,","cited_arxiv_id":null,"evidence_quote":"Gives Webster's method, the basis for the adaptive Webster's controller."}],"review_version":1}