{"id":"4388502a-0770-47ec-8832-9363dc4ff8ad","arxiv_id":"2506.19722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An integrated optimization framework for jointly choosing routes and departure times in autonomous mobility-on-demand fleets reduces total delay by up to 25% on a Manhattan case study with real taxi data.","lead":"This paper combines two traffic control ideas for self-driving taxi fleets: spreading trips across alternative routes and delaying departures to avoid congestion. On a simulated Manhattan network, the combined approach cuts total delay by up to 25% compared to selfish routing, and both the fleet and other traffic benefit even when only a fraction of vehicles are controlled.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 25% delay-reduction headline rests on a delay function the authors admit underestimates congestion, and Theorem 1's unbiasedness applies only to a linear model, not the polynomial used in the case study.","rationale":"The paper's strongest claim is empirical, and the empirical setup's congestion model is the weakest link: it is acknowledged to understate congestion, and the theoretical bridge to Vickrey's model is proved only for a different functional form. The failure mode is concrete: if true congestion is stronger or less separable than d(f), the relative benefit of coordinated routing may shrink because both the RDUO baseline and the optimized solution are evaluated under the same surrogate. The proposed recalibration test directly probes this. The reader's CONDITIONAL verdict already captures this; the concern reinforces it rather than changing it. The algorithmic and methodological contributions are independent of the quantitative headline.","tokens_in":26820,"tokens_out":6677,"duration_ms":69935,"concrete_test":"Re-run the 31 Manhattan instances with delay parameters recalibrated so the RDUO baseline matches the reported real-world ~25% peak-hour delay-to-travel-time ratio (e.g., increase alpha or lower beta), fixing all other algorithmic choices; compute the median INTEG total-delay reduction vs RDUO. If it moves by more than 5 percentage points relative to the reported 23%, the headline is calibration-dependent. Also verify whether Theorem 1 generalizes to the polynomial delay function; if not, the theorem should not be presented as justifying the case-study numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that integrated balancing and staggering reduces total delay by up to 25% is measured inside a congestion model, not against observed traffic. The case study uses d(f) = tau*alpha*[((f+beta)/tau)^gamma - (beta/tau)^gamma] with alpha=0.1, beta=35, gamma=3, calibrated to TLC statistics. Section 5 states the resulting RDUO has 12% delay share and median trip delay 0.7 min and 'may underestimate real-world congestion.' The theoretical support (Theorem 1) is for the linear model (3.3) with phi=(2-rho)^-1, matching an M/D/1 queue; it does not cover the polynomial delay function used in the experiments, and no independent validation of the polynomial is provided. Thus the 25%/35% numbers are properties of an admittedly low-congestion surrogate; they are not established for real congestion levels, and the paper itself lists higher-congestion evaluation as future work. This directly undermines the headline quantitative claim, though not the algorithmic contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the joint optimization of route choice and departure times (balanced and staggered routing) for centrally controlled autonomous mobility-on-demand (AMoD) fleets. It formulates a mixed-integer convex program, proves that, under Poisson arrivals, the linear congestion model (3.3) with φ=(2−ρ)^{-1} reproduces the expected travel time of an M/D/1 queue (discretized Vickrey bottleneck), and develops a large-neighborhood-search metaheuristic that initializes from a reactive dynamic user optimum and a greedy construction. On a Manhattan network with roughly 6,000 trips per instance derived from TLC taxi data, the integrated algorithm reduces total delay by a median of about 23% (up to 25%) and congestion delay by up to 35% relative to the selfish RDUO baseline; a mixed-traffic analysis with 10–100% controlled vehicles reports a win-win for both AMoD and baseload traffic under welfare- and profit-oriented objectives. The paper also benchmarks the staggering-only variant against an existing matheuristic from the literature.","tokens_in":27012,"tokens_out":11093,"duration_ms":103196,"significance":"The paper makes a useful algorithmic and modeling contribution: it is one of the few studies that simultaneously optimizes route and departure-time choices at vehicle level with explicit trip-by-trip congestion, and the LNS is designed to scale to thousands of trips. The connection to the discretized Vickrey bottleneck model, though limited to a linear special case, is a nice theoretical anchor, and the machine-implemented code is publicly available. The main quantitative claims (25%/35% reductions, win-win under partial control) are, however, demonstrated only inside a congestion surrogate that the authors themselves flag as possibly underestimating real congestion, and the unbiasedness theorem does not cover the polynomial delay function used in the case study.","major_comments":[{"comment":"The unbiasedness result in Theorem 1 is established only for the linear congestion model τ(t)=τa+φτa f(t). The case study in Section 5 uses the polynomial delay function d(f)=τa·α[((f+β)/τa)^γ−(β/τa)^γ] with α=0.1, β=35, γ=3, and the algorithm is explicitly designed for any convex non-decreasing delay function. No argument is given that the polynomial inherits the Vickrey/M/D/1 calibration, and the paragraph following (3.3) does not connect φ to the polynomial. Consequently, the abstract's statement that 'our congestion model yields an unbiased estimate of travel times derived from a discretized version of Vickrey's bottleneck model' is not supported for the model actually used in the numerical evaluation. Please either extend the theoretical justification to the calibrated delay function, or explicitly present the polynomial as a heuristic extension and soften the corresponding claims in the abstract and Section 3.","section":"Section 3, Theorem 1 and Equation (3.3); Section 5, Delay function parameterization"},{"comment":"The parameters α=0.1, β=35, γ=3 are calibrated to the same TLC dataset that supplies the trip origins, destinations, and departure times used in the evaluation, and the authors state that the resulting RDUO 'may underestimate real-world congestion' (median trip delay 0.7 min, delay share 12%). Under this surrogate, the reported reductions (median 23% total delay, up to 25%; congestion delay up to 35%) are internally consistent, but they are not established for realistic congestion levels. The manuscript should include a sensitivity analysis over the delay parameters (e.g., scaling α or the congestion multiplier) to show that the direction and approximate magnitude of the reductions persist, and the abstract and conclusion should report the reductions as properties of the calibrated model unless such robustness evidence is provided.","section":"Section 5, 'Delay function parameterization'; Section 6.2, Results 2–3"},{"comment":"The win-win conclusion—that both AMoD and baseload traffic benefit regardless of the operator's objective—is derived under the assumption that baseload trips are fixed at their RDUO routes and departure times ('Baseload trips follow fixed behavior, consistent with the RDUO solution'). The final paragraph of Section 6.4 acknowledges this and defers responsive baseload behavior to future work, but the abstract and conclusion state the win-win without this qualification. Since one of the paper's motivations is congestion driven by selfish rerouting, the claim should either be explicitly conditioned on non-responsive baseload traffic or supplemented with an experiment in which baseload traffic reacts to the controlled vehicles' choices (e.g., a simple selfish rerouting rule) to test whether the win-win survives.","section":"Section 6.4, 'Flow control analysis' and Result 6"}],"minor_comments":[{"comment":"In Equations (4.1e)–(4.1f), the start-time and completion-time variables are not distinguished typographically. The conflict condition in (4.1e) should use the completion time of the preceding trip in the rightmost term, and (4.1f) should set the completion time equal to the start time plus traversal time. As printed, the constraints are self-referential.","section":"Section 4.2, Equations (4.1e)–(4.1f)"},{"comment":"The proof divides by 1−φρ when solving for E[f(t)]; at ρ=1, φ=(2−ρ)^{-1}=1, so the denominator vanishes. Please state the assumption ρ<1 or treat ρ=1 via a limiting argument.","section":"Appendix A, proof of Theorem 1"},{"comment":"The term 'unbiased estimator' is nonstandard here: Theorem 1 equates expected travel times of two stochastic processes, not an estimator's bias in the statistical sense. Consider rephrasing as 'calibrated in expectation' or 'mean-consistent'.","section":"Throughout, especially Section 3 and abstract"},{"comment":"The sentence 'commercial mapping services report that travel between the Financial District and Lower Manhattan experiences a delay of approximately 25%...' lacks a citation; please add a source or remove the claim.","section":"Section 5, 'Delay function parameterization'"},{"comment":"The LNS parameter sensitivity analysis is conducted on five days (27–31); reporting the analysis on all 31 instances would strengthen confidence in the chosen parameter values.","section":"Appendix F"}],"recommendation":"major_revision","confidential_remarks":"This is a solid algorithmic paper with reproducible code and a realistic-scale case study; the method itself is likely of interest to the AMoD and dynamic traffic assignment communities. The main risk is overclaiming: the headline percentages come from a single self-calibrated surrogate that the authors themselves describe as potentially underestimating congestion, and the theoretical unbiasedness result does not cover the calibrated delay polynomial. I recommend requiring a robustness analysis around the congestion parameters and a careful qualification of the win-win statement before acceptance. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Coppola et al. integrate balanced and staggered routing for AMoD into one framework, and the algorithmic work is solid. The headline delay numbers, though, are computed inside a congestion model the authors admit underestimates real congestion, so treat them as indicative of the mechanism, not as a field measurement.\n\nWhat's new: prior work did staggering or balancing separately; here they're jointly optimized. The explicit modeling of baseload trips as individual trips instead of arc capacities is a genuine step forward. The LNS metaheuristic with incremental schedule updates is the real workhorse, and they benchmark it against the earlier matheuristic and show large gains in median delay reduction. Code and data are public. The Theorem 1 proof that the linear congestion model with phi=(2-rho)^-1 matches M/D/1 expected delay is correct, if narrow.\n\nSoft spots, in proportion. First, Theorem 1 covers only the linear delay model (3.3), but the case study uses the polynomial d(f) = tau*alpha*[((f+beta)/tau)^gamma - (beta/tau)^gamma] with alpha=0.1, beta=35, gamma=3. The unbiasedness result does not apply to that function. The abstract's phrasing suggests the congestion model generally is unbiased, but the proof is only for the linear instance. That mismatch should be fixed or clarified. Second, the delay parameters are calibrated to the same TLC data used for evaluation, and the paper itself states the RDUO delay share (about 12%, median trip delay 0.7 min) 'may underestimate real-world congestion.' So the 25% and 35% improvements are properties of a deliberately low-congestion surrogate. The relative comparison between their algorithm and RDUO under the same model is fair, but the absolute headline numbers are not validated. Third, the win-win result in the mixed-traffic analysis is presented for a single representative instance (day 21), not across all 31. That's weaker than the claim suggests.\n\nThese are limitations, not fatal flaws. The core idea holds: jointly optimizing routes and departure times reduces delay in the model, and the algorithmic machinery is a credible contribution. The paper is honest about the congestion underestimation and lists higher-congestion tests as future work.\n\nWho it's for: anyone working on AMoD fleet management, dynamic traffic assignment, or large-scale routing metaheuristics. It deserves a serious referee. I'd send it out with requests to clarify the scope of Theorem 1, report flow-control results across all instances, and ideally stress-test the delay function at higher congestion levels to see if the relative gains persist.\n\nRecommendation: engage with it; accept for peer review.","headline":"Solid algorithmic contribution integrating balanced and staggered routing, but the headline delay reductions are properties of a congestion model the authors admit underestimates reality, so treat them as indicative, not validated.","tokens_in":27588,"tokens_out":3072,"would_cite":true,"duration_ms":30689,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B20","90B06","90C59"],"pacs":[],"model":"deepseek-v4-flash","headline":"A centrally controlled fleet that jointly chooses routes and departure times cuts total traffic delay by up to 25 percent and congestion delay by up to 35 percent compared with selfish routing, and partial control still helps everyone.","keywords":["autonomous mobility-on-demand","balanced routing","staggered routing","congestion","large neighborhood search","Vickrey bottleneck model","route and departure time choice","Manhattan taxi data"],"falsifier":"Re-run the 31 Manhattan instances with the delay function re-fitted to the observed median peak delay of about 1.5 minutes rather than the aggregate fit that yields a 0.7-minute median, and compare the integrated solution with the selfish baseline; if the delay reduction drops below 23 percent or the congestion reduction disappears, the calibrated-congestion assumption is the cause.","tokens_in":1805,"feed_emoji":"🚦","tokens_out":3330,"duration_ms":127479,"temperature":0.7,"pith_summary":"This paper tries to establish that a centrally coordinated autonomous mobility-on-demand fleet should decide each trip's route and departure time in a single optimization, rather than handling the two decisions separately. On a Manhattan street network with roughly six thousand real taxi trips per instance, the integrated policy reduces total delay by about 23 percent in median and up to 25 percent, and congestion delay by up to 35 percent, relative to selfish routing. The paper also claims a formal grounding: its arc-delay function is an unbiased estimator of travel times under a discretized Vickrey bottleneck model. In mixed traffic, a welfare-oriented controller and a profit-oriented fleet controller both produce a win-win outcome, lowering delays for controlled and uncontrolled vehicles alike. If true, the result gives a practical upper bound on what central control of ride-hailing and robotaxi fleets could achieve before online uncertainty is introduced.","feed_headline":"Joint routing and timing cuts city traffic delay by 25%","feed_subtitle":"Combined balancing and staggering cuts congestion by 35 percent; partial control still helps everyone.","key_machinery":"The carrying mechanism is a joint route-and-departure-time optimization model in which every trip chooses from a precomputed set of limited-overlap alternative routes and a feasible departure time, while arc travel times are $\\tau_a + d(f^r_a)$ with $d$ any convex, non-decreasing delay function. The paper proves (Theorem 1) that when arrivals are Poisson with traffic intensity $\\rho$, the linear delay law $\\tau(t)=\\tau_a+(\\phi\\tau_a)f(t)$ with $\\phi=(2-\\rho)^{-1}$ reproduces the expected travel time of a discretized Vickrey bottleneck model, giving a queueing-theoretic calibration for the delay-function family used in computation. In the experiments that family is realized as the polynomial $d(f^r_a)=\\tau_a\\alpha[((f^r_a+\\beta)/\\tau_a)^\\gamma-(\\beta/\\tau_a)^\\gamma]$. Around this model sits a large-neighborhood-search metaheuristic whose insert, remove, local-search, and schedule-update operators evaluate route and staggering changes efficiently at the scale of thousands of trips.","core_discovery":"The central claim is that balanced routing and staggered routing are complements: spreading trips over alternative routes relieves spatial congestion, while delaying departures smooths temporal peaks, and optimizing both together yields delay reductions that neither achieves alone. The evidence is an algorithm that assigns each trip one of five limited-overlap alternative routes and a departure time within its allowable staggering window, minimizing total fleet travel time under a convex non-decreasing arc-delay function. In the full-control Manhattan experiments, the algorithm's median total delay reduction is 23 percent (10 to 17 hours saved), congestion delay falls by up to 35 percent, and the two mechanisms combine in a nearly additive way. The paper further claims that even a 10 percent controlled share captures about a quarter of the maximum delay reduction, and a 50 percent share captures three-quarters, in both welfare- and profit-oriented settings.","pith_inferences":["Because the model assumes full advance knowledge of demand and precomputed route sets, an online controller that learns requests as they arrive would likely recover a smaller share of the 25 percent; a rolling-horizon reoptimization on the same data would quantify the gap.","The authors note their delay parameterization may underestimate real congestion. If the true baseline congestion were closer to the observed 1.5-minute median delay, central coordination might produce larger absolute savings, but the relative 23 percent figure is not guaranteed to carry over.","The mixed-traffic win-win treats conventional traffic as fixed. If human drivers reroute in response to the fleet's new patterns, the gains to baseload traffic could change; simulating responsive baseload is the natural next experiment.","The unbiasedness theorem suggests the same delay-function form could be transferred to other cities by estimating $\\rho$ or recalibrating $\\alpha$, $\\beta$, and $\\gamma$ from local link data, without re-deriving the algorithm."],"forward_implications":["Under full centralized control, median total delay falls by roughly 23 percent on Manhattan-scale peak instances, corresponding to 10 to 17 hours saved per day.","Network congestion delay falls by up to 35 percent, and the integrated policy removes more congestion delay than balancing alone while adding less detour delay.","In mixed traffic, controlling only 10 percent of vehicles already yields about 25 percent of the maximum delay reduction, and 50 percent control yields about 75 percent.","Both welfare-oriented and profit-oriented operators produce lower delays for controlled and uncontrolled traffic at every tested control level.","The staggering-only variant beats the MILP-based matheuristic in delay reduction and robustness, giving a scalable building block for large instances."],"supporting_citations":[{"why":"Supplies the discretized Vickrey bottleneck model whose expected travel times the paper's congestion model is shown to match.","marker":"Otsubo & Rapoport 2008"},{"why":"Provides the M/D/1 queueing analysis used to compute expected bottleneck delay and identify the calibration $\\phi=(2-\\rho)^{-1}$.","marker":"Thomopoulos 2012"},{"why":"Introduces the bottleneck congestion framework that motivates the model of queueing delay at arcs.","marker":"Vickrey 1969"},{"why":"Defines the staggering-only problem and its MILP matheuristic, which serves as the NP-hardness source and as the baseline the new algorithm outperforms.","marker":"Coppola et al. 2025"},{"why":"Provides the SVP+ algorithm used to generate the limited-overlap alternative route sets that balancing assigns trips across.","marker":"Chondrogiannis et al. 2020"},{"why":"Supplies the real taxi trips whose origins, destinations, and departure times define the 31 case-study instances.","marker":"NYCTLC 2009"},{"why":"Supplies the hub-bound traffic estimate used to scale the number of trips in the study area.","marker":"NYMTC 2023"}],"fun_headline_variants":["Routing plus timing combo cuts delays up to 25%","Balanced and staggered rides slash congestion by 35%","Even partial AMoD control reaps large traffic benefits","Joint route-time optimization wins for all drivers","Two routing tricks together beat selfish driving"],"cache_read_input_tokens":29696,"weakest_assumption_plain":"The load-bearing premise is that the fitted polynomial delay curve captures how congestion actually builds on these streets; the authors note this may understate real-world congestion, so the reported reductions could shrink under stronger or differently shaped congestion.","fun_headline_variants_meta":{"raw":{"variants":["Routing plus timing combo cuts delays up to 25%","Balanced and staggered rides slash congestion by 35%","Even partial AMoD control reaps large traffic benefits","Joint route-time optimization wins for all drivers","Two routing tricks together beat selfish driving"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1366,"prompt_tokens":983,"completion_tokens":383,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":309}},"tokens_in":599,"tokens_out":383,"duration_ms":4137,"temperature":1.0,"reasoning_tokens":309,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:26:27.056984+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the 31 Manhattan instances with the delay function re-fitted to the observed median peak delay of about 1.5 minutes rather than the aggregate fit that yields a 0.7-minute median, and compare the integrated solution with the selfish baseline; if the delay reduction drops below 23 percent or the congestion reduction disappears, the calibrated-congestion assumption is the cause.","supporting_citations":[{"cited_title":", & author Rapoport, A","cited_arxiv_id":null,"evidence_quote":"Supplies the discretized Vickrey bottleneck model whose expected travel times the paper's congestion model is shown to match."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the M/D/1 queueing analysis used to compute expected bottleneck delay and identify the calibration $\\phi=(2-\\rho)^{-1}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the bottleneck congestion framework that motivates the model of queueing delay at arcs."},{"cited_title":", author Hiermann, G","cited_arxiv_id":null,"evidence_quote":"Defines the staggering-only problem and its MILP matheuristic, which serves as the NP-hardness source and as the baseline the new algorithm outperforms."},{"cited_title":", author Bouros, P","cited_arxiv_id":null,"evidence_quote":"Provides the SVP+ algorithm used to generate the limited-overlap alternative route sets that balancing assigns trips across."},{"cited_title":"title TLC Trip Record Data","cited_arxiv_id":null,"evidence_quote":"Supplies the real taxi trips whose origins, destinations, and departure times define the 31 case-study instances."},{"cited_title":"title Hub bound travel data 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the hub-bound traffic estimate used to scale the number of trips in the study area."}],"review_version":2}