Wise Goose Chase: A Predictive Path Planning Algorithm for Dynamic Rebalancing in Ride-Hailing Systems

Avalpreet Singh Brar; Christos G. Cassandras; Gioele Zardini; Rong Su

REVIEW 5 major objections 5 minor 13 references

Wise Goose Chase: A Predictive Path Planning Algorithm for Dynamic Rebalancing in Ride-Hailing Systems

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The Wise Goose Chase algorithm computes personalized cruising paths that minimize a driver's expected time to a passenger match, using edge-level forecasts of supply and demand.

desk verdict Plausible new algorithm with a genuine edge-level forecast framework, but the headline optimality claim does not hold in the fleet-wide deployment the experiments actually test. read the letter →

arxiv 2505.02603 v1 pith:BLJQAGTB submitted 2025-05-05 eess.SY cs.SY

Avalpreet Singh Brar , Rong Su , Christos G. Cassandras , Gioele Zardini This is my paper · ORCID

classification eess.SYcs.SY

keywords ride-hailingrebalancingpathplanningretardedfunctionaldifferentialequationssurvivalprobabilitydrivercompetitionexpectedallocationtimemobility-on-demandevent-triggeredpolicy

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 establish that idle ride-hailing drivers should be guided along full cruising paths rather than to fixed destinations, because allocations often happen while the driver is still on the road. It introduces the Wise Goose Chase (WGC) algorithm, which forecasts passenger queues and idle-driver counts at the road-segment level and then selects the path with the smallest expected time to allocation, including matches that occur mid-edge. Monte Carlo simulations on a synthetic 10x10 grid across fleet sizes from 100 to 5000 drivers show WGC with the lowest mean allocation time when compared with random-walk, greedy, and hotspot-guided baselines. If those results carry over to real platforms, event-triggered, path-level guidance is a practical alternative to destination-based rebalancing.

What carries the argument

The load-bearing object is the survival operator $G_{uv}(t)=\exp\left(-\int_{t-\tau_e}^{t} A_e(s)/D_e(s)\,ds\right)$, the probability that a driver crossing edge $(u,v)$ remains unallocated over the traversal interval. It appears in the system of retarded functional differential equations (delay differential equations with state history) that forecast idle-driver densities on edges and nodes, and it also defines the path objective: the survival curve $S(t)$ is the product of these factors along completed and current edges, and WGC integrates $S(t)$ to obtain the expected allocation time. A beam-search variant retains only the top $k$ partial paths, reducing evaluation complexity from $O(d^L \tau_{\max} L)$ to $O(k L d \tau_{\max})$.

What would settle it

Run a simulation in which half or more of the idle drivers follow WGC recommendations and compare the realized expected allocation time with the survival-probability forecast; if realized times systematically exceed the forecast, the open-loop fixed-transition-matrix assumption is falsified.

Watch

Extended reading notes

Core claim

The central claim is that rebalancing in ride-hailing is better posed as path planning over road segments than as destination assignment. WGC models each edge's passenger queue $Q_e(t)$ and idle-driver count $D_e(t)$, with instantaneous matches at rate $A_e(t)=\min(Q_e(t),D_e(t))$, and uses a system of retarded functional differential equations to predict how these quantities evolve. For a candidate path $\pi$, the probability of remaining unmatched up to time $t$ is the product of edge survival factors $\exp\left(-\int_{t-\tau_e}^{t} A_e(s)/D_e(s)\,ds\right)$, and the expected allocation time is $\int_0^{T_\pi} S(t)\,dt$; the recommended path minimizes that integral. In the reported Monte Carlo experiments, WGC attains the lowest mean allocation time at every fleet size tested, with the advantage over greedy routing widening as the fleet grows.

Load-bearing premise

The forecasts behind WGC's route choice assume every other idle driver keeps following the same fixed transition matrix, so a recommended path is optimal only for a single driver acting alone while everyone else's behavior is unchanged.

Editorial extensions

If this is right

Because WGC evaluates entire paths rather than endpoints, matches that occur while a driver is cruising along a road segment enter the optimization instead of being ignored.
The expected-allocation-time objective depends on the survival probability through the ratio $A_e/D_e$, so competition among idle drivers on each edge is explicitly priced into the route choice.
The event-triggered design computes a recommendation only when a driver requests one, avoiding periodic platform-wide broadcast decisions.
In the reported simulations WGC has the lowest mean allocation time at every fleet size from 100 to 5000; at $N=5000$ the means are 119.63 s for WGC versus 212.67 s for greedy routing.

Reading between the lines

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

An implication the authors leave implicit is that the forecast is open-loop: it does not model the feedback of WGC's own recommendations on $D_e(t)$ and $Q_e(t)$, so a version adopted by many drivers at once would need a closed-loop or equilibrium forecast.
The same survival-probability objective transfers to other task-cruising platforms, such as food delivery, courier services, or on-demand freight, with matching and patience parameters re-estimated for each setting.
A testable extension is to log predicted versus realized allocation times under increasing rates of driver compliance; the adoption level at which the open-loop forecast degrades would show where the model has to be re-closed.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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: this is a plausible new algorithm with a real gap between its promise and its proof. The edge-level RFDE model with delays and survival probabilities is a reasonable way to describe idle-driver competition and en-route matching, and the WGC path evaluation is clearly specified. I'd give the authors credit for the model and for an honest complexity analysis (the exponential path enumeration is acknowledged and a beam-search heuristic is proposed).

The soft spots are not fatal but they're real. The biggest one is exactly the stress-test concern: Eqs. (23) and (27) assume every idle driver follows the fixed CTMC Q. If a single driver uses WGC, that's consistent. But Table II compares fleet-wide policies, where everyone follows WGC—then the forecast used to compute survival probabilities is wrong on its own terms. The paper never states the deployment regime, so the claimed 'optimal' paths are only optimal in a single-driver deviation world. This needs to be stated and ideally modeled.

Second, the experiments give the algorithm perfect knowledge of the passenger arrival rates λe(t), and the simulation environment is the same RFDE model. That makes the comparison favorable to WGC in a way that won't hold with forecast error. The authors should at least add a comparison with noisy λ estimates.

Third, the objective in Eq. (33) is truncated at the path horizon. If a driver can survive the whole path without a match, the expected time to allocation is underestimated. That's a real bias, though it may be minor if typical survival probabilities are small by then.

The paper also doesn't compare against the closest prior method (MDM, [13]). No error bars in Table II, though the mean gaps are large enough that significance is plausible.

Overall, the central idea is interesting and the derivation is mostly clean, but the optimality claim is unverified in the fleet-wide regime. This deserves a serious referee, not a desk reject. The authors need to fix the deployment-regime story, add uncertainty, and report error bars before publication.

Referee Report

5 major / 5 minor

Summary. The paper proposes the Wise Goose Chase (WGC) algorithm, an event-triggered, driver-specific path planning framework for idle ride-hailing drivers. WGC forecasts spatio-temporal supply and demand at the road-segment level through a system of retarded functional differential equations (RFDEs) describing passenger queues, idle drivers on edges, and idle drivers at nodes. For a tagged driver, the algorithm evaluates candidate paths by computing survival probabilities from the forecast hazard rates and minimizes the integral of the survival probability over the planned path. Monte Carlo simulations on a 10x10 grid network compare WGC against random walk, greedy, and hotspot-guided baselines for fleet sizes from 100 to 5000, reporting lower mean and worst-case allocation times. The paper also provides a complexity analysis and a beam-search acceleration scheme. The central claim is that WGC computes personalized cruising paths that minimize each driver's expected time to allocation and consistently outperforms the tested baselines.

Significance. If the central claims are established, WGC would be a useful path-level alternative to destination-based rebalancing, with the distinctive feature of edge-level RFDE forecasting and explicit survival-probability path evaluation. The paper has clear strengths: it clearly frames the driver-specific, event-triggered setting; it provides explicit derivations for passenger abandonment and survival operators (Eqs. (2)-(20)); it gives reproducible algorithmic pseudocode (Algorithm 1); and it evaluates performance across six fleet sizes with Monte Carlo trials. However, the main technical claims are not yet supported because of a dimensional inconsistency in the matching rate, a truncated objective that does not equal the expected allocation time, and an unmodeled feedback loop between the WGC policy and the forecast dynamics. These issues are load-bearing for the claimed optimality and for the interpretation of the experimental comparison.

major comments (5)

[Section II-B, Eq. (10) and Section II-D, Eq. (30)] The matching rate is defined as A_e(t) = min(D_e(t), Q_e(t)), but D_e(t) and Q_e(t) are defined as total numbers of idle drivers and waiting passengers on edge e, so min(D_e, Q_e) is a count, not a rate. Inserting this expression into the differential equations (11), (23), and (27) makes the time derivatives dimensionally inconsistent, and all quantitative predictions, including the survival probability in Eq. (20), depend on an unspecified time-scale conversion. Please define A_e as a rate per unit time, e.g., A_e(t) = kappa_e min(D_e(t), Q_e(t)) with a calibrated kappa_e, or justify a fluid scaling in which min(D_e, Q_e) has the units of a rate.
[Section II-D, Eq. (33)] The objective E[T_alloc | pi] is defined as the integral of S(t) from 0 to T_pi, which is the expected allocation time truncated at the planned path end, not the true expected time to allocation. If the driver is still unmatched at T_pi, the integral assigns no contribution beyond the path, whereas the true expectation includes S(T_pi) times the expected remaining time plus T_pi. As written, Algorithm 1 can prefer a path with a lower truncated expectation even when that path leaves the driver stranded, so the claim that WGC 'minimizes each driver's expected time to allocation' is not supported by the stated objective. The post-path continuation rule must be specified and used consistently in both the optimization and the simulation.
[Section II-C, Eqs. (23) and (27), with Algorithm 1] The forecast dynamics assume that all idle drivers continue to select outgoing edges according to the fixed CTMC transition matrix Q, yet WGC is precisely a routing policy that changes those choices. The path evaluation in Algorithm 1 uses h(t_k) = A_e(t_k)/D_e(t_k) from forecasts that are valid only in the single-driver-deviation regime, where the tagged driver's action does not affect the aggregate state. The paper does not state whether the 'WGC' strategy in Table II is applied to a single tagged driver or to the entire fleet, and it does not test self-consistency, for example by iterating between the induced transition matrix and the forecast. The headline comparison therefore conflates the policy's performance with the accuracy of a forecast that the policy itself invalidates. Please specify the deployment regime and either prove or verify the single-driver-deviation property, or solve a closed-loop forecast that accounts for the routing-induced changes in D_e(t) and Q_e(t).
[Section II-A, Eq. (1) and Section II-C] The conservation law Eq. (1) includes occupied-driver variables \tilde D_e(t) and \tilde P_u(t), but the RFDE system provides no dynamics for these variables. The return of occupied drivers is inserted into Eq. (27) as sum_e R_{e to u} A_e(t - tau_{eu}), but without equations for \tilde D_e and \tilde P_u it is not demonstrated that the forecast preserves total driver count or that the occupied-driver component is consistent with the claimed state. Either derive the occupied-driver dynamics or explicitly state that Eq. (1) is an accounting identity that is not enforced by the forecast model.
[Section III-B/C, Table II] Table II reports only mean and worst-case allocation times over 100 trials, with no standard errors, confidence intervals, or significance tests. The qualitative ranking at large fleet sizes is plausible, but the claim of 'statistically robust performance estimates' is not supported, and it remains unclear whether WGC's advantage comes from the forecast model or from the specific path objective. Please report error bars or confidence intervals and, ideally, compare against a state-of-the-art path-based method such as the MDM approach cited as [13].

minor comments (5)

[Section II-C, opening paragraph] The transition-probability normalization is written as sum_{w in G+(v)} Q_{wv} = 1, but the indices appear reversed; it should likely be sum_{w in G+(v)} Q_{vw} = 1 for outgoing transitions.
[Algorithm 1 and Eq. (38)] When D_e(t_k) = 0, the hazard h(t_k) = A_e(t_k)/D_e(t_k) is undefined; the implementation needs a guard, for example setting the hazard to zero when there are no idle drivers on the edge.
[Figures 2 and 3] The axis labels and units are missing from the figures as presented; please add them and state explicitly what 'convergence' means in Figure 2.
[Section III-D] The claim that beam search achieves 'negligible loss in path optimality' is not quantified; please provide an experiment or a bound to support this statement.
[Section III, simulation setup] The simulation description does not specify how often WGC recommendations are recomputed, how drivers respond if they deviate from a recommendation, or how the inputs lambda_e(t), Q, and R_{e to u} are estimated in practice; these are treated as known, but the assumptions should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: WGC's path scores follow from a self-contained RFDE forecast and standard survival analysis; no fitted input is relabeled as a prediction.

full rationale

The WGC derivation is self-contained. The passenger-queue and idle-driver dynamics in Eqs. (11), (23), and (27) are conservation laws driven by exogenous inputs λe(t), the CTMC matrix Q, edge traversal times τe, and patience rate µ. The survival probability G_uv(t) in Eq. (20) is derived from the hazard A_e/D_e rather than assumed equal to the objective, and the expected allocation time in Eq. (33) is the standard integral of the survival function, not a fitted quantity. The optimal path in Eq. (34) minimizes that integral, so the claimed minimization holds by the definition of the objective, not by circular reduction to the inputs. The Monte Carlo comparison in Table II is an empirical evaluation, not a construction that forces WGC's score. Author self-citations [4], [8], [11], [12] are contextual related-work references and are not used to justify the WGC model or its optimality. The forecast's reliance on a fixed CTMC for other drivers is a mean-field self-consistency limitation that would matter under fleet-wide adoption, but it is a modeling-accuracy issue, not a case of the prediction being equivalent to its inputs by construction.

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

The central claim rests on an uncalibrated simulation model. The paper pulls many inputs (lambda_e, Q, R, patience rate, algorithm constants) from choices rather than data, and the main unstated load-bearing assumption is that the forecast remains valid when drivers act on the recommendations. The ledger shows that the actual contribution is a modeling-and-optimization framework, not a data-derived discovery.

free parameters (5)

Passenger patience rate mu = 0.1 per second (mean patience 10 seconds)
Chosen as a simulation input; directly controls the abandonment term mu Q_e(t) in Eq. (11) and therefore affects queue dynamics and matching opportunities.
Per-edge demand arrival profiles lambda_e(t) = Unspecified; base rates from a uniform distribution, hotspot edges with elevated rates, and a sinusoidal perturbation
Central input to the WGC forecast in Algorithm 1. WGC is assumed to know these future rates exactly, but the exact values, hotspot multipliers, and perturbation parameters are not reported.
CTMC transition matrix Q = Unspecified
Governs how idle drivers move in the predicted dynamics in Eqs. (23) and (27). Without its values the simulation and forecast cannot be reproduced.
Destination popularity distribution R_{e to u} = Unspecified
Determines where occupied drivers return after a trip and feeds the node-level idle driver inflow term in Eq. (27).
Max path length L and beam width k = Not specified
Algorithm 1 enumerates simple paths up to L and the beam search retains k partial paths. Both affect path quality and runtime, but their experimental values are not reported.

assumptions (5)

domain assumption Matching occurs immediately at rate A_e(t) = min(D_e(t), Q_e(t)) whenever both an idle driver and a waiting passenger are present
Assumed in Eq. (10). This ignores spatial distribution within an edge, queueing order, platform matching delays, and the fact that a single driver cannot match multiple passengers at once.
domain assumption Each idle driver on an edge is equally likely to be matched, so the individual allocation hazard is A_e(t) / D_e(t)
Used to derive the survival probability G_uv(t) in Eq. (20). It requires D_e(t) > 0 and uniform mixing of drivers and passengers on the edge.
ad hoc to paper Idle drivers follow a fixed CTMC with transition matrix Q, independent of WGC recommendations
The forecast equations (23) and (27) use Q for all idle drivers. When the recommended policy changes driver choices, the forecast is not self-consistent.
domain assumption The platform knows the future arrival rates lambda_e(t) exactly
The numerical trajectory generation in Algorithm 1 uses the true lambda_e(t) profiles. No forecasting error or estimation procedure is modeled.
ad hoc to paper The expected allocation time may be truncated at T_pi with zero contribution after the path ends
Eq. (33) integrates survival only up to T_pi, so paths shorter than the true allocation horizon are systematically favored and the stated objective does not equal expected time to allocation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Wise Goose Chase: A Predictive Path Planning Algorithm for Dynamic Rebalancing in Ride-Hailing Systems." pith.science (2026). https://pith.science/paper/BLJQAGTB

@misc{pith2026250502603,
  author       = {Pith},
  title        = {Pith review of: Wise Goose Chase: A Predictive Path Planning Algorithm for Dynamic Rebalancing in Ride-Hailing Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BLJQAGTB}},
  note         = {Machine review of arXiv:2505.02603}
}

read the original abstract

Traditional rebalancing methods in ride-hailing systems direct idle drivers to fixed destinations, overlooking the fact that ride allocations frequently occur while cruising. This destination-centric view fails to exploit the path-dependent nature of modern platforms, where real-time matching depends on the entire trajectory rather than a static endpoint. We propose the Wise Goose Chase (WGC) algorithm, an event-triggered, driver-specific path planning framework that anticipates future matching opportunities by forecasting spatio-temporal supply and demand dynamics. WGC uses a system of Retarded Functional Differential Equations (RFDEs) to model the evolution of idle driver density and passenger queues at the road-segment level, incorporating both en-route matching and competition among drivers. Upon request, WGC computes personalized cruising paths that minimize each driver's expected time to allocation. Monte Carlo simulations on synthetic urban networks show that WGC consistently outperforms baseline strategies, highlighting the advantage of predictive, context-aware rebalancing in dynamic mobility systems.

Figures

Figures reproduced from arXiv: 2505.02603 by the authors.

**Figure 1.** This figure provides an overview of the WGC system dynamics. The top part presents the set of coupled differential equations describing the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 3.** Comparison of expected allocation time ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 6 canonical work pages

[13]

Route recommendations for idle taxi drivers: Find me the shortest route to a customer!

N. Garg and S. Ranu, “Route recommendations for idle taxi drivers: Find me the shortest route to a customer!” in Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 1425–1434

work page 2018
[1]

Hunting or waiting: Earning more by understanding taxi service strategies,

C. Chen, D. Zhang, Y . Wang, H. Huang, C. Chen, D. Zhang, Y . Wang, and H. Huang, “Hunting or waiting: Earning more by understanding taxi service strategies,” Enabling Smart Urban Services with GPS Trajectory Data, pp. 71–94, 2021

work page 2021
[2]

Robotic load balancing for mobility-on-demand systems,

M. Pavone, S. L. Smith, E. Frazzoli, and D. Rus, “Robotic load balancing for mobility-on-demand systems,” The International Journal of Robotics Research , vol. 31, no. 7, pp. 839–854, 2012

work page 2012
[3]

Vehicle rebalancing for mobility-on-demand systems with ride-sharing,

A. Wallar, M. Van Der Zee, J. Alonso-Mora, and D. Rus, “Vehicle rebalancing for mobility-on-demand systems with ride-sharing,” in 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2018, pp. 4539–4546

2018
[4]

Ensuring service fairness in taxi fleet man- agement,

A. S. Brar and R. Su, “Ensuring service fairness in taxi fleet man- agement,” in 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC) . IEEE, 2020, pp. 1–6

2020
[5]

On re-balancing self-interested agents in ride-sourcing transportation networks,

A. Sadeghi and S. L. Smith, “On re-balancing self-interested agents in ride-sourcing transportation networks,” in 2019 IEEE 58th Conference on Decision and Control (CDC) . IEEE, 2019, pp. 5119–5125

2019
[6]

Driver positioning and incentive budgeting with an escrow mechanism for ride-sharing platforms,

H. Y . Ong, D. Freund, and D. Crapis, “Driver positioning and incentive budgeting with an escrow mechanism for ride-sharing platforms,” INFORMS Journal on Applied Analytics , vol. 51, no. 5, pp. 373–390, 2021

work page 2021
[7]

Analysis and control of autonomous mobility-on-demand systems,

G. Zardini, N. Lanzetti, M. Pavone, and E. Frazzoli, “Analysis and control of autonomous mobility-on-demand systems,” Annual Review of Control, Robotics, and Autonomous Systems , vol. 5, no. 1, pp. 633– 658, 2022

2022

Show all 13 references

[8]

Vehicle rebalancing under adherence uncertainty,

A. S. Brar, R. Su, and G. Zardini, “Vehicle rebalancing under adherence uncertainty,” 2024. [Online]. Available: https://arxiv.org/ abs/2412.16632

2024 arXiv
[9]

i-rebalance: Personalized vehicle repositioning for supply demand balance,

H. Chen, P. Sun, Q. Song, W. Wang, W. Wu, W. Zhang, G. Gao, and Y . Lyu, “i-rebalance: Personalized vehicle repositioning for supply demand balance,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 1, 2024, pp. 46–54

2024
[10]

Vehicle/employee rebalancing and charging scheduling in one-way car sharing systems,

G. Guo, M. Kang, and T. Sun, “Vehicle/employee rebalancing and charging scheduling in one-way car sharing systems,” IEEE Trans- actions on Intelligent Transportation Systems , vol. 24, no. 10, pp. 10 665–10 675, 2023

2023
[11]

Supply-demand balancing model for ev rental fleet,

A. S. Brar, P. Kasture, and R. Su, “Supply-demand balancing model for ev rental fleet,” in 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC) . IEEE, 2022, pp. 1350– 1355

2022
[12]

Dynamic supply-demand balancing policy for cmod fleet,

A. S. Brar and R. Su, “Dynamic supply-demand balancing policy for cmod fleet,” in 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) . IEEE, 2021, pp. 2435–2440

2021

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [13]

Route recommendations for idle taxi drivers: Find me the shortest route to a customer!

N. Garg and S. Ranu, “Route recommendations for idle taxi drivers: Find me the shortest route to a customer!” in Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, 2018, pp. 1425–1434

work page 2018

[2] [1]

Hunting or waiting: Earning more by understanding taxi service strategies,

C. Chen, D. Zhang, Y . Wang, H. Huang, C. Chen, D. Zhang, Y . Wang, and H. Huang, “Hunting or waiting: Earning more by understanding taxi service strategies,” Enabling Smart Urban Services with GPS Trajectory Data, pp. 71–94, 2021

work page 2021

[3] [2]

Robotic load balancing for mobility-on-demand systems,

M. Pavone, S. L. Smith, E. Frazzoli, and D. Rus, “Robotic load balancing for mobility-on-demand systems,” The International Journal of Robotics Research , vol. 31, no. 7, pp. 839–854, 2012

work page 2012

[4] [3]

Vehicle rebalancing for mobility-on-demand systems with ride-sharing,

A. Wallar, M. Van Der Zee, J. Alonso-Mora, and D. Rus, “Vehicle rebalancing for mobility-on-demand systems with ride-sharing,” in 2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2018, pp. 4539–4546

2018

[5] [4]

Ensuring service fairness in taxi fleet man- agement,

A. S. Brar and R. Su, “Ensuring service fairness in taxi fleet man- agement,” in 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC) . IEEE, 2020, pp. 1–6

2020

[6] [5]

On re-balancing self-interested agents in ride-sourcing transportation networks,

A. Sadeghi and S. L. Smith, “On re-balancing self-interested agents in ride-sourcing transportation networks,” in 2019 IEEE 58th Conference on Decision and Control (CDC) . IEEE, 2019, pp. 5119–5125

2019

[7] [6]

Driver positioning and incentive budgeting with an escrow mechanism for ride-sharing platforms,

H. Y . Ong, D. Freund, and D. Crapis, “Driver positioning and incentive budgeting with an escrow mechanism for ride-sharing platforms,” INFORMS Journal on Applied Analytics , vol. 51, no. 5, pp. 373–390, 2021

work page 2021

[8] [7]

Analysis and control of autonomous mobility-on-demand systems,

G. Zardini, N. Lanzetti, M. Pavone, and E. Frazzoli, “Analysis and control of autonomous mobility-on-demand systems,” Annual Review of Control, Robotics, and Autonomous Systems , vol. 5, no. 1, pp. 633– 658, 2022

2022