Pith. sign in

REVIEW 5 major objections 5 minor 12 references

Randomized routing strategies of fleets of CAVs may prove market efficient

T0 review · 5 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Randomized routing can win CAV fleets more customers than system-optimal routing, a simulation study claims, and mixing mean travel time into fleet payoffs can steer competition back toward social welfare.

desk verdict A useful but overclaimed CAV routing benchmark: the one-fleet randomized routing result is suggestive, yet the headline mechanism (unpredictable travel times for HDVs) is absent from the model. read the letter →

arxiv 2607.14859 v1 pith:ODRHOMXK submitted 2026-07-16 cs.MA

Grzegorz Jamróz , {L}ukasz Gorczyca , Rafa{l} Kucharski This is my paper
classification cs.MA MSC 91A8090B20
keywords autonomousdrivingmarketdesignrandomizedroutingsharetraveltimevariabilityday-to-daydynamicsfleetcompetitiontrafficsimulation
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 asks what routing algorithms competing CAV fleet operators should use when they are paid by market share, and whether that market can be steered toward city goals. It models day-to-day choices of a fixed population of drivers who can drive themselves (HDVs) or join one of two fleets, with travel times set by congestion. The paper claims that randomized routing—deliberately making travel times unpredictable for independent drivers—lets a fleet win more customers than routing proportional to system optimum or user equilibrium, provided driver attitudes toward fleets vary widely. It further claims that adding the system's average travel time to the fleets' payoff, with enough weight, restores the advantage of socially optimal routing while keeping competition alive.

What carries the argument

The central objects are the randomized routing heuristics RFlexV and RFlex and the utility/credibility update of drivers. RFlexV sorts fleet members by discount factor γ and routes the highest-γ drivers to a randomly chosen route at a level that keeps them 'happy' (their disutility under fleet routing stays below HDV's). RFlex computes each driver's minimal share s_i of days on the faster route needed to keep that driver, then randomizes with target shares s_i/σ. The objective Obj_μ=(1−μ)share + μτ_avg is the device that penalizes randomization as μ grows.

What would settle it

Run the same day-to-day benchmark but give fleet operators only observable choice histories—not the drivers' true discount factors—and require them to estimate γ per driver. If randomized routing then fails to outperform SO/UE in market share, the central claim is falsified. Alternatively, re-run with driver attitudes drawn from a low-variance distribution (e.g., all γ near 0.7); the paper's own logic predicts the randomized advantage should largely disappear.

Watch

Extended reading notes

Core claim

In a benchmark of one OD pair with two parallel routes and 200 drivers, the authors simulate two fleet operators maximizing market share. They show that the randomized heuristics RFlexV and RFlex—which keep existing customers 'happy' by routing high-discount-factor drivers on the faster route a calibrated share of days while randomizing the rest—capture more market share than proportional SO/UE routing, especially when the other fleet sticks to SO. The randomization strips human drivers of the information that one route will be faster on a given day, making independent driving less attractive. However, when the fleet objective is augmented with mean systemwide travel time (Obj_μ = (1−μ) shar

Load-bearing premise

The fleets are assumed to know each driver's discount factor toward each fleet exactly, rather than having to infer it from observed behavior; the paper's headline advantage of randomized routing is demonstrated under that knowledge, and if operators must learn attitudes from choices, that advantage is not supported.

Editorial extensions

If this is right

  • If randomized routing is as effective as simulated, fleet operators maximizing market share will adopt it, and an operator facing a randomized rival cannot stay competitive with SO routing.
  • Market share alone as a remuneration rule invites travel-time volatility; adding average travel time to the objective, at weight roughly 0.5, makes randomization unprofitable and lets SO routing compete.
  • The benchmark provides a reproducible setup (BPR delay functions, logit HDV route choice, credibility update) for comparing future routing algorithms.
  • SUMO-based microsimulations with 150 vehicles also show the randomized algorithm beating SO, suggesting the qualitative result may extend beyond the abstract BPR setting.

Reading between the lines

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

  • The advantage of randomization likely hinges on the assumption that fleet operators know each driver's discount factor exactly; if attitudes must be learned from observed choices, the edge may shrink or vanish, and the market-design conclusion would need re-testing.
  • The same logic could apply beyond CAV markets: any platform that can commit to deliberately noisy service (e.g., unpredictable pricing or delivery times) may extract market share from competitors who optimize average outcomes—suggesting regulators should watch for 'strategic opacity' as a competitive weapon.
  • The Obj_μ idea could be tested as a regulatory contract: city pays operators partly for low mean travel time, which the simulations suggest tames randomization without killing competition; a natural next experiment is a market with more than two routes where the SO/UE distinction matters more.
  • Because the paper uses two parallel identical routes, the claim that randomization is 'more efficient' is demonstrated for that topology; on asymmetric or network topologies, the effect of randomization on HDV information could differ.
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

5 major / 5 minor

Summary. The paper studies a simulated market in which human drivers (HDVs) choose between driving independently and joining one of two competing CAV fleets, with fleet revenue proportional to market share. The authors propose several routing algorithms: proportional system-optimum (SO) and user-equilibrium (UE) baselines, and two randomized heuristics (RFlexV and RFlex) that route high-discount-factor drivers to the faster route while randomizing assignments. In a one-OD-pair, two-route BPR network, they report that randomized algorithms achieve higher market share than SO/UE when drivers' discount factors are known and diverse. They also propose a combined objective Obj_mu = (1-mu) n_f + mu tau_avg and show that for mu approx 0.5 the SO-based algorithm becomes competitive, which they interpret as discouraging antisocial randomization. The paper claims that randomized routing works by making travel times unpredictable for HDVs, and that a socially oriented objective preserves competition while limiting this behavior.

Significance. If the central claim held, the paper would make a useful contribution to the design of future CAV routing markets: it would show that market-share-maximizing fleets can exploit randomized routing to attract reluctant HDV drivers, and that a simple modification of the fleet objective can mitigate the resulting system inefficiency. The benchmark framework and explicit algorithm descriptions are valuable, and the SUMO microsimulation appendix is a positive step toward external validity. However, the significance is substantially weakened by the mismatch between the stated causal mechanism and the model, the mixed results in two-fleet scenarios, and the lack of statistical rigor (no error bars, no sensitivity analysis, parameter choices made after inspecting results). The paper introduces a useful starting point, but the evidence as presented does not support the strength of the abstract's claims.

major comments (5)
  1. [Section 2, Eq. (2)] If the authors intend the 'unpredictability' to be mediated by the mean travel times themselves, they should demonstrate that a deterministic cycling schedule with identical per-driver route frequencies fails to reproduce the advantage; otherwise the claim is unsupported.
  2. [Experiment 3, Obj_mu] Additionally, the normalized tau_avg uses t_SO_avg as the benchmark, which may mechanically favor SO algorithms; a different normalization (e.g., free-flow time) would provide a more neutral test.
  3. [Abstract and Experiment 2] This is not a request for perfection, but the discrepancy between the abstract and the body is load-bearing for the paper's central claim.
  4. [Table 1 / Section 3, Experiment 2] Without this, the 'strictly superior' claim in the Introduction is not supported by the reported evidence.
  5. [Section 3, DiscFknown=True] This is not a request to solve the learning problem, but the paper should not imply that the mechanism is robust to realistic information constraints.
minor comments (5)
  1. [Throughout] There are several typos and awkward phrases: 'decreseas' (p.3), 'unrelisti' (p.5), 'dependece' (p.13), 'advanteges' (p.12), 'paper paper' (p.13). Please proofread carefully.
  2. [Table 2] The row 'Your Algorithm ? ?' is unprofessional in a submitted manuscript; it appears to be a template placeholder. Please remove it and include a proper description of the proposed benchmark entry.
  3. [Figure 4] The bottom panel of Figure 4 is difficult to read because the subplot structure is not clearly labeled. Please enlarge the panels and add clear axis titles for all subplots, including the one-fleet cases.
  4. [Section 3, Experiment 1] The statement 'for further experiments we choose to set the logit parameter to 0.2' is presented as a modeling decision, but the rationale is not fully quantified. Please report the actual travel-time variance for different P values so the reader can assess the trade-off.
  5. [References] Some references lack full bibliographic details (e.g., de Almeida Correia et al. 2019 is missing volume/page if available). Please check journal style.

Circularity Check

1 steps flagged · score 6.0 of 10

Experiment 3's policy conclusion is built into the defined objective; main benchmark results are conditional and not otherwise circular.

  1. self definitional [Section 3, Experiment 3: SO component in objective (Obj_µ definition and Fig. 6 discussion)]
    "Let τ j avg := t SO avg /t j avg be the normalized average travel time and let n f,j = |I f,j |/N D be the share of fleet f on day j. Fig. 6 shows the objectives we propose (to be maximized): Obj µ = (1−µ)n f,j +µτ j avg ... This changes with increasing µ and for µ≈0.5 the plain algorithm SO− becomes competitive enough and randomization is discouraged; indeed using randomization resulting in increased average travel times reduces the objective, i.e. payoff for the fleet operator."

    Obj_μ is defined as (1−μ)n_f + μτ_avg, so μ directly controls how much the objective rewards low normalized average travel time. Since Experiment 2 already establishes that randomized algorithms increase average travel times, the claim that increasing μ makes SO− competitive and discourages randomization is a definitional property of the objective, not an emergent result. The policy recommendation 'augment the market-share objective with mean systemwide travel time' is therefore equivalent, by construction, to 'reward low average travel time if you want low average travel time.' The threshold discussion at μ≈0.5 is a numerical consequence of this built-in weighting.

full rationale

The benchmark comparisons in Experiments 1–2 are empirical simulations under explicit assumptions (two equal routes, BPR congestion, discount factors known to fleets). Using γ to route high-attitude drivers is part of the algorithm definition rather than a hidden fit; the results are honestly conditional on that assumption, and the paper itself lists unknown discount factors as future work. The self-citations (Jamróz et al. 2025; Hoffmann et al. 2025) are contextual and not load-bearing: neither is used to forbid alternatives or to supply the core efficiency claim. The clear circularity is concentrated in Experiment 3: the proposed objective is literally a weighted sum of market share and normalized average travel time, so the finding that SO−-type routing becomes competitive as μ grows, and that randomization is discouraged, is an algebraic consequence of the objective definition plus the earlier observation that randomization raises average travel time. The abstract's causal attribution of the advantage to 'unpredictable travel times' is not supported by the HDV utility model, which uses only mean travel times, but that is a model-validity concern rather than an additional circularity. Taking the main simulation comparisons at face value, the central circular step is the policy conclusion of Experiment 3, yielding a score of 6.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The model's conclusions depend on the BPR congestion form, the Gaussian discount-factor assumptions, and the known-to-fleet assumption; the heuristics contain four hand-set parameters (β, κ, σ, α) whose sensitivity is not tested. No code or data are shipped.

free parameters (4)
  • logit parameter β = 0.2
    Chosen after comparing to P=1.0 in Experiment 1; the authors state 0.2 gives 'better, stable travel times'. Selection is post hoc and directly affects the variability that drives the randomized-routing result.
  • offer discount κ = 0.5 for RFlexV-/RFlex-, 0.8 for SO-
    The paper limits further experiments to 'unrealistically short travel times' variants because the advantage is long-term only there (Experiment 1); κ values are ad hoc.
  • RFlex heuristic parameter σ = 0.4
    Set as 'a parameter to optimize' in Algorithm 4 with no sensitivity analysis; target ratio sh_target = sh*/σ.
  • credibility update rate α = 0.2
    Assumed in Eq. 4 with no justification or sensitivity analysis.
assumptions (6)
  • domain assumption BPR travel-time function t_BPR(q)=5min*(1+(q/(0.5*ND))^2) on each route (Eq. 1).
    All travel-time outcomes and algorithm calculations are generated from this congestion model.
  • domain assumption Driver discount factors are drawn from known truncated Gaussian distributions N(0.7,0.2^2) plus N(0,0.15^2).
    The diversity of attitudes, central to the claim, is assumed from a stated-preference reference and a fleet-specific term.
  • ad hoc to paper Fleet operators know every driver's discount factor (Table 1, DiscFknown=True).
    RFlexV/RFlex sort drivers by γF,f_i and identify 'unhappy' drivers using this exact knowledge; the claim would not follow if attitudes had to be inferred.
  • domain assumption HDV route choice is logit with β=0.2 while mode choice is deterministic minimum disutility.
    The behavioral model combines stochastic route choices with deterministic mode switching, a modeling choice not independently validated.
  • ad hoc to paper In RFlexV, non-fleet drivers are assumed to split 50-50 between the two symmetric routes.
    The calculation of n_faster* in Algorithm 3 depends on this split, which real HDV logit routing would not exactly produce.
  • ad hoc to paper Credibility of each fleet updates via Cred_i ← (1−α)Cred_i + α*O_i/t_i with α=0.2 (Eq. 4).
    The update rule and rate are assumed; results may depend on this learning speed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Randomized routing strategies of fleets of CAVs may prove market efficient." pith.science (2026). https://pith.science/paper/ODRHOMXK

@misc{pith2026260714859,
  author       = {Pith},
  title        = {Pith review of: Randomized routing strategies of fleets of CAVs may prove market efficient},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODRHOMXK}},
  note         = {Machine review of arXiv:2607.14859}
}
read the original abstract

In future cities every driver may own a vehicle which could be either independently driven (HDV), or autonomously routed and piloted (CAV). The autonomous operations could be handled by a few competing companies. What is the market structure which would make this market aligned with city goals? In this paper we discuss a variant of the emerging market of collectively routed fleets of CAVs, where revenue for fleet operators is proportional to market share. We provide benchmark scenarios to compare the routing algorithms. We present several routing algorithms and demonstrate that, when the attitudes of human drivers towards CAVs exhibit significant diversity, randomised CAV routing, resulting in unpredictable travel times for HDVs, is more efficient than routing proportional to system optimum/user equilibrium. Based on this, we propose to improve the design of the market by augmenting the market-share objective with mean systemwide travel time in order to limit antisocial randomised strategies of fleet operators and drive the competition towards social welfare oriented cooperation.

Figures

Figures reproduced from arXiv: 2607.14859 by the authors.

Figure 1
Figure 1. Incentives to switch to CAV. Compared to independent driving, AV technology [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Dynamics of the system at a glance. Every day each driver [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of properties of the system for different logit parameters of human [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Competition of fleet operators in a benchmark scenario. Randomized algorithms [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Usage of HDV or Fleet 0 or Fleet 1 on day [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Smoothed-out objective (50-day moving average) for different values of [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Screenshot of SUMO-based simulation experiment. The two equivalent routes [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Dependence of average travel time on routes (blue - upper route, orange - lower [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Modal split on days 1 − 200 in the SUMO-based experiment. Each circle sector depicts the chosen modes of a driver (angle 0 - day 1, angle 3π/2 - day 200). The advantage of randomised algorithms over SO- is clearly present in this setting as well. 13 [PITH_FULL_IMAGE:f…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 1 linked inside Pith

  1. [1]

    Microscopic Traffic Simulation using SUMO , year=

    Lopez, Pablo Alvarez and Behrisch, Michael and Bieker-Walz, Laura and Erdmann, Jakob and Flötteröd, Yun-Pang and Hilbrich, Robert and Lücken, Leonhard and Rummel, Johannes and Wagner, Peter and Wiessner, Evamarie , booktitle=. Microscopic Traffic Simulation using SUMO , year=

  2. [2]

    Transportation Research Part A: Policy and Practice , volume=

    On the impact of vehicle automation on the value of travel time while performing work and leisure activities in a car: Theoretical insights and results from a stated preference survey , author=. Transportation Research Part A: Policy and Practice , volume=. 2019 , publisher=

    2019

  3. [3]

    Perspectives on Politics , volume=

    Disrupting regulation, regulating disruption: The politics of Uber in the United States , author=. Perspectives on Politics , volume=. 2018 , publisher=

    2018

  4. [4]

    American Economic Review , volume=

    Disruptive change in the taxi business: The case of Uber , author=. American Economic Review , volume=. 2016 , publisher=

    2016

  5. [5]

    Yale lJ , volume=

    Amazon's antitrust paradox , author=. Yale lJ , volume=. 2016 , publisher=

    2016

  6. [6]

    Oxford Review of Economic Policy , volume=

    An invitation to market design , author=. Oxford Review of Economic Policy , volume=. 2017 , publisher=

    2017

  7. [7]

    2024 , url =

    GoldmanSachs , title =. 2024 , url =

    2024

  8. [8]

    American Economic Review , volume=

    Marketplaces, markets, and market design , author=. American Economic Review , volume=. 2018 , publisher=

    2018

Show all 12 references
  1. [9]

    arXiv preprint arXiv:2512.03524 , year=

    Market share maximizing strategies of CAV fleet operators may cause chaos in our cities , author=. arXiv preprint arXiv:2512.03524 , year=

  2. [10]

    arXiv preprint arXiv:2507.19675 , year=

    Wardropian Cycles make traffic assignment both optimal and fair by eliminating price-of-anarchy with Cyclical User Equilibrium for compliant connected autonomous vehicles , author=. arXiv preprint arXiv:2507.19675 , year=

    arXiv

  3. [11]

    Greenwade

    George D. Greenwade. The C omprehensive T ex A rchive N etwork ( CTAN ). TUGBoat. 1993

    1993

  4. [12]

    Jama cardiology , volume=

    Accuracy of wrist-worn heart rate monitors , author=. Jama cardiology , volume=. 2017 , publisher=

    2017

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.