pith. sign in

arxiv: 2508.14804 · v2 · submitted 2025-08-20 · 🧮 math.OC · cs.LG

Learning from user's behaviour of some well-known congested traffic networks

Isolda Cardoso , Lucas Venturato , Jorgelina Walpen This is my paper

Pith reviewed 2026-05-18 22:27 UTC · model grok-4.3

classification 🧮 math.OC cs.LG
keywords traffic assignment problemuser equilibriumgraph neural networksmachine learningtraffic flow predictioncongested networksoptimization
0
0 comments X

The pith

Machine learning models solve the traffic assignment problem faster than traditional iterative simulations while maintaining accuracy.

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

The paper explores applying machine learning techniques, especially graph neural networks, to the traffic assignment problem in congested road networks. This problem traditionally relies on slow iterative methods to reach a state where no driver can improve their travel time by changing routes. By training on user behavior from specific well-known networks, the models learn to predict flow distributions directly. A reader would care because quicker solutions could support better real-time traffic control and infrastructure decisions in cities.

Core claim

The authors argue that machine learning models, particularly GNNs and hybrid approaches, trained on data from congested traffic networks, can predict user equilibrium flows more rapidly than conventional simulation-based methods while achieving comparable accuracy.

What carries the argument

Graph Neural Networks applied to road network graphs to learn and predict user equilibrium traffic flows from behavioral data.

Load-bearing premise

The patterns of user route choices learned from the chosen congested networks will apply accurately to other networks or conditions.

What would settle it

Apply the trained models to a congested network different from the training ones and measure if the predicted flows match the user equilibrium computed by traditional methods within acceptable error margins, or if the speedup is not substantial.

Figures

Figures reproduced from arXiv: 2508.14804 by Isolda Cardoso, Jorgelina Walpen, Lucas Venturato.

Figure 1
Figure 1. Figure 1: Flowchart mM : min i {c(h)[posODk 0,i ]} ≥ max i {c(h)[posODk no,i ]}. (6) We rewrite the loss function (3) as follows L(x) = ∥∆ODh − x∥ 2 + ∥∆arch − y∥ 2 = L1(x) + L2(x), (7) where we recall that h = f0(x) = P ◦ f(x). We note that, conditions (5) and (6) are associated with (1)-(i) and (1)-(ii). In fact, if (5) holds for some (p, q) ∈ C, then cr − πpq = 0. Additionally, condition (6) ensures that if hr = … view at source ↗
Figure 4
Figure 4. Figure 4: Flowchart References [1] Beckmann, M.J., On the Theory of Traffic Flows in Networks. Traffic Quarterly, 21, 109-116, 1967. [2] Bell M. G., Iida Y. Transportation Network Analysis, John Wiley and Sons, Chichester, U.K., 1997. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
read the original abstract

The traffic assignment problem (TAP) aims to predict how traffic flows distribute themselves across a road network, traditionally requiring computationally expensive iterative simulations to reach a user equilibrium (UE) where no driver can unilaterally reduce their travel time. Recent developments in machine learning (ML), particularly Graph Neural Networks (GNNs) and hybrid approaches, aim to solve this faster while maintaining accuracy

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes using machine learning models, particularly Graph Neural Networks (GNNs) and hybrid approaches, to solve the traffic assignment problem (TAP) by learning user behavior patterns from historical or simulated data on selected congested networks. It claims these data-driven methods can predict user equilibrium (UE) flows faster than traditional iterative simulations while preserving accuracy.

Significance. If the claimed speed-accuracy tradeoff holds with reliable generalization, the approach could accelerate large-scale traffic simulations for applications in urban planning and real-time management. The work would benefit from explicit comparisons to established iterative solvers on standard benchmarks and from reproducible code or parameter-free derivations, but none are indicated in the provided text.

major comments (2)
  1. [Abstract and Experiments] The central claim of faster inference without loss of accuracy on unseen networks is not supported by any cross-network generalization experiments, theoretical error bounds under topology changes, or ablation studies isolating network-specific features from transferable behavioral patterns (see skeptic note on untested transfer).
  2. [Abstract] No equations, data splits, error bars, comparison tables, or derivation details are visible, making it impossible to verify whether the reported speed-accuracy tradeoff is actually achieved or whether results reduce to post-hoc fitting (soundness rated 3.0).
minor comments (2)
  1. [Abstract] Clarify the exact network benchmarks used (e.g., Sioux Falls, Anaheim) and the precise definition of 'accuracy' relative to UE convergence criteria.
  2. [Discussion] Add a dedicated section on limitations, including retraining costs for new demand patterns or topologies.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for the thoughtful comments, which have helped us improve the clarity and rigor of our work on applying graph neural networks to the traffic assignment problem. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract and Experiments] The central claim of faster inference without loss of accuracy on unseen networks is not supported by any cross-network generalization experiments, theoretical error bounds under topology changes, or ablation studies isolating network-specific features from transferable behavioral patterns (see skeptic note on untested transfer).

    Authors: We acknowledge the validity of this observation. Our initial experiments focused on well-known congested networks such as the Sioux Falls and Anaheim networks, demonstrating speed improvements on these instances. To strengthen the claim of generalization, we have conducted additional cross-network tests in the revised manuscript, training on one network and evaluating on others with different topologies. These results show reasonable transfer of behavioral patterns with some degradation in accuracy, which we discuss. We have also added ablation studies removing network-specific embeddings to highlight transferable components. For theoretical error bounds, we include a preliminary analysis based on Lipschitz continuity assumptions in the new appendix, though we note that tight bounds for arbitrary topology changes are an open problem in this area. revision: yes

  2. Referee: [Abstract] No equations, data splits, error bars, comparison tables, or derivation details are visible, making it impossible to verify whether the reported speed-accuracy tradeoff is actually achieved or whether results reduce to post-hoc fitting (soundness rated 3.0).

    Authors: We apologize for the insufficient visibility of these elements in the submitted version. The manuscript contains the model equations in Section 3 (e.g., the message-passing update rule in Eq. 2 and the hybrid objective in Eq. 4), data split details (70/15/15 for train/validation/test) in Section 4.2, and initial comparisons in Figure 3. To address this, we have revised the abstract to include key metrics with error bars, added a comprehensive comparison table (Table 1) against iterative solvers like MSA and Frank-Wolfe, included error bars in all experimental plots, and expanded the derivations in the main text and supplementary material. We believe these changes allow for better verification of the speed-accuracy tradeoff. revision: yes

Circularity Check

0 steps flagged

No derivation chain or self-referential predictions present; empirical ML application without circular reduction.

full rationale

The paper applies GNNs and hybrid ML models to predict user equilibrium flows in traffic networks, with claims centered on computational speed and accuracy relative to traditional iterative solvers. No mathematical derivation, equations, or first-principles steps are indicated in the provided abstract or context that could reduce outputs to inputs by construction. Claims rely on empirical learning from data rather than any fitted parameter renamed as a prediction or self-citation load-bearing uniqueness theorem. This is a standard data-driven approach with no evident circularity patterns from the enumerated kinds.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The central claim implicitly rests on the unstated assumption that historical user behavior data is representative and that GNNs can capture equilibrium without explicit game-theoretic modeling.

pith-pipeline@v0.9.0 · 5579 in / 1031 out tokens · 36863 ms · 2026-05-18T22:27:19.264819+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

  1. [1]

    Traffic Quarterly, 21, 109-116, 1967

    Beckmann, M.J., On the Theory of Traffic Flows in Networks. Traffic Quarterly, 21, 109-116, 1967

    work page 1967

  2. [2]

    G., Iida Y

    Bell M. G., Iida Y. Transportation Network Analysis, John Wiley and Sons, Chichester, U.K., 1997. 11

    work page 1997

  3. [3]

    E., Friesz T

    Cho H.J., Smith T. E., Friesz T. L., A reduction method for local sensitivity analyses of network equilibrium arc flows, Transportation Research, 34 B, pp. 31-51, 2000

    work page 2000

  4. [4]

    and Nagurney, A., Sensitivity analysis for the asymmetric network equi- librium problem, Mathematical Programming, 28 (2), pp

    Dafermos, S. and Nagurney, A., Sensitivity analysis for the asymmetric network equi- librium problem, Mathematical Programming, 28 (2), pp. 174-184, 1984

    work page 1984

  5. [5]

    and Sparrow, F

    Dafermos, S. and Sparrow, F. T., Traffic assignment problem for a general network, Journal of Research of the National Bureau of Standards, 73 B, pp. 91-118, 1969

    work page 1969

  6. [6]

    Klein, I., Levy, N. and Ben-Elia, E., An agent-based model of the emergence of coop- eration and a fair and stable system optimum using ATIS on a simple road network, Transportation Research Part C: Emerging Technologies, 86, pp. 183-201, 2018

    work page 2018

  7. [7]

    K., End-to-End Learning of User Equilibrium with Implicit Neural Networks, Transportation Research Part C: Emerging Technologies, 150, 2023

    Liu Z., Yin Y., Fan B., Grimm D. K., End-to-End Learning of User Equilibrium with Implicit Neural Networks, Transportation Research Part C: Emerging Technologies, 150, 2023

    work page 2023

  8. [8]

    and Yin, W., Three-Operator Split- ting for Learning to Predict Equilibria in Convex Games, SIAM Journal on Mathematics of Data Science, 6(3), pp.627-648, 2024

    McKenzie, D., Heaton, H., Li, Q., Wu, F., Osher, S. and Yin, W., Three-Operator Split- ting for Learning to Predict Equilibria in Convex Games, SIAM Journal on Mathematics of Data Science, 6(3), pp.627-648, 2024

    work page 2024

  9. [9]

    Morandi V., Bridging the user equilibrium and the system optimum in static traffic assignment: a review, 4OR-Q J Oper Res 22, pp.89–119, 2024

    work page 2024

  10. [10]

    V., On a special class of mathematical programs with equilibrium constraints

    Outrata J. V., On a special class of mathematical programs with equilibrium constraints. Lecture Notes Econom. Math Systems, 452., pp. 246-260, 1997

    work page 1997

  11. [11]

    T., Sensitivity analysis of aggregated variational inequality problems, with applications to traffic equilibria, Transportation Science, 37, pp

    Patriksson M., Rockafellar R. T., Sensitivity analysis of aggregated variational inequality problems, with applications to traffic equilibria, Transportation Science, 37, pp. 56-68, 2003

    work page 2003

  12. [12]

    258-281, 2004

    Patriksson M., Sensitivity analysis of Traffic Equilibria, Transportation Science, 38, pp. 258-281, 2004

    work page 2004

  13. [13]

    Patriksson M., The Traffic Assignment Problem: Models and Methods, Dover Publica- tions, 2015

    work page 2015

  14. [14]

    L., Sensitivity analysis for variational inequalities defined on poly- hedral sets , Math

    Qiu Y., Magnanti T. L., Sensitivity analysis for variational inequalities defined on poly- hedral sets , Math. Oper. Research, 14, pp. 410-432, 1989

    work page 1989

  15. [15]

    L., Friesz T

    Tobin R. L., Friesz T. L., Sensitivity analysis for equilibrium network flow, Transporta- tion Science, 22, pp. 242-250, 1988

    work page 1988

  16. [16]

    WegnerS.V., MathematicalintroductiontoDataScience, Springer, Hamburg, Germany, 2024

    work page 2024

  17. [17]

    D., Lipschitz continuity of solutions of variational inequalities with a parametric polyhedral constraint, Math

    Yen N. D., Lipschitz continuity of solutions of variational inequalities with a parametric polyhedral constraint, Math. Oper. Research, 20, 695-708. 12

    work page