Pith. sign in

REVIEW 3 major objections 5 minor 27 references

Semi-on-Demand Transit Feeders with Shared Autonomous Vehicles and Reinforcement-Learning-Based Zonal Dispatching Control

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Zonal RL dispatch lets semi-on-demand bus feeders carry 16% more riders

desk verdict Useful RL dispatch study, but the headline 16% passenger gain is partly an artifact of an uncalibrated linear demand model; the RL-vs-nominal comparison is cleaner but statistically untested. read the letter →

arxiv 2509.01883 v1 pith:NUMZMWIZ submitted 2025-09-02 cs.LG cs.SYeess.SYmath.OC

Max T.M. Ng , Roman Engelhardt , Florian Dandl , Hani S. Mahmassani , Klaus Bogenberger This is my paper · ORCID
classification cs.LGcs.SYeess.SYmath.OC
keywords sharedautonomousvehiclessemi-on-demandtransitreinforcementlearningzonaldispatchingfeederserviceProximalPolicyOptimizationdemand-responsivetransportfirst-mile-last-mile
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

This paper claims that a transit feeder service combining a fixed-route segment with an on-demand flexible zone can beat a traditional fixed-route bus on passenger throughput when a reinforcement-learning controller decides which flexible zone each shared autonomous vehicle should serve. In simulations of a real Munich bus route during off-peak hours, the semi-on-demand service with RL-based zonal dispatching serves 16% more passengers than the fixed-route baseline, at 13% higher generalized cost (user time plus operator cost). The RL controller itself is responsible for a 2.4% passenger gain over a simple zonal assignment rule, at only 1.4% higher cost. The paper frames this as an answer to the first-mile/last-mile problem: flexible service that is still predictable and cheap to operate.

What carries the argument

The semi-on-demand (SoD) route structure: vehicles follow a fixed-route segment from the terminus, then enter a predetermined flexible zone where they make on-demand stops, then return to the fixed-route schedule. The zonal dispatch is posed as a Markov decision process with four actions (regular route, Zone 1, Zone 2, or hold), and solved with the policy-gradient algorithm Proximal Policy Optimization (PPO), which clips policy updates to keep training stable. The reward is the negative count of request rejections due to excessive waiting.

What would settle it

Run the same simulation with the demand model modified to keep demand fixed regardless of walking time, holding all other settings identical. If the RL-zonal service then serves roughly the same number of passengers as the fixed-route baseline (rather than 16% more), the headline gain is an artifact of the assumed demand elasticity rather than of the dispatching policy.

Watch

Extended reading notes

Core claim

The central claim is that dynamic zonal dispatching—deciding, at each departure, whether a shared autonomous vehicle should serve the regular route, flexible Zone 1, or flexible Zone 2—is best handled by a deep reinforcement-learning policy trained with Proximal Policy Optimization. The policy observes 18 state variables (vehicle counts, demand forecast, unassigned requests, scheduled flexible time, and boarding/alighting loads per zone) and acts to minimize rejected requests. After training on 2,000 demand instances and evaluating on 100 separate instances, the RL-zonal service serves 16% more passengers than the fixed-route service and 2.4% more than a nominal zonal service, with generaliz

Load-bearing premise

Demand is assumed to grow linearly as walking time decreases (up to 10 minutes), but the slope of that relationship is not calibrated to any observed data; if real ridership responds less strongly to reduced walking, the reported 16% passenger gain shrinks or disappears.

Editorial extensions

If this is right

  • If the claim holds, transit operators can extend fixed-route service into low-density areas without a full demand-responsive system, using spare autonomous vehicles to serve flexible zones on demand.
  • The RL policy learns to hold vehicles at the terminus to avoid bunching and to serve future demand, a behavior not explicitly programmed, implying that learned dispatch can adapt to demand fluctuation.
  • The 16% passenger gain includes induced demand from reduced walking access: semi-on-demand service eliminates walking in the flexible zone, which attracts riders who would otherwise not use the service.
  • The cost breakdown shows the trade-off is mostly access time saved versus waiting and riding time increased; the RL controller keeps the extra cost small while adding riders.
  • The framework is demonstrated on a directional feeder corridor in off-peak conditions; the same structure could generalize to other directional feeders with predictable demand patterns.

Reading between the lines

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

  • The measured 16% gain depends on the paper's assumption that demand rises linearly with reduced walking time (up to 10 minutes), but the slope of that demand response is not calibrated to data; if real-world response is flatter, the service may still improve access but serve fewer additional passengers.
  • A field test or a stated-preference study that measures actual ridership change under a semi-on-demand feeder could directly validate the demand-induction assumption; that result is the most falsifiable piece of the claim.
  • The RL controller's advantage over nominal zonal assignment is modest (2.4% more passengers, 1.4% higher cost), suggesting that the bigger win is the hybrid route design itself, not the learning algorithm—though the learning algorithm also avoids the need to hand-tune zone dispatch frequencies.
  • The state representation and action space are small and the training budget modest (24 hours, 2.88 million steps); the policy might transfer to other feeder routes with similar demand shapes, but that transfer is not demonstrated.
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

3 major / 5 minor

Summary. This paper proposes a semi-on-demand (SoD) feeder service using shared autonomous vehicles, in which vehicles first serve fixed-route stops and then serve a flexible-route area on demand. A proximal policy optimization (PPO) agent dynamically decides whether to dispatch an available vehicle to one of two zones, to the regular SoD route, or to hold the vehicle. The agent-based simulation uses FleetPy on Munich bus route 193 with demand adapted from a boarding/alighting dataset. The paper reports that RL-zonal SoD serves 16% more passengers than fixed-route service at 13% higher generalized cost, and that the RL control contributes 2.4% more passengers at 1.4% higher cost relative to nominal zonal dispatching. The central claims are that the SoD concept increases served passengers and that RL zonal control improves operational efficiency relative to fixed zonal rules.

Significance. The methodological core is sound: the MDP formulation is reasonable; PPO is implemented with standard clipping, GAE, and reported hyperparameters; training uses 2,000 demand instances with a separate 100-instance held-out evaluation; and the simulation is built on the open-source FleetPy framework with all key hyperparameters listed. If the empirical claims hold, the paper provides a practical, transferable control design for an important first-mile/last-mile problem. The main caveat is that the passenger-serving advantage is conditional on a demand model whose elasticity is neither calibrated nor reported, so the headline improvement is not fully identified.

major comments (3)
  1. [Section III-A] The demand model is stated as 'a linear function of walking time (maximum 10 min)' but no slope, intercept, calibration source, or functional form is given. Because SoD removes nearly all walking/access time relative to the fixed-route baseline, this assumption mechanically generates additional trip requests; Section III-B attributes 10% of the passenger gain to SoD before any RL effect. The headline 16% gain over fixed route therefore mixes an unquantified induced-demand elasticity with operational efficiency. Please report the demand model parameters and generated request counts, and add a sensitivity analysis with fixed demand or flatter demand responses so the reader can separate supply-side gains from demand-side assumptions.
  2. [Section III-B, Figs. 7-9] The reported percentages are averages over 100 simulation runs, but the paper presents only boxplots and no confidence intervals or paired-difference tests for the mean passenger counts or generalized costs. Given the stochasticity of the demand instances, the '16% more' and '2.4% more' claims could be within sampling noise. Please report the means with standard errors or confidence intervals, and ideally paired differences between control types on the same demand instances.
  3. [Section III-A] The paper states that trip origins and destinations are 'adapted from a boarding and alighting dataset' but does not describe the transformation: no scaling, spatial aggregation, temporal profile, or procedure for applying the walking-time elasticity is provided. This makes the simulation non-reproducible and prevents an independent check of the linear demand assumption. Please add a data-generation appendix or pseudocode, and clarify whether the demand elasticity applies to trip generation, OD selection, or both.
minor comments (5)
  1. [Section III-A] The long list of simulation and RL parameters is given in prose. A table would improve readability and reduce the chance of transcription errors.
  2. [Section II-B] The paper mentions 18 state variables but does not itemize them. Please provide the explicit list, including how the 15-min demand forecast is represented.
  3. [Fig. 4] 'M.A.' in the caption should be defined. Also, the reward plot would benefit from a pre-training or random-policy baseline.
  4. [Abstract / Section III-B] The phrase 'efficient training' is not supported by a comparison of training budgets or sample efficiency; consider softening or adding evidence.
  5. [Equation (2)] Please define each cost component in the text immediately after the equation, especially the relationship between access time, waiting time, and riding time for a request.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RL dispatch comparison is an out-of-sample simulation result; demand model is an explicit input assumption, not a fitted predictor.

full rationale

The paper's central claim is that RL zonal dispatching serves 2.4% more passengers at 1.4% higher costs compared to nominal zonal dispatching. This result comes from 100 demand instances separate from the RL training set, as stated: "The results of each control type are from simulations on 100 demand instances (separate from the RL training set)." The RL policy is trained on 2000 instances and then evaluated statically, so the comparison is genuinely out-of-sample and not a fitted artifact. The larger 16% gain over fixed-route service is partly driven by the demand model: "demand is assumed to be a linear function of walking time (maximum 10 min)" (Section III-A). This is an explicit input assumption, not a hidden fit. The paper transparently attributes the difference to convenience: "The convenience of SoD service to pick up and drop off passengers attracts more demand, resulting in 10% more passengers served on average compared to the fixed route." While the demand model's parameters are unreported, that is a reproducibility weakness, not circularity, because the passenger counts are neither equivalent to the demand function nor statistically forced from a fitted parameter. The authors' self-citations (e.g., [5] for zone sizes, [21] for the simulation adaptation) are references to prior tools and parameter choices, not to the claimed efficiency result. No equation or result reduces to its own input by construction. The RL reward is based on rejections, and the reported passenger counts and costs are simulated outputs, not the reward itself. Thus, no circular step is present.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims rest on two types of upstream content: a simulator taken as ground truth (FleetPy, developed by the same research group) and a demand model whose functional form is assumed without calibration. No new physical entities are postulated. The most consequential assumption is the unquantified linear demand response to walking time, because it directly generates the passenger gains that form the headline result.

free parameters (4)
  • Demand response coefficients (linear demand as a function of walking time)
    The simulator generates induced demand based on walking time with a linear form and a 10-minute cap, but the slope and intercept are not reported or estimated. This parameter largely determines the passenger gains attributed to SoD and RL zonal service.
  • Generalized cost weights (gamma_A, gamma_W, gamma_T, gamma_O, gamma_V) = 33 $/h, 24.75 $/h, 16.5 $/h, 0.694 $/km, 7.59 $/h
    These coefficients in Eq. (2) are adopted from Tirachini and Antoniou [26] and assumed to apply to the Munich scenario. The 13% and 1.4% cost comparisons are directly proportional to these values.
  • PPO hyperparameters = epsilon=0.2, gamma=0.99, lambda=0.95, learning rate=0.003, batch size=64, epochs=10, network 64x64
    Chosen by tuning, but the tuning procedure is not documented; policy quality may depend on these choices.
  • Insertion objective penalties (gamma_R, gamma_S) = 10^6 each
    Hand-set large values in Eq. (1) to force the insertion heuristic to prioritize request satisfaction; they affect which insertions are accepted and hence passenger counts.
assumptions (4)
  • domain assumption The FleetPy agent-based simulator faithfully represents SAV operations, including vehicle movement, request insertion, and schedule adherence.
    The authors rely on their own FleetPy framework [20,21] as ground truth; no independent validation of the simulation against field data is provided.
  • ad hoc to paper Trip demand is a linear function of walking time with a 10-minute cap.
    This demand model is introduced for this study to represent induced demand; no coefficients or calibration source are given.
  • domain assumption The 18 state variables and the 15-minute demand forecast satisfy the Markov property for the dispatching MDP.
    The paper asserts the state is parsimonious and Markov but does not specify how the demand forecast is generated.
  • domain assumption Generalized-cost coefficients from the literature apply unchanged to the Munich scenario.
    Cost weights in Eq. (2) are imported from [26] without scenario-specific calibration.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semi-on-Demand Transit Feeders with Shared Autonomous Vehicles and Reinforcement-Learning-Based Zonal Dispatching Control." pith.science (2026). https://pith.science/paper/NUMZMWIZ

@misc{pith2026250901883,
  author       = {Pith},
  title        = {Pith review of: Semi-on-Demand Transit Feeders with Shared Autonomous Vehicles and Reinforcement-Learning-Based Zonal Dispatching Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NUMZMWIZ}},
  note         = {Machine review of arXiv:2509.01883}
}
read the original abstract

This paper develops a semi-on-demand transit feeder service using shared autonomous vehicles (SAVs) and zonal dispatching control based on reinforcement learning (RL). This service combines the cost-effectiveness of fixed-route transit with the adaptability of demand-responsive transport to improve accessibility in lower-density areas. Departing from the terminus, SAVs first make scheduled fixed stops, then offer on-demand pick-ups and drop-offs in a pre-determined flexible-route area. Our deep RL model dynamically assigns vehicles to subdivided flexible-route zones in response to real-time demand fluctuations and operations, using a policy gradient algorithm - Proximal Policy Optimization. The methodology is demonstrated through agent-based simulations on a real-world bus route in Munich, Germany. Results show that after efficient training of the RL model, the semi-on-demand service with dynamic zonal control serves 16% more passengers at 13% higher generalized costs on average compared to traditional fixed-route service. The efficiency gain brought by RL control brings 2.4% more passengers at 1.4% higher costs. This study not only showcases the potential of integrating SAV feeders and machine learning techniques into public transit, but also sets the groundwork for further innovations in addressing first-mile-last-mile problems in multimodal transit systems.

Figures

Figures reproduced from arXiv: 2509.01883 by the authors.

Figure 1
Figure 1. Illustration of various routes as a feeder service (SoD: semi-on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Components and data flows between the RL model and simulation [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Sample trajectories of SAVs Simulation runs from 9 p.m. to midnight for off-peak demand, with the first hour ignored in metric evaluation for warm-up. Trip origins and destinations are adapted from a boarding and alighting dataset of a local public transit operator, whereas demand is assumed to be a linear function of walking time (maximum 10 min). GIS dataset of transit alignment and stop locations [24] and road ne… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Training metrics of the RL model (M.A.: 10-step moving average) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 8
Figure 8. Figure 8: Vehicle distance traveled of four control types [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 6
Figure 6. Figure 6: Densities of RL model actions [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Requests satisfied of four control types [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 24 canonical work pages

  1. [1]

    Intermodal Autonomous Mobility-on-Demand,

    M. Salazar, N. Lanzetti, F. Rossi, M. Schiffer, and M. Pavone, “Intermodal Autonomous Mobility-on-Demand,” IEEE Transactions on Intelligent Transportation Systems , vol. 21, no. 9, pp. 3946–3960, Sept. 2020

    work page 2020

  2. [2]

    Enhanced Mobility With Con- nectivity and Automation: A Review of Shared Autonomous Vehicle Systems,

    L. Zhao and A. A. Malikopoulos, “Enhanced Mobility With Con- nectivity and Automation: A Review of Shared Autonomous Vehicle Systems,” IEEE Intelligent Transportation Systems Magazine , vol. 14, no. 1, pp. 87–102, Jan. 2022

    work page 2022

  3. [3]

    A survey on demand-responsive public bus systems,

    P. Vansteenwegen, L. Melis, D. Aktas ¸, B. D. G. Montenegro, F. Sar- tori Vieira, and K. S ¨orensen, “A survey on demand-responsive public bus systems,” Transportation Research Part C: Emerging Technolo- gies, vol. 137, p. 103573, 2022

    work page 2022

  4. [4]

    Redesigning large-scale multimodal transit networks with shared autonomous mobility services,

    M. T. M. Ng, H. S. Mahmassani, O. Verbas, T. Cokyasar, and R. Engelhardt, “Redesigning large-scale multimodal transit networks with shared autonomous mobility services,” Transportation Research Part C: Emerging Technologies , p. 104575, Mar. 2024

    work page 2024

  5. [5]

    Autonomous Minibus Service With Semi-on-Demand Routes in Grid Networks,

    M. T. M. Ng and H. S. Mahmassani, “Autonomous Minibus Service With Semi-on-Demand Routes in Grid Networks,” Transportation Research Record, vol. 2677, no. 1, pp. 178–200, 2023

    work page 2023

  6. [6]

    A survey on planning semi-flexible transit systems: Methodological issues and a unifying framework,

    F. Errico, T. G. Crainic, F. Malucelli, and M. Nonato, “A survey on planning semi-flexible transit systems: Methodological issues and a unifying framework,” Transportation Research Part C: Emerging Technologies, vol. 36, pp. 324–338, 2013

    work page 2013

  7. [7]

    Flexing service schedules: Assessing the potential for demand-adaptive hybrid transit via a stated preference approach,

    C. Frei, M. Hyland, and H. S. Mahmassani, “Flexing service schedules: Assessing the potential for demand-adaptive hybrid transit via a stated preference approach,” Transportation Research Part C: Emerging Technologies, vol. 76, pp. 71–89, Mar. 2017

    work page 2017

  8. [8]

    Semi-on-Demand Hybrid Transit Route Design with Shared Autonomous Mobility Services,

    M. T. M. Ng, F. Dandl, H. S. Mahmassani, and K. Bogenberger, “Semi-on-Demand Hybrid Transit Route Design with Shared Autonomous Mobility Services,” in 103rd Transportation Research Board Annual Meeting , Washington, D.C., 2024. [Online]. Available: http://arxiv.org/abs/2403.15804

    work page arXiv 2024

Show all 27 references
  1. [9]

    2-Vehicle zone optimal design for feeder transit services,

    X. Li and L. Quadrifoglio, “2-Vehicle zone optimal design for feeder transit services,” Public Transport , vol. 3, no. 1, pp. 89–104, Feb. 2011

    2011

  2. [10]

    Optimal Zonal Design for Flexible Bus Service Under Spatial and Temporal Demand Uncertainty,

    E. Lee, H. K. Lo, and M. Li, “Optimal Zonal Design for Flexible Bus Service Under Spatial and Temporal Demand Uncertainty,” IEEE Transactions on Intelligent Transportation Systems , vol. 25, no. 1, pp. 251–262, Jan. 2024

    2024

  3. [11]

    SAMoD: Shared Autonomous Mobility- on-Demand using Decentralized Reinforcement Learning,

    M. Gueriau and I. Dusparic, “SAMoD: Shared Autonomous Mobility- on-Demand using Decentralized Reinforcement Learning,” in 2018 21st International Conference on Intelligent Transportation Systems (ITSC). Maui, HI: IEEE, Nov. 2018, pp. 1558–1563

    2018

  4. [12]

    Dispatch of autonomous ve- hicles for taxi services: A deep reinforcement learning approach,

    C. Mao, Y . Liu, and Z.-J. M. Shen, “Dispatch of autonomous ve- hicles for taxi services: A deep reinforcement learning approach,” Transportation Research Part C: Emerging Technologies , vol. 115, p. 102626, June 2020

    2020

  5. [13]

    Context-Aware Taxi Dispatching at City- Scale Using Deep Reinforcement Learning,

    Z. Liu, J. Li, and K. Wu, “Context-Aware Taxi Dispatching at City- Scale Using Deep Reinforcement Learning,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 3, pp. 1996–2009, Mar. 2022

    1996

  6. [14]

    Deep dispatching: A deep reinforcement learning approach for vehicle dispatching on on- line ride-hailing platform,

    Y . Liu, F. Wu, C. Lyu, S. Li, J. Ye, and X. Qu, “Deep dispatching: A deep reinforcement learning approach for vehicle dispatching on on- line ride-hailing platform,” Transportation Research Part E: Logistics and Transportation Review , vol. 161, p. 102694, May 2022

    2022

  7. [15]

    Reinforcement learning for ridesharing: An extended survey,

    Z. T. Qin, H. Zhu, and J. Ye, “Reinforcement learning for ridesharing: An extended survey,” Transportation Research Part C: Emerging Technologies, vol. 144, p. 103852, Nov. 2022

    2022

  8. [16]

    Deep Reinforcement Learning for In- telligent Transportation Systems: A Survey,

    A. Haydari and Y . Yilmaz, “Deep Reinforcement Learning for In- telligent Transportation Systems: A Survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 1, pp. 11–32, Jan. 2022

    2022

  9. [17]

    A Multiline Customized Bus Planning Method Based on Reinforcement Learning and Spatiotemporal Clustering Algorithm,

    W. Li, L. Zheng, L. Liao, X. Yang, D. Sun, and W. Liu, “A Multiline Customized Bus Planning Method Based on Reinforcement Learning and Spatiotemporal Clustering Algorithm,” IEEE Transactions on Computational Social Systems , pp. 1–15, 2024

    2024

  10. [18]

    Multi-agent deep rein- forcement learning based real-time planning approach for responsive customized bus routes,

    B. Wu, X. Zuo, G. Chen, G. Ai, and X. Wan, “Multi-agent deep rein- forcement learning based real-time planning approach for responsive customized bus routes,” Computers & Industrial Engineering, vol. 188, p. 109840, Feb. 2024

    2024

  11. [19]

    Proximal Policy Optimization Algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal Policy Optimization Algorithms,” Aug. 2017. [Online]. Available: http://arxiv.org/abs/1707.06347

    2017 arXiv

  12. [20]

    FleetPy: A Modular Open-Source Simulation Tool for Mobility On-Demand Services,

    R. Engelhardt, F. Dandl, A.-A. Syed, Y . Zhang, F. Fehn, F. Wolf, and K. Bogenberger, “FleetPy: A Modular Open-Source Simulation Tool for Mobility On-Demand Services,” 2022. [Online]. Available: http://arxiv.org/abs/2207.14246

    2022 arXiv

  13. [21]

    Simulating Semi-on-Demand Hybrid Route Transit Feeders with Shared Autonomous Mobility Services,

    M. T. Ng, R. Engelhardt, F. Dandl, K. Bogenberger, and H. S. Mah- massani, “Simulating Semi-on-Demand Hybrid Route Transit Feeders with Shared Autonomous Mobility Services,” in12th Symposium of the European Association for Research in Transportation , Espoo, 2024

    2024

  14. [22]

    A Closer Look at Invalid Action Masking in Policy Gradient Algorithms,

    S. Huang and S. Onta ˜n´on, “A Closer Look at Invalid Action Masking in Policy Gradient Algorithms,” The International FLAIRS Conference Proceedings, vol. 35, May 2022

    2022

  15. [23]

    High-Dimensional Continuous Control Using Generalized Advantage Estimation,

    J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, “High-Dimensional Continuous Control Using Generalized Advantage Estimation,” Oct. 2018. [Online]. Available: http://arxiv.org/abs/1506.02438 [Accessed: 2024-04-22]

    2018 arXiv

  16. [24]

    MVG timetable data in GTFS format,

    MVG, “MVG timetable data in GTFS format,” 2024. [Online]. Avail- able: https://www.mvg.de/services/fahrgastservice/fahrplandaten.html [Accessed: 2024-01-26]

    2024

  17. [25]

    OSMnx: New methods for acquiring, constructing, an- alyzing, and visualizing complex street networks,

    G. Boeing, “OSMnx: New methods for acquiring, constructing, an- alyzing, and visualizing complex street networks,” Computers, Envi- ronment and Urban Systems , vol. 65, pp. 126–139, 2017

    2017

  18. [26]

    The economics of automated public transport: Effects on operator cost, travel time, fare and subsidy,

    A. Tirachini and C. Antoniou, “The economics of automated public transport: Effects on operator cost, travel time, fare and subsidy,” Economics of Transportation , vol. 21, p. 100151, 2020

    2020

  19. [27]

    Stable-baselines3: Reliable reinforcement learning im- plementations,

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann, “Stable-baselines3: Reliable reinforcement learning im- plementations,” Journal of Machine Learning Research , vol. 22, no. 268, pp. 1–8, 2021

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.