Pith. sign in

REVIEW 4 major objections 6 minor 19 references

Machine Learning Predictions for Traffic Equilibria in Road Renovation Scheduling

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An XGBoost surrogate predicts total travel time from road closures to 11% MAPE, enough to screen maintenance schedules before traffic simulation.

desk verdict Solid empirical comparison of surrogate models for traffic equilibrium prediction, but the headline XGBoost margin rests on an underspecified feature-selection procedure and inconsistent reported numbers. read the letter →

arxiv 2506.05933 v1 pith:OZT5GJ2O submitted 2025-06-06 cs.LG

Robbert Bosch , Wouter van Heeswijk , Patricia Rogetzer , Martijn Mes This is my paper
classification cs.LG
keywords surrogatemodelingtrafficassignmentproblemroadmaintenanceschedulingtotaltraveltimepredictionXGBoostonlinelearningfeatureengineeringSiouxFallsnetwork
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Long-term road maintenance planning needs to know how much simultaneous road closures will clog the network, but each traffic-equilibrium simulation is expensive and evaluating thousands of schedules is impractical. This paper tries to show that a machine-learning surrogate can estimate the total travel time of a closure scenario accurately and cheaply enough to prune poor schedules before simulation. Using the Sioux Falls network and an online protocol in which training data grows by 1,000 scenarios per batch, the authors compare linear, ensemble, probabilistic, and neural regressors against three heuristics. They report that the Costliest Subset Heuristic is a strong conservative baseline under scarce data, while XGBoost, trained on one-hot encodings, 12 engineered features, and the heuristic output, reaches an 11% mean absolute percentage error and the lowest pinball loss among all models. If that holds, the surrogate can reduce the number of traffic assignments required in maintenance scheduling optimization.

What carries the argument

The argument is carried by two mechanisms working together. The Costliest Subset Heuristic (CSH) searches the set of already-simulated configurations for the most expensive subset of a candidate closure set and returns its observed total travel time, giving a cheap, conservative lower-bound estimate that also becomes an input feature to the regressors. The second mechanism is feature construction: one-hot encodings of which roads close, pairwise indicators, and 12 engineered features selected by forward and backward sequential selection on validation $R^2$ are combined into the training representation. The claim that XGBoost works better than the alternatives rests on this feature set plus online retraining every 1,000-scenario batch, with pinball loss at quantile $\tau = 0.05$ used to enforce underestimation.

What would settle it

A concrete audit is to rerun the online experiment while recomputing feature selection on each batch's training data only, then compare XGBoost's MAPE and pinball loss with the reported values and with the Costliest Subset Heuristic. If the XGBoost advantage shrinks or vanishes, the original result depended on seeing future batches; a simpler check is whether any of the 12 selected features are statistics computed over all 200,000 scenarios, which would directly encode target information.

Watch

Extended reading notes

Core claim

The paper's central discovery is a model ranking: most regressors cannot beat the Costliest Subset Heuristic, but gradient-boosted trees in the form of XGBoost predict network-wide total travel time with markedly better accuracy and conservative bias. The paper reports an 11% MAPE for XGBoost and a pinball loss at the 0.05 quantile lower than every other completed model, with the heuristic as the best baseline when training data is scarce. The authors conclude that a surrogate trained on a combination of one-hot closure indicators, selected engineered features, and heuristic output is accurate and data-efficient enough to support simulation-free screening of road renovation schedules.

Load-bearing premise

The load-bearing premise is that the 12 engineered features were chosen using only data available before each evaluation batch; if the selection looked at all 200,000 scenarios, the reported accuracy on unseen batches would be inflated because the features would encode information about those very scenarios.

Editorial extensions

If this is right

  • Planners can use XGBoost to screen candidate maintenance schedules, simulating only the ones the surrogate does not flag as excessively congested.
  • The Costliest Subset Heuristic supplies a strong default prediction in the early iterations of optimization, when too few simulations have been run to train a reliable regressor.
  • Because XGBoost trains in about 39 seconds per iteration on this benchmark, it fits the online retraining loop, whereas neural networks, random forests, and quantile regression exceed the 10-minute iteration limit.
  • The 0.05 pinball loss makes XGBoost's predictions deliberately conservative, so a promising schedule is less likely to be discarded because of an overestimate; the trade-off is that some bad schedules survive screening and still get simulated.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next step, which the paper motivates but does not run, is to embed the XGBoost surrogate in a full renovation scheduler and measure how many traffic assignments are actually skipped before solution quality degrades; that would turn the accuracy gains into an end-to-end runtime win.
  • The same surrogate pipeline should transfer to other fixed networks, since the feature set is built from per-link baseline attributes rather than schedule-specific data; a quick benchmark on a second network would test that generality.
  • The paper's monotonicity check (200,000 scenarios with no Braess-paradox reversal) is specific to Sioux Falls; on a network where closing a road can reduce total travel time, the Costliest Subset Heuristic would lose its lower-bound interpretation and XGBoost's relative advantage could change.
  • The quantile level $\tau=0.05$ is a modeling choice; tuning $\tau$ to the decision context could trade off the cost of simulating bad schedules against the cost of accidentally discarding good ones.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper investigates machine-learning surrogate models for predicting the total travel time (TTT) of a road network under renovation scenarios, using the Sioux Falls network as a case study. The targets are generated by the Frank-Wolfe algorithm for the traffic assignment problem, and the authors compare nine regression models and three heuristics under a simulated online evaluation in which data arrives in batches of 1000 scenarios. The principal empirical claim is that XGBoost, trained on a combination of one-hot encodings, engineered features, and the Costliest Subset Heuristic (CSH) output, substantially outperforms all alternatives, achieving a MAPE of 11% and large relative improvements in pinball loss and MAPE.

Significance. If the empirical comparison is leakage-free and the reported margins are correct, the paper offers a practically useful surrogate for the Road Network Maintenance Scheduling Problem, potentially reducing the number of expensive traffic assignment simulations. The work has clear strengths: the regression targets come from an external solver, so there is no derivation-from-input circularity; CSH is a sensible and interpretable baseline; and the online evaluation protocol is, in principle, well suited to the intended application. However, the headline numerical claim is internally inconsistent, and the evaluation may be affected by target leakage in feature selection and by post hoc outlier removal, so the magnitude of XGBoost's advantage is not yet established.

major comments (4)
  1. [Abstract, §5.3, Table 5, §6] The central numerical claim is internally inconsistent. The abstract states that XGBoost achieves a MAPE of 11%, while Table 5 reports a MAPE of 15% for XGBoost, and the conclusion repeats the 11% figure. The relative improvement over the next-best model is given as 39% for MAPE and 20% for pinball loss in Section 5.3, but as 39% and 19% in the conclusion, and as 20% and 38% in the abstract. Since the size of XGBoost's advantage is the headline result, all these numbers must be reconciled and reported consistently with respect to Table 5.
  2. [§4.2, §5.1, §5.3] The feature selection procedure is a potential source of label leakage. Section 4.2 selects a fixed set of 12 features using forward and backward sequential feature selection that maximizes a validation R2, but the manuscript never states that this selection was restricted to data available before each online evaluation batch. Section 5.1 describes a single experiment with 200,000 datapoints, and if the feature selection used the full dataset, the 'unseen' batches in Section 5.3 are not truly unseen for any model using those selected features. This would directly inflate the reported performance of XGBoost and every other model that uses the selected features. The authors must either clarify that feature selection used only the first training batch or a time-aware validation split, or re-run the evaluation under that restriction.
  3. [§5.3, Table 5 footnote] The footnote to Table 5 states that for LogBaggingRidge, LogBayesianRidge, and LogOLS, 'outliers during testing were removed for averages.' No objective criterion for identifying these outliers is given, and the removal is applied after seeing the results and only to some models. This makes the averaged metrics non-comparable across models and can artificially widen XGBoost's margin. The authors should define the outlier-removal rule a priori and apply it identically to all models, or report results both with and without outlier removal.
  4. [§4.3, §4.4, §5.1] The CSH output is used as a regression feature, but the definition of the known-dataset D in Equation (6) is not tied to the temporal split of the online protocol. If D includes configurations from the test batch or from scenarios generated after the current training cutoff, the CSH feature leaks information about the target. The manuscript should state explicitly that, in the online evaluation, D contains only configurations that were accepted into the training set before the test batch was evaluated.
minor comments (6)
  1. [Title page] The affiliation contains a typo: 'Entrepeneurship' should be 'Entrepreneurship'.
  2. [§4.2] The first sentence, 'To improve reduce redundancy,' is ungrammatical; it should read 'To reduce redundancy.'
  3. [§5.2, Table 5] The heuristic introduced as 'Costliest Additive Subset (CASH)' in Section 4.3 is referred to as 'AdditiveSubset' in Section 5.2 and Table 5; the naming should be standardized.
  4. [§5.3, Table 5] The text refers to the third heuristic as 'CostliestSuperset', while Section 4.3 and Table 5 call it 'CheapestSuperset' (CSupH); the terminology should be aligned.
  5. [Figure 2] The caption uses 'q=0.05' while the text and Table 4 use \(\tau=0.05\); use a single consistent notation for the quantile parameter.
  6. [§4.1, §4.4] The pairwise encoding described in Section 4.1 is not mentioned in the final feature set used in Section 4.4; the authors should clarify whether pairwise features were included in the reported experiments or only in preliminary ones.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: targets come from an external Frank-Wolfe solver and held-out batches are compared against those simulation outputs.

full rationale

The TTT targets are produced by a Frank-Wolfe traffic assignment solver (Section 3.2, Algorithm 1), an external computation independent of the regression models. Surrogates are trained on these simulation outputs and evaluated on sequential unseen batches (Section 5.1), so the central comparison is against externally generated ground truth, not against the models' own inputs. The CSH heuristic feature, Equation (6), exactly equals the target when a configuration is already in the known set D, so every training point can carry a feature that coincides with its target. However, at evaluation time the next batch is not in D, so the CSH feature for test points is a genuine estimate based on previously seen configurations; it does not encode the test target. This is a training-time feature choice, not a definitional equivalence between prediction and target. The Section 4.2 feature-selection step uses validation R-squared without stating whether the split is temporally restricted; if the full 200,000-scenario dataset informed the chosen features, that would be an evaluation-leakage risk, but leakage is an experimental-design concern rather than circular derivation. There are no load-bearing self-citations or imported uniqueness theorems. Thus the derivation chain is self-contained and the score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the correctness of the traffic simulation labels, on the feature selection and online split being leakage-free, and on the heuristic monotonicity observation. The models introduce no new physical entities; the free parameters are standard ML hyperparameters and the conservatism quantile. The main unverified load-bearing choices are the BPR coefficients, the unstated Frank-Wolfe convergence tolerance, and the timing of feature selection relative to the online data boundary.

free parameters (3)
  • XGBoost hyperparameters = Not reported.
    Learning rate, tree depth, number of estimators, and regularization are not stated in the paper; these choices are fitted implicitly or selected without disclosure and directly affect the reported MAPE and pinball loss gains.
  • Pinball quantile tau = 0.05.
    Chosen to define conservative predictions; all pinball loss comparisons depend on this arbitrarily selected quantile.
  • Feature selection cut-off = 12 features.
    The cut-off where marginal validation R2 improvement 'diminishes significantly' is a hand-chosen stopping rule, and the selected feature set is fitted to the data.
assumptions (5)
  • domain assumption BPR link cost function with a=0.15, b=4.
    Used in Eq. 3 to define travel costs; coefficients are taken from standard practice, not estimated for Sioux Falls.
  • domain assumption User equilibrium via Frank-Wolfe has converged for all 200,000 scenarios.
    The regression targets y=TTT(A') are generated by Algorithm 1, but no convergence tolerance or verification is reported, so noisy equilibria would bias the labels.
  • ad hoc to paper Monotonicity of TTT under road closures (no Braess cases in the 200,000 dataset).
    Stated in Section 4.3 to justify the Costliest Subset Heuristic; it is an empirical observation on the generated dataset, not a theorem, and may not hold on other networks.
  • ad hoc to paper The 12 engineered features are selected without looking at future online evaluation batches.
    The paper does not state that feature selection is restricted to data available at each iteration; if full-data selection was used, the online test is leaked.
  • domain assumption Scenarios are drawn from the same distribution in training and test batches and are 'unseen' with no duplication.
    Section 5.1 assumes the next batch is unseen, but no scenario sampling distribution or deduplication rule is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Learning Predictions for Traffic Equilibria in Road Renovation Scheduling." pith.science (2026). https://pith.science/paper/OZT5GJ2O

@misc{pith2026250605933,
  author       = {Pith},
  title        = {Pith review of: Machine Learning Predictions for Traffic Equilibria in Road Renovation Scheduling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZT5GJ2O}},
  note         = {Machine review of arXiv:2506.05933}
}
read the original abstract

Accurately estimating the impact of road maintenance schedules on traffic conditions is important because maintenance operations can substantially worsen congestion if not carefully planned. Reliable estimates allow planners to avoid excessive delays during periods of roadwork. Since the exact increase in congestion is difficult to predict analytically, traffic simulations are commonly used to assess the redistribution of the flow of traffic. However, when applied to long-term maintenance planning involving many overlapping projects and scheduling alternatives, these simulations must be run thousands of times, resulting in a significant computational burden. This paper investigates the use of machine learning-based surrogate models to predict network-wide congestion caused by simultaneous road renovations. We frame the problem as a supervised learning task, using one-hot encodings, engineered traffic features, and heuristic approximations. A range of linear, ensemble-based, probabilistic, and neural regression models is evaluated under an online learning framework in which data progressively becomes available. The experimental results show that the Costliest Subset Heuristic provides a reasonable approximation when limited training data is available, and that most regression models fail to outperform it, with the exception of XGBoost, which achieves substantially better accuracy. In overall performance, XGBoost significantly outperforms alternatives in a range of metrics, most strikingly Mean Absolute Percentage Error (MAPE) and Pinball loss, where it achieves a MAPE of 11% and outperforms the next-best model by 20% and 38% respectively. This modeling approach has the potential to reduce the computational burden of large-scale traffic assignment problems in maintenance planning.

Figures

Figures reproduced from arXiv: 2506.05933 by the authors.

Figure 1
Figure 1. Illustration of the Sioux Falls traffic network. Left shows the graph of the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Pinball loss of regression models over iterations. Some models had outliers [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages

  1. [1]

    Gazi University Journal of Science 34(3), 710– 716 (Sep 2021)

    Aksoy, I.C., Mutlu, M.M., Alver, Y.: Urban Road Network Maintenance Scheduling Using Ant Colony Optimization. Gazi University Journal of Science 34(3), 710– 716 (Sep 2021). https://doi.org/10.35378/gujs.789519, number: 3 Publisher: Gazi University

    work page doi:10.35378/gujs.789519 2021

  2. [2]

    Computer- Aided Civil and Infrastructure Engineering 33(10), 833–848 (2018)

    Bagloee, S.A., Sarvi, M., Patriksson, M., Asadi, M.: Optimization for Roads’ Construction: Selection, Prioritization, and Scheduling. Computer- Aided Civil and Infrastructure Engineering 33(10), 833–848 (2018). https://doi.org/10.1111/mice.12370

    work page doi:10.1111/mice.12370 2018

  3. [3]

    Yale University Press (1956), https://trid.trb.org/View/91120, number: 226 pp Machine Learning Predictions for Traffic Equilibria 15

    Beckmann, M., McGuire, C.B., Winsten, C.B.: Studies in the Economics of Trans- portation. Yale University Press (1956), https://trid.trb.org/View/91120, number: 226 pp Machine Learning Predictions for Traffic Equilibria 15

    work page 1956

  4. [4]

    Transportation Science 39(4), 446–450 (2005), https://www.jstor.org/stable/25769266, publisher: INFORMS

    Braess, D., Nagurney, A., Wakolbinger, T.: On a Paradox of Traffic Planning. Transportation Science 39(4), 446–450 (2005), https://www.jstor.org/stable/25769266, publisher: INFORMS

    work page arXiv 2005

  5. [5]

    Transportation Research Part C: Emerging Technologies 167, 104838 (Oct 2024)

    Chen, Y., Zheng, L., Tan, Z.: Roadside LiDAR placement for cooperative traffic de- tection by a novel chance constrained stochastic simulation optimization approach. Transportation Research Part C: Emerging Technologies 167, 104838 (Oct 2024). https://doi.org/10.1016/j.trc.2024.104838

    work page arXiv 2024

  6. [6]

    Computer-Aided Civil and Infrastructure En- gineering 19(6), 446–455 (2004)

    Cheu, R.L., Wang, Y., Fwa, T.F.: Genetic Algorithm-Simulation Methodology for Pavement Maintenance Scheduling. Computer-Aided Civil and Infrastructure En- gineering 19(6), 446–455 (2004). https://doi.org/10.1111/j.1467-8667.2004.00369.x

    work page arXiv 2004

  7. [7]

    International Journal of Transportation Science and Technology 5(1), 17–27 (Aug 2016)

    Gong, L., Fan, W.: Optimizing scheduling of long-term highway work zone projects. International Journal of Transportation Science and Technology 5(1), 17–27 (Aug 2016). https://doi.org/10.1016/j.ijtst.2016.06.003

    work page doi:10.1016/j.ijtst.2016.06.003 2016

  8. [8]

    Computer-Aided Civil and Infrastructure Engineering 34(8), 638–653 (Aug 2019)

    Gu, Z., Waller, S.T., Saberi, M.: Surrogate-based toll optimization in a large- scale heterogeneously congested network. Computer-Aided Civil and Infrastructure Engineering 34(8), 638–653 (Aug 2019). https://doi.org/10.1111/mice.12444

    work page doi:10.1111/mice.12444 2019

Show all 19 references
  1. [9]

    Transportmetrica A: Transport Science 11(1), 74–101 (Jan 2015)

    Jiang, Y., Szeto, W.: Time-dependent transportation network design that considers health cost. Transportmetrica A: Transport Science 11(1), 74–101 (Jan 2015). https://doi.org/10.1080/23249935.2014.927938

    2015

  2. [10]

    Transportmetrica A: Transport Science 14(4), 346–371 (Apr 2018)

    Kumar, A., Mishra, S.: A simplified framework for sequencing of transportation projects considering user costs and benefits. Transportmetrica A: Transport Science 14(4), 346–371 (Apr 2018). https://doi.org/10.1080/23249935.2017.1387827

    2018

  3. [11]

    Transportation Research 9(5), 309–318 (Oct 1975)

    LeBlanc, L.J., Morlok, E.K., Pierskalla, W.P.: An efficient approach to solving the road network equilibrium traffic assignment problem. Transportation Research 9(5), 309–318 (Oct 1975). https://doi.org/10.1016/0041-1647(75)90030-1

    1975 doi

  4. [12]

    Smart and Resilient Transportation3(2), 118–130 (Jan 2021)

    Li, Y., Fan, W.: Bi-level optimization of long-term highway work zone scheduling considering elastic demand. Smart and Resilient Transportation3(2), 118–130 (Jan 2021). https://doi.org/10.1108/SRT-01-2021-0004

    2021 doi

  5. [13]

    PLOS ONE 11(10), e0164780 (Oct 2016)

    Lu, G., Xiong, Y., Ding, C., Wang, Y.: An Optimal Schedule for Urban Road Network Repair Based on the Greedy Algorithm. PLOS ONE 11(10), e0164780 (Oct 2016). https://doi.org/10.1371/journal.pone.0164780

    2016 doi

  6. [14]

    Computer-Aided Civil and Infrastructure Engineering 37(4), 427–450 (2022)

    Miralinaghi, M., Davatgari, A., Seilabi, S.E., Labi, S.: Contract bundling consider- ations in urban road project scheduling. Computer-Aided Civil and Infrastructure Engineering 37(4), 427–450 (2022). https://doi.org/10.1111/mice.12740

    2022 doi

  7. [15]

    https://doi.org/10.48550/arXiv.2501.11057

    Natterer, E., Engelhardt, R., H¨ orl, S., Bogenberger, K.: Machine Learning Surro- gates for Optimizing Transportation Policies with Agent-Based Models (Feb 2025). https://doi.org/10.48550/arXiv.2501.11057

    2025 doi arXiv

  8. [16]

    Transportation Science 8(3), 203–216 (1974), https://www.jstor.org/stable/25767747, publisher: INFORMS

    Nguyen, S.: An Algorithm for the Traffic Assignment Problem. Transportation Science 8(3), 203–216 (1974), https://www.jstor.org/stable/25767747, publisher: INFORMS

    1974

  9. [17]

    Roads, U.S.B.o.P.: Traffic Assignment Manual for Application with a Large, High Speed Computer. U.S. Department of Commerce, Bureau of Public Roads, Office of Planning, Urban Planning Division (1964), google-Books-ID: AvNUR O JEcC

    1964

  10. [18]

    Song, Z., He, Y., Liu, Z.: Rehabilitation Project Selection and Scheduling in Trans- portation Networks. Tech. Rep. MPC-18-358, Mountain Plains Consortium (Dec 2018), https://rosap.ntl.bts.gov/view/dot/42564

    2018

  11. [19]

    IEEE Trans- actions on Intelligent Transportation Systems 15(6), 2595–2604 (Dec 2014)

    Zheng, H., Nava, E., Chiu, Y.C.: Measuring Networkwide Traffic Delay in Sched- ule Optimization for Work-Zone Planning in Urban Networks. IEEE Trans- actions on Intelligent Transportation Systems 15(6), 2595–2604 (Dec 2014). https://doi.org/10.1109/TITS.2014.2318299

    2014

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.