A Driving Regime-Embedded Deep Learning Framework for Modeling Intra-Driver Heterogeneity in Multi-Scale Car-Following Dynamics

Jiying Yan; Junfang Tian; Shiquan Zhong; Shirui Zhou; Tao Wang; Yongfu Li

REVIEW 4 major objections 6 minor 46 references

A Driving Regime-Embedded Deep Learning Framework for Modeling Intra-Driver Heterogeneity in Multi-Scale Car-Following Dynamics

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 A hybrid GRU-LSTM model that first classifies a driver's current regime, then predicts motion, roughly halves acceleration error on freeway trajectories.

desk verdict A useful regime-embedding architecture with a plausible single-step win, undermined by the authors' own closed-loop results and a missing online regime check. read the letter →

arxiv 2506.05902 v1 pith:3K52SJPS submitted 2025-06-06 cs.LG physics.soc-ph

Shirui Zhou , Jiying Yan , Junfang Tian , Tao Wang , Yongfu Li , Shiquan Zhong This is my paper

classification cs.LGphysics.soc-ph

keywords car-followingintra-driverheterogeneitydrivingregimestrajectorysegmentationLSTMGRUtrafficoscillationsNGSIM

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

The paper sets out to show that the heterogeneity inside a single driver's behavior, the switching between accelerating, decelerating, steady following, and cruising, can be captured by making discrete driving regimes explicit inputs to a neural car-following model. It builds LSTM-DR, a two-branch architecture in which a GRU branch classifies the current regime from recent kinematics and an LSTM branch predicts the follower's acceleration, speed, and spacing from continuous kinematics plus that regime label. On an independent freeway trajectory test set, the model reports acceleration mean squared error of 0.375 against 0.835 for a plain LSTM, with smaller gains for speed and spacing, and narrower error distributions than all baselines. The authors also show platoon-level simulations in which the model reproduces stop-and-go wave propagation and oscillations. If the claim holds, a simple regime-labeling step is a low-cost way to add behavioral realism to data-driven traffic simulation and control.

What carries the argument

The load-bearing object is the one-hot driving-regime label $DR_t$ in the input tensor $X_i(t-kT) = [\Delta d, \Delta v, v, DR]$. Offline, regimes are produced by bottom-up slope-segmentation of speed profiles, Dynamic Time Warping alignment of leader-follower pairs to extract time delay and safety spacing, and an 85th-percentile threshold on time delay that separates car-following from free-flow, with a $\pm 0.5\,\mathrm{m/s^2}$ slope criterion for submodes. Online, the GRU branch reconstructs $DR_t$ from kinematics and past labels; the LSTM branch embeds it and concatenates it with continuous features, and a two-phase local-then-global training loop with three curriculum stages coordinates the two branches. The mechanism carries the argument because the regime label is the only new information the proposed model adds relative to the LSTM baseline.

What would settle it

Compute the GRU's per-step one-hot regime prediction on the held-out lane-5 trajectories and compare it with the offline DTW and threshold labels; if online accuracy is close to chance, or if the offline labels depend on future leader positions, the regime pathway cannot explain the reported gain. A complementary check is to train a plain LSTM with the same parameter count and input history on the same data and see whether its acceleration MSE approaches 0.375, which would attribute the gain to capacity rather than to regime embedding.

Watch

Extended reading notes

Core claim

The central claim is that intra-driver heterogeneity is not noise to be averaged away but structure that can be labeled, and that feeding those labels into a recurrent predictor improves accuracy. Concretely, the paper claims that LSTM-DR, a GRU branch that outputs one of six one-hot driving regimes, an embedding layer, and an LSTM branch that fuses the embedded regime with spacing, relative speed, and follower speed to predict acceleration, outperforms IDM, RNN, GRU, and plain LSTM on held-out test data. The reported numbers are $\mathrm{MSE}_a = 0.375$ versus $0.835$ for LSTM, $\mathrm{MSE}_v = 0.684$ versus $0.811$, and $\mathrm{MSE}_x = 19.25$ versus $26.05$, with data-driven baselines sharing the same layer sizes, learning rate, and batch size. The same model, run in closed loop over platoons, reproduces oscillation propagation and hysteresis more faithfully than the baselines, which the paper takes as evidence that the discrete-continuous fusion captures both microscopic and macroscopic traffic behavior.

Load-bearing premise

The framework assumes that the driving-regime labels, produced offline by a segmentation pipeline with thresholds calibrated on NGSIM data, can be reconstructed online by the GRU from past kinematics alone; the paper does not report regime-classification accuracy on the held-out lane, so the claimed single-step gain may not transfer to real-time use.

Editorial extensions

If this is right

Single-step acceleration prediction error drops from $\mathrm{MSE}_a = 0.835$ to $0.375$ relative to a plain LSTM on the held-out lane, a reduction of roughly 55 percent.
Speed and spacing errors also fall, from $0.811$ to $0.684$ and from $26.05$ to $19.25$, so the regime input improves more than the one output it directly conditions.
The error distribution tightens, with an interquartile range of 19.61 versus 22.66 for LSTM and 40.03 for IDM, indicating more stable trajectory predictions across speed conditions.
Closed-loop platoon simulations reproduce oscillation propagation and hysteresis with lower velocity error than LSTM and IDM, supporting the model's use in traffic-flow studies.
Because the model outputs acceleration before integrating to speed and position, it aligns with a control-oriented interface for automated driving.

Reading between the lines

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

Editorial inference: the offline regime labels are produced by segmenting full trajectories before training, so they may encode information from the whole trajectory; a real-time test requires labels computable causally and a report of the GRU's online classification accuracy.
Editorial inference: the regime-embedding idea should transfer to other piecewise-behavioral prediction problems, such as lane changing or eco-driving, whenever a cheap classifier can supply the discrete mode.
Editorial inference: because the 85th-percentile threshold and slope criteria are calibrated on NGSIM data, applying the framework to another site likely needs recalibration; the paper lists validation on other datasets as future work.

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.

Desk Editor's Note

read the letter

Short version: the LSTM-DR architecture is a sensible integration of bottom-up segmentation, DTW labeling, and a GRU/LSTM stack, and the single-step numbers on a held-out lane are worth taking seriously. But the paper's own closed-loop tables contradict the headline acceleration claim, and the authors do not report the one number that would tell you whether the regime module works online. I would send it to review, not desk reject, but only with major revisions.

The genuinely new part is the integration: a GRU that predicts six discrete regimes and feeds an LSTM kinematic predictor, trained in three curriculum stages and evaluated on cleaned NGSIM data. The single-step results (MSE_a 0.375 vs 0.835 for LSTM) are large and, if reproducible, practically meaningful. The boxplot comparison and the qualitative stop-and-go reproduction are pluses. This is not just retrofitting a known model.

The soft spots are real and load-bearing. First, Table I shows LSTM-DR beating LSTM on acceleration in single-step prediction, but Tables II and III, the closed-loop platoon simulations, show the reverse: mean MSE_a 0.339 vs 0.306 and 0.411 vs 0.340. That is a 10-20% degradation in the setting the paper itself says matters for real-time use. The text in Section III-B touts the IDM comparison but does not mention that LSTM beats LSTM-DR on acceleration. That is selective reporting, and it directly contradicts the "statistically significant improvements over all baselines" phrasing in Section III-A. Second, "statistically significant" appears several times, but there are no confidence intervals, multiple seeds, or significance tests anywhere. Third, the driving regimes are a thresholded quantization of acceleration slope, and the CF/FF threshold is calibrated on the same dataset; the GRU's job is to reconstruct those labels online, yet the paper reports no regime classification accuracy. If the GRU mislabels states at runtime, the single-step gain may not transfer, and Tables II-III suggest exactly that kind of erosion.

A smaller gap: the comparison set omits existing regime-aware models such as Ma and Andreasson 2007 and Xu and Laval 2020, so the claim that regimes are "seldom incorporated" is overstated.

Bottom line: this paper is for traffic modelers who want an interpretable regime-state predictor and a benchmark to beat. It has a real architectural idea and a clean held-out evaluation setup, but the internal inconsistency needs a direct explanation or a corrected claim. The authors should be asked for code, regime classification accuracy, and uncertainty estimates. I would accept it for peer review because the idea is worth shepherding, but it needs major revision.

Referee Report

4 major / 6 minor

Summary. The paper proposes LSTM-DR, a hybrid car-following model that embeds discrete driving regimes into a GRU/LSTM architecture. Regimes are obtained by a bottom-up segmentation algorithm combined with Dynamic Time Warping, with car-following/free-flow separated by an 85th-percentile time-delay threshold and operational submodes classified by a speed-slope threshold. The GRU predicts the driving regime and the LSTM predicts acceleration, from which velocity and position are integrated. Training uses a two-phase local/global curriculum, and evaluation is on NGSIM I-80 with lane 5 held out. The paper claims large single-step MSE reductions relative to IDM, RNN, GRU, and LSTM (e.g., acceleration MSE 0.375 vs 0.835 for LSTM) and closed-loop reproduction of stop-and-go oscillations. My reading is that the single-step result is plausible and the held-out lane is a real strength, but the unqualified accuracy claims are contradicted by the paper's own closed-loop acceleration results and are not supported by any uncertainty quantification or significance testing.

Significance. The idea of explicitly modeling intra-driver heterogeneity through discrete regimes is a relevant and timely extension of data-driven car-following, and the hybrid discrete-continuous architecture is a reasonable design. The paper has several concrete strengths: independent lane-5 test data, comparisons against IDM/RNN/GRU/LSTM, DTW-based regime segmentation, and platoon-level closed-loop simulations that aim to reproduce oscillation propagation. If the single-step gains survive a corrected evaluation that uses predicted regimes online and includes repeated-seed statistics, the framework could be useful for trajectory prediction and control-oriented applications. At present, however, the load-bearing claims are stronger than the evidence: the acceleration gain reverses in the closed-loop fleet simulations, and the repeated use of 'statistically significant' is unsupported by any reported test. These issues are fixable, but they require substantive revision rather than minor editing.

major comments (4)

[III-B (Tables II and III)] The closed-loop fleet results contradict the unqualified improvement claim made in Section III-A and the Abstract. In scenario a (lead vehicle 539), the mean acceleration MSE is 0.339 for LSTM-DR versus 0.306 for LSTM, i.e., 10.8% worse; in scenario b (lead vehicle 745), it is 0.411 versus 0.340, i.e., 20.9% worse. The accompanying text reports only the 46.1% reduction relative to IDM and does not mention that LSTM-DR is worse than plain LSTM on acceleration in both scenarios. This is an internal inconsistency: the single-step MSE_a advantage in Table I (0.375 vs 0.835) does not transfer to recursive multi-step prediction. Please report and discuss this reversal explicitly, and restrict the improvement claim to the single-step setting or to the metrics where the closed-loop results are actually favorable (MSE_x and MSE_v).
[III-A and throughout] The phrase 'statistically significant improvements' is used repeatedly, but no confidence intervals, repeated-seed statistics, or hypothesis tests are reported anywhere. Table I and Figure 4 appear to come from a single training run, and the text does not state how many random initializations or train/validation splits were used. The boxplot in Figure 4 reports a different quantity ('prediction loss') whose relationship to the MSE values in Table I is not defined. Please provide multiple-seed standard errors and a formal significance test (e.g., a paired bootstrap over trajectories), or remove the word 'significant' and make the claims descriptive.
[II-B (Eqs. (1)-(9)) and II-C (Eq. (12))] The regime labels used as training targets and input features are a discretization of the speed slope, and the threshold |omega_0| = 0.5 m/s^2 is effectively an acceleration threshold, i.e., the same signal the model is asked to predict; the CF/FF threshold P85_tau is also calibrated on the same NGSIM dataset used for training and testing. The paper does not report the GRU's regime-classification accuracy on the held-out lane, so it is unknown whether the Table I gains reflect causal, real-time prediction of regimes or an advantage from oracle/offline labels. Please report the GRU classification accuracy and add an ablation comparing (a) ground-truth DR as input, (b) predicted DR as input, and (c) no DR input, and clarify that no future trajectory information is used in constructing the labels available at test time.
[II-C (Eqs. (11)-(12)) and Algorithm 2] The manuscript is ambiguous about whether the DR features fed to the LSTM during inference are ground-truth offline labels, GRU predictions, or a mixture. Equation (11) includes DR_i(t-kT) in the input tensor, while Algorithm 2's local phase constructs X_local from [d_true, v_true, Delta_v_true] only and then uses the GRU output. This is load-bearing because the claimed real-time applicability depends on the GRU's predictions, not on offline labels. Please specify the exact feature vector used at every training and inference stage, including the closed-loop simulations, and confirm that the sequence of DR labels is always generated causally from past information.

minor comments (6)

[II-B, Eq. (3)] The set-update notation 'S' = S' {s_j, s_k} U {merge(s_j, s_k)}' is missing the set-difference symbol; please write S' <- (S' \ {s_j, s_k}) U {merge(s_j, s_k)}.
[II-B and References] The text attributes the 0.5 m/s^2 threshold to 'Ozaki's stability criterion' and cites [46], but reference [46] is Guan et al. (2023); the citation does not match the claimed source.
[III-A and III-B] The in-text references to 'Table 2' and 'Table 3' do not match the manuscript's Table I, Table II, and Table III; please renumber and update the cross-references.
[II-C] The symbol T is used both for the observation time step/reaction delay in Eqs. (10)-(11) and for the total trajectory duration in Eq. (2); this overloaded notation makes the definitions of N and the history window difficult to follow.
[III-A (speed regimes and Fig. 5)] The speed bins appear inverted: the text calls v > 15 m/s 'low-speed' and v < 8 m/s 'high-speed', while the Figure 5 captions say high, medium, and low-speed conditions; please correct the inconsistency.
[III-A] There is a typo in 'speeDRegimes', and the sentence 'These results demonstrate that the LSTM-DR model effectively resolves vehicular dynamics across diverse speeDRegimes' should end with a period and a full rewrite.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: the DR feature is a quantized acceleration proxy, but it is predicted by the GRU from causal kinematics rather than read from the target, so the reported gains are empirical comparisons rather than identity reductions.

full rationale

The paper's derivation chain is a standard supervised learning pipeline. Offline segmentation in Section II-B labels regimes by piecewise-linear speed slopes (Eqs. 1-3 and the |omega_0|=0.5 threshold, i.e., acceleration), and Eq. (11) includes those regimes in the predictor's input tensor. This is the strongest circularity candidate: the input contains a discretized version of the target acceleration. However, Algorithms 1 and 2 train a GRU to produce DR_t from the kinematic state and use the GRU's argmax output, not the offline label, when forming the LSTM input and computing both L_local and L_global. The model is therefore not handed the ground-truth target at inference, and the comparison against a plain LSTM is an empirical architecture comparison rather than a definitional identity. The CF/FF threshold in Eq. (9) is calibrated from NGSIM statistics, but the test set is a separate lane and the threshold is a fixed preprocessing transform, not a parameter fitted to the target. The only self-citation found is reference [26] (Tian as co-author), used as background for the importance of intra-driver heterogeneity; it is corroborated by independent references and is not load-bearing. The paper's use of 'statistically significant' without reporting significance tests is an evidentiary weakness, and the closed-loop acceleration results in Tables II-III being worse for LSTM-DR than in Table I is a consistency and generalization concern, but neither is a circularity. No load-bearing step reduces to its own input by construction.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical entities or latent mechanisms; the driving regimes are thresholds over existing kinematic variables. The main ledger items are hand-set or data-calibrated thresholds used to construct the regime labels, plus the assumption that those offline labels can be predicted causally by the GRU at runtime. These parameters and assumptions, rather than any new theoretical principle, carry most of the paper's explanatory weight.

free parameters (7)

Steady-state slope threshold |omega0|=0.5 m/s^2 = 0.5 m/s^2 (absolute value)
Used to separate steady-state from acceleration/deceleration states in Section II-B.3; hand-set with reference to Ozaki's 0.05g criterion and validated on NGSIM, but no sensitivity analysis.
CF/FF time-delay percentile P85_tau = 85th percentile of empirical tau_n distribution
Eq. (9) classifies car-following versus free flow using a quantile threshold calibrated from the same NGSIM dataset; this is a fitted data threshold.
Slope merging tolerance epsilon = 0.01
Eq. (3) merges adjacent segments when slope differences are below this value; chosen by hand and not subjected to sensitivity analysis.
Minimum segment length Lmin = 0.5 s
Eq. (2) sets the segmentation termination Nmax=floor(2T); picked to align with NGSIM's 10 Hz sampling, but no robustness check.
Historical observation window N = 10
Section II-C.1 sets N=10 based on prior studies; the exact choice affects the input tensor and prediction accuracy.
Stage transition MSE threshold epsilon = 0.05 m/s^2
Section II-C.3 switches from local to global optimization when single-step MSE drops below this threshold; it is an empirically set training hyperparameter.
Training hyperparameters = eta=1e-3, batch=128, beta1=0.95, beta2=0.9999, label smoothing=0.1, 6 layers, 16 hidden units
Set by grid search or prior work in Section II-C.2; they shape all reported results and are not benchmarked for sensitivity.

assumptions (6)

domain assumption The Montanino-Punzo reconstructed NGSIM trajectories are accurate enough to serve as ground truth.
Section II-A treats the corrected I-80 dataset as observed truth; all reported errors are relative to these reconstructed trajectories.
domain assumption Newell's simplified car-following relation and an 85th-percentile time-delay threshold separate car-following from free-flow sections.
Section II-B.2 uses Eq. (4) and Eq. (9) to define CF/FF; the threshold is empirical and dataset-specific.
ad hoc to paper Six discrete driving regimes with a fixed plus or minus 0.5 m/s^2 slope threshold are sufficient to represent intra-driver heterogeneity.
Section II-B.3 assigns every trajectory segment to one of six states using an acceleration threshold; this is the core modeling assumption of the framework.
ad hoc to paper Regime labels computed offline over complete trajectories are predictable in real time by the GRU from only historical inputs.
Section II-C.1 and Algorithm 2 rely on this causal predictability, but no regime prediction accuracy is reported.
standard math Standard backpropagation and Adam optimization converge to a useful local optimum for the hybrid architecture.
Training details in Section II-C.3 assume standard deep learning convergence behavior.
standard math DTW alignment between leader and follower trajectories extracts meaningful time delays and safety spacings.
Section II-B.2 uses DTW as the alignment tool; its optimal path is treated as the stimulus-response lag, a domain interpretation beyond the mathematical definition of DTW.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Driving Regime-Embedded Deep Learning Framework for Modeling Intra-Driver Heterogeneity in Multi-Scale Car-Following Dynamics." pith.science (2026). https://pith.science/paper/3K52SJPS

@misc{pith2026250605902,
  author       = {Pith},
  title        = {Pith review of: A Driving Regime-Embedded Deep Learning Framework for Modeling Intra-Driver Heterogeneity in Multi-Scale Car-Following Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3K52SJPS}},
  note         = {Machine review of arXiv:2506.05902}
}

read the original abstract

A fundamental challenge in car-following modeling lies in accurately representing the multi-scale complexity of driving behaviors, particularly the intra-driver heterogeneity where a single driver's actions fluctuate dynamically under varying conditions. While existing models, both conventional and data-driven, address behavioral heterogeneity to some extent, they often emphasize inter-driver heterogeneity or rely on simplified assumptions, limiting their ability to capture the dynamic heterogeneity of a single driver under different driving conditions. To address this gap, we propose a novel data-driven car-following framework that systematically embeds discrete driving regimes (e.g., steady-state following, acceleration, cruising) into vehicular motion predictions. Leveraging high-resolution traffic trajectory datasets, the proposed hybrid deep learning architecture combines Gated Recurrent Units for discrete driving regime classification with Long Short-Term Memory networks for continuous kinematic prediction, unifying discrete decision-making processes and continuous vehicular dynamics to comprehensively represent inter- and intra-driver heterogeneity. Driving regimes are identified using a bottom-up segmentation algorithm and Dynamic Time Warping, ensuring robust characterization of behavioral states across diverse traffic scenarios. Comparative analyses demonstrate that the framework significantly reduces prediction errors for acceleration (maximum MSE improvement reached 58.47\%), speed, and spacing metrics while reproducing critical traffic phenomena, such as stop-and-go wave propagation and oscillatory dynamics.

Figures

Figures reproduced from arXiv: 2506.05902 by the authors.

**Figure 2.** One Example of Driving Regimes Identification. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Hybrid Deep Learning Architecture for Car-following Modeling. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Comparative Distribution Analysis of Simulation Errors [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Single Trajectory Simulation under (a) high (b) medium [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 7.** Figure 7: Fleet Simulation of scenarios a and b. To evaluate the model’s generalization capability under traffic oscillation conditions, simulation experiments were conducted on platoons with distinct oscillation scenario. As illustrated in Fig.7, comparative spatiotemporal traj…

**Figure 6.** Figure 6: One Example of Dynamic Coupling between Speed [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 45 canonical work pages

[1]

Dynamical model of traffic congestion and numerical simulation,

M. Bando, K. Hasebe, A. Nakayama, A. Shibata, and Y . Sugiyama, “Dynamical model of traffic congestion and numerical simulation,” Physical Review E, vol. 51, no. 2, pp. 1035–1042, Feb. 1995

work page 1995
[2]

Congested Traffic States in Empirical Observations and Microscopic Simulations,

M. Treiber, A. Hennecke, and D. Helbing, “Congested Traffic States in Empirical Observations and Microscopic Simulations,”Physical Review E, vol. 62, no. 2, pp. 1805–1824, Aug. 2000

work page 2000
[3]

Full velocity difference model for a car- following theory,

R. Jiang, Q. Wu, and Z. Zhu, “Full velocity difference model for a car- following theory,”Physical Review E, vol. 64, no. 1, p. 017101, Jun. 2001

work page 2001
[4]

A simplified car-following theory: a lower order model,

G. F. Newell, “A simplified car-following theory: a lower order model,” Transportation Research Part B- Methodological, 2002

work page 2002
[5]

Trajectory data- based traffic flow studies: A revisit,

L. Li, R. Jiang, Z. He, X. M. Chen, and X. Zhou, “Trajectory data- based traffic flow studies: A revisit,”Transportation Research Part C: Emerging Technologies, vol. 114, pp. 225–240, May 2020

work page 2020
[6]

Analysis of asymmetric driving behavior using a self-learning approach,

D. Wei and H. Liu, “Analysis of asymmetric driving behavior using a self-learning approach,”Transportation Research Part B- Methodologi- cal, vol. 47, pp. 1–14, Jan. 2013

work page 2013
[7]

Towards Data-Driven Car- Following Models,

V . Papathanasopoulou and C. Antoniou, “Towards Data-Driven Car- Following Models,”Transportation Research Part C-Emerging Tech- nologies, vol. 55, no. SI, pp. 496–509, Jun. 2015. 12

work page 2015
[8]

Car-following behavior with instantaneous driver–vehicle reaction delay: A neural-network-based methodology,

J. Zheng, K. Suzuki, and M. Fujita, “Car-following behavior with instantaneous driver–vehicle reaction delay: A neural-network-based methodology,”Transportation Research Part C- Emerging Technologies, vol. 36, pp. 339–351, Nov. 2013

work page 2013

Show all 46 references

[9]

A Car-Following Model Considering Asymmetric Driving Behavior Based on Long Short-Term Memory Neu- ral Networks,

X. Huang, J. Sun, and J. Sun, “A Car-Following Model Considering Asymmetric Driving Behavior Based on Long Short-Term Memory Neu- ral Networks,”Transportation Research Part C-Emerging Technologies, vol. 95, pp. 346–362, Oct. 2018

2018
[10]

Memory, attention and prediction: a deep learning architecture for car-following,

Y . Wu, H. Tan, X. Chen, and B. Ran, “Memory, attention and prediction: a deep learning architecture for car-following,”Transportmetrica B: Transport Dynamics, vol. 7, no. 1, pp. 1553–1571, Dec. 2019

2019
[11]

Human-like Autonomous Car- Following Model with Deep Reinforcement Learning,

M. Zhu, X. Wang, and Y . Wang, “Human-like Autonomous Car- Following Model with Deep Reinforcement Learning,”Transportation Research Part C-Emerging Technologies, vol. 97, pp. 348–368, Dec. 2018

2018
[12]

A sequence to sequence learning based car- following model for multi-step predictions considering reaction delay,

L. Ma and S. Qu, “A sequence to sequence learning based car- following model for multi-step predictions considering reaction delay,” Transportation Research Part C- Emerging Technologies, vol. 120, p. 102785, Nov. 2020

2020
[13]

On the Impact of Prior Experiences in Car-Following Models: Model Development, Computational Efficiency, Comparative Analyses, and Extensive Applications,

Y . Yu, Z. He, and X. Qu, “On the Impact of Prior Experiences in Car-Following Models: Model Development, Computational Efficiency, Comparative Analyses, and Extensive Applications,”IEEE Transactions on Cybernetics, vol. 53, no. 3, pp. 1405–1418, Mar. 2023

2023
[14]

Capturing Car-Following Behaviors by Deep Learning,

X. Wang, R. Jiang, L. Li, Y . Lin, X. Zheng, and F.-Y . Wang, “Capturing Car-Following Behaviors by Deep Learning,”IEEE Transactions on Intelligent Transportation Systems, vol. 19, no. 3, pp. 910–920, 2018

2018
[15]

Incorporating human-factors in car- following models: A review of recent developments and research needs,

M. Saifuzzaman and Z. Zheng, “Incorporating human-factors in car- following models: A review of recent developments and research needs,” Transportation Research Part C- Emerging Technologies, vol. 48, pp. 379–403, Nov. 2014

2014
[16]

Hysteresis in traffic flow revisited: An improved measure- ment method,

J. A. Laval, “Hysteresis in traffic flow revisited: An improved measure- ment method,”Transportation Research Part B- Methodological, vol. 45, no. 2, pp. 385–391, Feb. 2011

2011
[17]

On the periodicity of traffic oscillations and capacity drop: The role of driver characteristics,

D. Chen, S. Ahn, J. Laval, and Z. Zheng, “On the periodicity of traffic oscillations and capacity drop: The role of driver characteristics,” Transportation Research Part B- Methodological, vol. 59, pp. 117–136, Jan. 2014

2014
[18]

A behavioral car-following model that captures traffic oscillations,

D. Chen, J. Laval, Z. Zheng, and S. Ahn, “A behavioral car-following model that captures traffic oscillations,”Transportation Research Part B- Methodological, vol. 46, no. 6, pp. 744–761, Jul. 2012

2012
[19]

Microscopic traffic hys- teresis in traffic oscillations: A behavioral perspective,

D. Chen, J. A. Laval, S. Ahn, and Z. Zheng, “Microscopic traffic hys- teresis in traffic oscillations: A behavioral perspective,”Transportation Research Part B- Methodological, vol. 46, no. 10, pp. 1440–1453, Dec. 2012

2012
[20]

Heterogeneity in car-following behav- ior: Theory and empirics,

S. Ossen and S. P. Hoogendoorn, “Heterogeneity in car-following behav- ior: Theory and empirics,”Transportation Research Part C- Emerging Technologies, vol. 19, no. 2, pp. 182–195, Apr. 2011

2011
[21]

Formalizing the heterogeneity of the vehicle-driver system to reproduce traffic oscillations,

M. Makridis, L. Leclercq, B. Ciuffo, G. Fontaras, and K. Mattas, “Formalizing the heterogeneity of the vehicle-driver system to reproduce traffic oscillations,”Transportation Research Part C-Emerging Technolo- gies, vol. 120, p. 102803, Nov. 2020

2020
[22]

Investigating the long- and short-term driving characteristics and incorporating them into car-following models,

X. Chen, J. Sun, Z. Ma, J. Sun, and Z. Zheng, “Investigating the long- and short-term driving characteristics and incorporating them into car-following models,”Transportation Research Part C- Emerging Technologies, vol. 117, p. 102698, Aug. 2020

2020
[23]

Multi-anticipation and heterogeneity in car-following empirics and a first exploration of their implications,

S. Ossen and S. Hoogendoorn, “Multi-anticipation and heterogeneity in car-following empirics and a first exploration of their implications,” in 2006 IEEE Intelligent Transportation Systems Conference. Toronto, Ont.: IEEE, 2006, pp. 1615–1620

2006
[24]

Analyzing fluctuations in car-following,

P. Wagner, “Analyzing fluctuations in car-following,”Transportation Research Part B- Methodological, vol. 46, no. 10, pp. 1384–1392, Dec. 2012

2012
[25]

A mechanism to describe the formation and propagation of stop-and-go waves in congested freeway traffic,

J. A. Laval and L. Leclercq, “A mechanism to describe the formation and propagation of stop-and-go waves in congested freeway traffic,”Philo- sophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, vol. 368, no. 1928, pp. 4519–4541, Oct. 2010

1928
[26]

Experimental study and modeling of car-following behavior under high speed situation,

Y .-X. Huang, R. Jiang, H. Zhang, M.-B. Hu, J.-F. Tian, B. Jia, and Z.- Y . Gao, “Experimental study and modeling of car-following behavior under high speed situation,”Transportation Research Part C-Emerging Technologies, vol. 97, pp. 194–215, Dec. 2018

2018
[27]

Driving Style Analysis Using Primitive Driving Patterns With Bayesian Nonparametric Approaches,

W. Wang, J. Xi, and D. Zhao, “Driving Style Analysis Using Primitive Driving Patterns With Bayesian Nonparametric Approaches,”IEEE Transactions on Intelligent Transportation Systems, vol. 20, no. 8, pp. 2986–2998, Aug. 2019

2019
[28]

Method for investigating intradriver heterogeneity using vehi- cle trajectory data: A Dynamic Time Warping approach,

J. Taylor, “Method for investigating intradriver heterogeneity using vehi- cle trajectory data: A Dynamic Time Warping approach,”Transportation Research Part B- Methodological, 2015

2015
[29]

Treiber and A

M. Treiber and A. Kesting,Traffic Flow Dynamics. Berlin, Heidelberg: Springer Berlin Heidelberg, 2013

2013
[30]

Is more always better? The impact of vehicular trajectory completeness on car-following model calibration and validation,

A. Sharma, “Is more always better? The impact of vehicular trajectory completeness on car-following model calibration and validation,”Trans- portation Research Part B- Methodological, 2019

2019
[31]

A pattern recognition algorithm for assessing trajectory completeness,

A. Sharma, Z. Zheng, and A. Bhaskar, “A pattern recognition algorithm for assessing trajectory completeness,”Transportation Research Part C- Emerging Technologies, vol. 96, pp. 432–457, Nov. 2018

2018
[32]

Behavior Measurement, Analysis, and Regime Classification in Car Following,

X. Ma and I. Andreasson, “Behavior Measurement, Analysis, and Regime Classification in Car Following,”IEEE Transactions on Intelli- gent Transportation Systems, vol. 8, no. 1, pp. 144–156, Mar. 2007

2007
[33]

Statistical Analysis of Driver Behavior Data in Different Regimes of the Car-Following Stage,

X. Ma and I. Andr ´easson, “Statistical Analysis of Driver Behavior Data in Different Regimes of the Car-Following Stage,”Transportation Research Record: Journal of the Transportation Research Board, vol. 2018, no. 1, pp. 87–96, Jan. 2007

2018
[34]

Statistical inference for two-regime stochastic car- following models,

T. Xu and J. Laval, “Statistical inference for two-regime stochastic car- following models,”Transportation Research Part B: Methodological, vol. 134, pp. 210–228, 2020

2020
[35]

A Recurrent Neural Network Based Microscopic Car Following Model to Predict Traffic Oscillation,

M. Zhou, X. Qu, and X. Li, “A Recurrent Neural Network Based Microscopic Car Following Model to Predict Traffic Oscillation,”Trans- portation Research Part C-Emerging Technologies, vol. 84, pp. 245–264, Nov. 2017

2017
[36]

Learning Car-Following Behaviors for a Connected Automated Vehicle System: An Improved Sequence-to-Sequence Deep Learning Model,

W. Lu, Z. Yi, B. Liang, Y . Rui, and B. Ran, “Learning Car-Following Behaviors for a Connected Automated Vehicle System: An Improved Sequence-to-Sequence Deep Learning Model,”IEEE Access, vol. 11, pp. 28 076–28 089, 2023

2023
[37]

A Sequence-to- Sequence Car-Following Model for Addressing Driver Reaction Delay and Cumulative Error in Multi-Step Prediction,

N. Xu, C. Chen, Y . Zhang, J. Wang, Q. Liu, and C. Guo, “A Sequence-to- Sequence Car-Following Model for Addressing Driver Reaction Delay and Cumulative Error in Multi-Step Prediction,”IEEE Transactions on Intelligent Transportation Systems, pp. 1–13, 2024

2024
[38]

On the assessment of vehicle trajectory data accuracy and application to the Next Generation SIMulation (NGSIM) program data,

V . Punzo, M. T. Borzacchiello, and B. Ciuffo, “On the assessment of vehicle trajectory data accuracy and application to the Next Generation SIMulation (NGSIM) program data,”Transportation Research Part C- Emerging Technologies, vol. 19, no. 6, pp. 1243–1262, Dec. 2011

2011
[39]

Trajectory data reconstruction and simulation-based validation against macroscopic traffic patterns,

M. Montanino and V . Punzo, “Trajectory data reconstruction and simulation-based validation against macroscopic traffic patterns,”Trans- portation Research Part B- Methodological, vol. 80, pp. 82–106, Oct. 2015

2015
[40]

SEGMENTING TIME SERIES: A SURVEY AND NOVEL APPROACH,

E. Keogh, S. Chu, D. Hart, and M. Pazzani, “SEGMENTING TIME SERIES: A SURVEY AND NOVEL APPROACH,” inSeries in Machine Perception and Artificial Intelligence. WORLD SCIENTIFIC, Jun. 2004, vol. 57, pp. 1–21

2004
[41]

Statistical Test for 85th and 15th Percentile Speeds with Asymptotic Distribution of Sample Quantiles,

Y . Hou, C. Sun, and P. Edara, “Statistical Test for 85th and 15th Percentile Speeds with Asymptotic Distribution of Sample Quantiles,” Transportation Research Record: Journal of the Transportation Research Board, vol. 2279, no. 1, pp. 47–53, Jan. 2012

2012
[42]

Long memory is important: A test study on deep-learning based car-following model,

X. Wang, R. Jiang, L. Li, Y .-L. Lin, and F.-Y . Wang, “Long memory is important: A test study on deep-learning based car-following model,” Physica A: Statistical Mechanics and its Applications, vol. 514, pp. 786– 795, Jan. 2019

2019
[43]

Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling,

J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling,” Dec. 2014, arXiv:1412.3555 [cs]

2014 arXiv
[44]

Analyses of the stability and wave properties of a new con- tinuum traffic theory,

H. Zhang, “Analyses of the stability and wave properties of a new con- tinuum traffic theory,”Transportation Research Part B: Methodological, vol. 33, no. 6, pp. 399–415, Aug. 1999

1999
[45]

Learning-Based Adaptive Opti- mal Control for Connected Vehicles in Mixed Traffic: Robustness to Driver Reaction Time,

M. Huang, Z.-P. Jiang, and K. Ozbay, “Learning-Based Adaptive Opti- mal Control for Connected Vehicles in Mixed Traffic: Robustness to Driver Reaction Time,”IEEE Transactions on Cybernetics, vol. 52, no. 6, pp. 5267–5277, Jun. 2022

2022
[46]

Integrated Decision and Control: Toward Interpretable and Computationally Efficient Driving Intelligence,

Y . Guan, Y . Ren, Q. Sun, S. E. Li, H. Ma, J. Duan, Y . Dai, and B. Cheng, “Integrated Decision and Control: Toward Interpretable and Computationally Efficient Driving Intelligence,”IEEE Transactions on Cybernetics, vol. 53, no. 2, pp. 859–873, Feb. 2023

2023

Pith tools

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

[1] [1]

Dynamical model of traffic congestion and numerical simulation,

M. Bando, K. Hasebe, A. Nakayama, A. Shibata, and Y . Sugiyama, “Dynamical model of traffic congestion and numerical simulation,” Physical Review E, vol. 51, no. 2, pp. 1035–1042, Feb. 1995

work page 1995

[2] [2]

Congested Traffic States in Empirical Observations and Microscopic Simulations,

M. Treiber, A. Hennecke, and D. Helbing, “Congested Traffic States in Empirical Observations and Microscopic Simulations,”Physical Review E, vol. 62, no. 2, pp. 1805–1824, Aug. 2000

work page 2000

[3] [3]

Full velocity difference model for a car- following theory,

R. Jiang, Q. Wu, and Z. Zhu, “Full velocity difference model for a car- following theory,”Physical Review E, vol. 64, no. 1, p. 017101, Jun. 2001

work page 2001

[4] [4]

A simplified car-following theory: a lower order model,

G. F. Newell, “A simplified car-following theory: a lower order model,” Transportation Research Part B- Methodological, 2002

work page 2002

[5] [5]

Trajectory data- based traffic flow studies: A revisit,

L. Li, R. Jiang, Z. He, X. M. Chen, and X. Zhou, “Trajectory data- based traffic flow studies: A revisit,”Transportation Research Part C: Emerging Technologies, vol. 114, pp. 225–240, May 2020

work page 2020

[6] [6]

Analysis of asymmetric driving behavior using a self-learning approach,

D. Wei and H. Liu, “Analysis of asymmetric driving behavior using a self-learning approach,”Transportation Research Part B- Methodologi- cal, vol. 47, pp. 1–14, Jan. 2013

work page 2013

[7] [7]

Towards Data-Driven Car- Following Models,

V . Papathanasopoulou and C. Antoniou, “Towards Data-Driven Car- Following Models,”Transportation Research Part C-Emerging Tech- nologies, vol. 55, no. SI, pp. 496–509, Jun. 2015. 12

work page 2015

[8] [8]

Car-following behavior with instantaneous driver–vehicle reaction delay: A neural-network-based methodology,

J. Zheng, K. Suzuki, and M. Fujita, “Car-following behavior with instantaneous driver–vehicle reaction delay: A neural-network-based methodology,”Transportation Research Part C- Emerging Technologies, vol. 36, pp. 339–351, Nov. 2013

work page 2013