{"id":"bd4b101f-9cfc-4bdd-ae4f-23de383b02b7","arxiv_id":"2507.04520","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"DURO uses a graph-LSTM network to output demand prediction intervals, then solves a robust rebalancing model; in single-morning NYC simulations it reduces waiting times versus deterministic baselines and runs about 60 times faster than DRO.","lead":"This paper combines a deep neural network that predicts demand uncertainty intervals with robust optimization to rebalance autonomous ride-hailing vehicles. In tests on New York City taxi data, the method matches or beats a distributionally robust baseline while solving far faster.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DURO engine is under-specified: the robust counterpart of the rebalancing model is asserted but never derived, and Equation (7) still contains the uncertain demand r, so the actual optimization solved in the experiments is undefined.","rationale":"The reader's conditional verdict already flags that the robust reformulation is asserted without derivation, but the reader's weakest_assumption focuses on benchmark parameter selection on a single test morning. My stress-test elevates the missing robust counterpart to the primary load-bearing concern: the central claim that DURO is effective and efficient depends on the actual optimization model being well-defined. If Equation (7) is not the true robust counterpart, then every reported performance number could be an artifact of an ad hoc or incorrect reformulation. This concern is concrete, addressable, and does not require new experiments in the first instance; it can be settled by an analytical derivation and code inspection. I do not move the verdict because the concern is consistent with the reader's CONDITIONAL recommendation: the paper should be accepted only if the authors supply the missing derivation and clarify the implemented model. The benchmark-selection issue remains important but secondary, as it concerns the strength and generalizability of the empirical comparison rather than the validity of the method itself.","tokens_in":17438,"tokens_out":3567,"duration_ms":43843,"concrete_test":"Independently derive the robust counterpart of problem (5) under uncertainty set (6a)-(6b): write the epigraph form min_{x,y} t subject to t >= max_{r in U} [objective(x,y,r)] and the deterministic constraints, then compute the worst-case r and any required auxiliary variables. Check whether the resulting model is equivalent to Equation (7) with r replaced by its worst-case value. If the equivalence fails or auxiliary variables are missing, request the authors' implementation code to confirm what objective and constraints were actually solved for the reported DURO experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.C presents the MIVR model with demand uncertainty set (6a)-(6b), then states that 'the robust optimization model can be reframed into the following format' (7). But Equation (7) is not a robust counterpart: it still contains the uncertain variable r_i directly in the objective, and no worst-case maximization, epigraph reformulation, or dual transformation is given. The objective coefficient on r_i is positive (through gamma * T_i), so the adversarial demand is not trivially the upper bound when the budget constraint (6b) is active; a correct robust reformulation requires nontrivial handling of the box-plus-budget uncertainty set. Without this derivation, the DURO and RO engines are not fully specified, and the numerical results in Table II and Figures 10-14 cannot be reproduced or audited. This is more foundational than the benchmark-selection concern: even a perfectly designed experiment would not validate an optimization model whose actual solved form is unknown.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DURO, a vehicle-rebalancing framework for autonomous mobility-on-demand systems in which a graph-LSTM neural network predicts distributional parameters of demand, the predicted distribution is turned into a box-plus-budget uncertainty set, and a robust optimization model is solved in a model-predictive-control loop. The authors benchmark DURO against deterministic optimization with historical-average and neural-network point forecasts, against classical robust optimization with a manually specified uncertainty set, and against a moment-based distributionally robust optimization model. Experiments on Manhattan taxi data from June 2019 report that DURO reduces passenger waiting time and leaving rate relative to deterministic baselines, is competitive with RO and DRO, and is substantially faster than DRO.","tokens_in":17674,"tokens_out":2693,"duration_ms":32943,"significance":"If the claimed results held, the paper would make a useful practical contribution: it couples a data-driven uncertainty-quantification module with robust optimization, avoids manual uncertainty-set tuning in RO, and avoids the computational cost of DRO. The idea of feeding neural-network-derived prediction intervals into a robust rebalancing model is timely and relevant. However, the supporting evidence is currently incomplete: the robust counterpart of the optimization model is not derived, the numerical comparison is based on a single 2-hour test window with parameters selected on that same window, and the DRO comparison is run under a different, reduced setting. These issues make the central claim 'on par with DRO, but with superior computational efficiency' not yet established. The paper does not provide code or machine-checked proofs, but the empirical workflow is clearly described and the authors honestly acknowledge in the conclusion that the regularity of their dataset obscures DURO's potential.","major_comments":[{"comment":"Equation (7) is not a robust counterpart of the model with uncertainty set (6a)-(6b): it still contains the uncertain demand r_i directly in the objective, and no worst-case maximization, epigraph reformulation, or dual transformation is provided. Since the objective coefficient on r_i is positive through the unsatisfied-trip term gamma * T_i, the adversarial demand is not trivially the upper bound when the budget constraint (6b) is active. The actual optimization problem solved in the experiments is therefore undefined, and the numerical results in Table II and Figures 10-14 cannot be reproduced or audited. This is a load-bearing gap that must be fixed before the method can be evaluated.","section":"Section III.C, Eq. (7)"},{"comment":"The paper reports the best-performing parameter combination (PI=75%, Gamma=9 for DURO; the best rho for RO) chosen by evaluating on the same single test morning (June 27, 7:00-9:00) that is used for all comparisons. This is selection on the test set, and with only one 2-hour window there are no repeated trials or error bars to quantify variability. The claim that DURO consistently outperforms DOHV and DONN, and is competitive with RO, is therefore not supported by the presented evidence. The authors' own conclusion that the dataset's regularity 'obscure[s] the full potential of DURO' reinforces the need for evaluation on multiple days or repeated demand scenarios before drawing general conclusions.","section":"Section IV.D, Figs. 10-13"},{"comment":"The DRO comparison is not apples-to-apples. In Table II the DRO* entry rebalances only in 'a few regions' because of computational complexity, which does not represent the same control problem as the other engines. In the later DRO comparison the number of rebalancing regions is reduced from 63 to 30, the ambiguity set is augmented with neighboring 5-minute data to achieve feasibility, and DURO is run with only the 75% and 95% intervals because the 50% interval is infeasible. These changes mean the DRO results in Fig. 14 are obtained in a different experimental setting from the other benchmark results, and the conclusion that DURO is 'on par with DRO' is not justified by the reported evidence.","section":"Section IV.D.6, Table II, Fig. 14"},{"comment":"The reformulation issue also affects the constraints: constraint (3b), y_sum <= r_i, contains the uncertain r_i, and the transition constraints (4a)-(4d) couple rebalancing and matching decisions across look-ahead intervals. Even if one accepted Eq. (7) as the objective, the robust version of the full feasible set must be specified, for example by enforcing the constraints for all r in the uncertainty set or by moving to a two-stage formulation. The paper does not state which robust semantics (here-and-now, adjustable, or constraint-wise) are being used, so the reported optimization results are unverifiable.","section":"Section III.C, Constraints (3b) and (4a)-(4d)"}],"minor_comments":[{"comment":"There are multiple typos and formatting artifacts, including 'constrains' for 'constraints', 'X-ais' for 'X-axis', 'Possion' for 'Poisson', and the unresolved LaTeX macro 'glsabk:dro' in Section IV.D.6. These should be corrected before publication.","section":"Global"},{"comment":"The objective (5) is written as a formula followed by a sentence that explains beta and gamma, but the equation itself is not numbered consistently with its reference in the text; the same applies to the display for (7). Please number all displayed equations and refer to them consistently.","section":"Section III.C, objective (5)"},{"comment":"The sentence 'this result serves as the baseline for later comparison' appears to refer to Fig. 10, but the baseline terminology is ambiguous because later comparisons report percentage reductions relative to DOHV and DONN, not to the DURO parameter sweep. Clarify what the baseline is in each comparison.","section":"Section IV.D.5"},{"comment":"The conclusion describes the uncertainty set of RO as 'momentum-based', while Section III.A uses the same phrase for the framework in [41]. This is likely a typo for 'moment-based' or 'momentum-based' in a different sense. Please align the terminology with the body of the paper.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea, but the missing derivation of the robust counterpart makes the current experimental results unauditable, and the single-window, test-set-tuned comparison is not enough to support the central performance claims. I would recommend asking the authors to (1) derive or explicitly state the robust reformulation (or provide code that defines the exact model), (2) move parameter selection to a validation set or report repeated trials with error bars across multiple days, and (3) make the DRO comparison use the same problem setting as the other benchmarks or clearly separate the two experimental regimes. The conclusion's acknowledgment of the dataset's regularity is an honest limitation statement, and it supports a major-revision rather than a rejection, but the current manuscript is not yet acceptable in its present form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the integration is sensible, but the paper never actually specifies the robust problem it solves. Equation (7) is offered as the reframed robust model, yet it keeps the uncertain demand r_i in the objective with no worst-case operator, and the stated constraint list drops (6a)-(6b). As written, the DURO and RO engines are not defined; the numerical results cannot be audited. This is a load-bearing flaw.\n\nWhat is genuinely new: the use of a probabilistic GCN-LSTM to produce demand uncertainty intervals (trained by NLL, validated with PICP/MPIW) that feed into a box-plus-budget robust rebalancing model. The robust model itself is from the authors' earlier paper [12]; the new piece is replacing manual interval choices with learned ones. That is a legitimate, incremental step. The computational-efficiency claim is credible: under 5 seconds per interval for DURO versus over 5 minutes for DRO.\n\nThe other weaknesses are real but secondary. Best-performing parameters (PI=75%, Gamma=9, best rho for RO) are selected on the same single test morning; there are no error bars or repeated trials. The DRO comparison is internally inconsistent: Table II lists a degraded DRO* while Figure 14 shows DRO beating DURO on both metrics, and the text admits DRO's flexibility gives it an edge—contradicting the abstract's 'on par' claim. The authors also concede the regular demand data obscures DURO's potential.\n\nWho it is for: researchers working on prediction-then-optimize for mobility fleets will find the pipeline worth discussing. It deserves a serious referee: the idea is new enough and the efficiency angle is practically relevant. But the referee should require a full derivation of the robust counterpart, multi-day experiments with variance reporting, and a repaired DRO comparison. Without those, the headline performance claims should not be taken at face value.","headline":"Sensible idea—learned uncertainty sets for robust rebalancing—but the paper never writes down the actual robust problem it solves: Equation (7) keeps the uncertain demand in the objective and drops the uncertainty constraints.","tokens_in":18193,"tokens_out":5829,"would_cite":true,"duration_ms":63731,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B06","90C15","90C47"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces DURO, a neural-network uncertainty-interval predictor feeding a robust rebalancing model, claiming it beats deterministic forecasts and matches slower distributionally robust optimization.","keywords":["Autonomous mobility-on-demand","vehicle rebalancing","demand uncertainty","robust optimization","distributionally robust optimization","deep learning","uncertainty quantification","graph neural networks"],"falsifier":"Fix DURO's percentile and budget along with the robust baseline's deviation parameter using only data before June 27, then run the simulator on each remaining weekday in June and on several irregular-demand days, such as rainy days or holidays. If DURO's average waiting time and leaving rate are not consistently below the deterministic engines and within DRO's range once parameters are locked, the paper's performance claim would be refuted. A cheaper check is to compute the prediction-interval coverage of the network's intervals on those held-out days: coverage far below the intended 95 percent would show the uncertainty intervals are not faithful.","tokens_in":17211,"feed_emoji":"🚕","tokens_out":9024,"duration_ms":97949,"temperature":0.7,"pith_summary":"The paper tries to establish that a fleet operator can make vehicle-rebalancing decisions under uncertain demand without hand-picking an uncertainty set or paying the computational cost of distributionally robust optimization. DURO trains a graph-convolutional LSTM network to estimate demand distribution parameters, converts them into per-region, per-time prediction intervals, and feeds those intervals directly into a robust optimization model that rebalances idle vehicles and estimates passenger matching. On New York City Manhattan weekday data, the authors report that DURO lowers average passenger waiting time relative to deterministic baselines and is on par with DRO, while solving a full 63-region rebalancing step in under five seconds versus more than five minutes for DRO. If true, real-time robust rebalancing becomes practical, and the same learned-interval-then-robustly-optimize pattern could extend to other decision problems with uncertain inputs. The paper also acknowledges that the tested dataset's regularity may hide DURO's full advantage over baselines.","feed_headline":"Neural-net demand intervals beat deterministic fleet rebalancing","feed_subtitle":"A new framework, DURO, matches slower data-driven optimization in Manhattan tests while cutting solve time from minutes to seconds.","key_machinery":"The load-bearing object is the per-region, per-time demand interval $(LB_k^i, UB_k^i)$, produced by a probabilistic graph-convolutional LSTM network whose output parametrizes a Poisson demand distribution trained by negative log-likelihood. The interval enters the downstream optimization module through a box constraint, while a polyhedral constraint with a budget parameter $\\Gamma$ limits the total demand deviation from the forecast across all regions. This machinery replaces the manually chosen $\\rho$ of classical robust optimization with a data-driven interval, preserving linear-programming tractability and avoiding the ambiguity set that makes DRO computationally heavy. The downstream module is a matching-integrated rebalancing model solved on a rolling horizon, which decides rebalancing flows and estimated matching simultaneously while penalizing unsatisfied trips.","core_discovery":"On the paper's own terms, the central claim is that demand uncertainty for vehicle rebalancing can be quantified automatically by a neural network, and that the resulting prediction intervals, when plugged into a box-plus-budget robust optimization model, produce rebalancing decisions that beat deterministic forecast-based rebalancing and match distributionally robust optimization at a fraction of its computational cost. In the NYC morning-peak experiments, DURO's average passenger waiting time is lower than both deterministic engines, historical-average demand and neural point prediction, across all simulated parameter combinations. In the head-to-head with DRO on a reduced 30-region setting, the paper reports DRO attaining somewhat lower waiting times and leaving rates, but DURO achieving comparable results once its percentile and budget parameter are chosen, while solving far faster and at full scale. The paper further claims the framework is general, applying to any decision problem with uncertain parameters, not only to autonomous mobility-on-demand rebalancing.","pith_inferences":["Beyond the paper, the same learned-interval-plus-robust-optimization pattern could be transplanted to other real-time decisions with uncertain inputs, such as ride-hailing pricing, charging-station dispatch, or warehouse task assignment, provided the downstream problem is a linear or convex program that accepts box-plus-budget uncertainty.","Beyond the paper, the performance comparison would be more decisive on days when the historical average is a poor forecast; the paper itself notes its weekdays-only, regular-demand data may hide DURO's advantage, so testing on rainy days, holidays, or weekends is a natural next experiment.","Beyond the paper, one could decouple the network's contribution from the distributional assumption by replacing the Poisson likelihood with Bayesian quantile regression, a direction the authors suggest, and comparing interval sharpness and final rebalancing metrics; this would show how much of the gain comes from learning versus from the chosen distribution.","Beyond the paper, a validation-style experiment that fixes the percentile and budget on a validation set and then evaluates on unseen test days would quantify how much tuning effort actually remains in DURO, since the reported results use the best-performing parameters selected on the test morning."],"forward_implications":["If the central claim holds, fleet operators no longer need to hand-craft robust optimization uncertainty sets: the neural network supplies per-region intervals, and only a percentile and a budget need to be set, both with clearer operational meaning than a raw deviation parameter.","A fleet running DURO on the tested 63-region Manhattan case can make a rebalancing decision in under five seconds per interval, whereas DRO required more than five minutes even after restricting rebalancing to 30 high-demand regions.","DURO's average passenger waiting time is reported lower than both deterministic engines, historical-average demand and neural point forecast, across all simulated parameter combinations in the morning-peak test.","In the direct comparison with DRO on the reduced 30-region setting, DURO is within a small performance gap once its percentile and budget parameter are chosen, and the paper frames this gap as an acceptable trade for the large computational saving.","The framework is presented as general-purpose rather than AMoD-specific: any decision problem with uncertain parameters could use the same architecture of learned uncertainty intervals upstream of a robust optimization model."],"supporting_citations":[{"why":"Supplies the robust matching-integrated rebalancing model and transition-matrix approximation that DURO's optimization module builds on.","marker":"[12]"},{"why":"Introduces the prediction-then-optimize rebalancing framework with mean-and-standard-deviation uncertainty sets that DURO automates; serves as a conceptual baseline.","marker":"[41]"},{"why":"Establishes data-driven uncertainty-set construction for robust taxi dispatch, representing the robust optimization baseline DURO benchmarks against.","marker":"[43]"},{"why":"Defines a data-driven distributionally robust vehicle-balancing approach with ambiguity sets; serves as the main performance and efficiency counterpart.","marker":"[39]"},{"why":"Provides the moment-based ambiguity set formulation used to construct the DRO engine in the comparison.","marker":"[32]"},{"why":"Defines the mean prediction interval width metric used to evaluate the quality of the uncertainty intervals produced by the prediction module.","marker":"[48]"}],"fun_headline_variants":["Neural uncertainty intervals speed up fleet rebalancing","DURO: neural net robust optimization for AMoD rebalancing","Faster robust rebalancing with neural demand uncertainty","Neural intervals beat point forecasts in fleet balancing","Robust rebalancing matches DRO with neural speed boost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the performance measured on one regular weekday morning in Manhattan, with DURO's uncertainty parameters selected on that same morning, is representative of how the method would perform on other days and less regular demand patterns.","fun_headline_variants_meta":{"raw":{"variants":["Neural uncertainty intervals speed up fleet rebalancing","DURO: neural net robust optimization for AMoD rebalancing","Faster robust rebalancing with neural demand uncertainty","Neural intervals beat point forecasts in fleet balancing","Robust rebalancing matches DRO with neural speed boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00013,"raw_usage":{"total_tokens":1138,"prompt_tokens":974,"completion_tokens":164,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":83}},"tokens_in":590,"tokens_out":164,"duration_ms":2661,"temperature":1.0,"reasoning_tokens":83,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:46:07.728056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix DURO's percentile and budget along with the robust baseline's deviation parameter using only data before June 27, then run the simulator on each remaining weekday in June and on several irregular-demand days, such as rainy days or holidays. If DURO's average waiting time and leaving rate are not consistently below the deterministic engines and within DRO's range once parameters are locked, the paper's performance claim would be refuted. A cheaper check is to compute the prediction-interval coverage of the network's intervals on those held-out days: coverage far below the intended 95 percent would show the uncertainty intervals are not faithful.","supporting_citations":[{"cited_title":"Robust matching-integrated vehicle rebalancing in ride-hailing system with uncertain demand,","cited_arxiv_id":null,"evidence_quote":"Supplies the robust matching-integrated rebalancing model and transition-matrix approximation that DURO's optimization module builds on."},{"cited_title":"Data-driven vehicle rebalancing with predictive prescriptions in the ride-hailing system,","cited_arxiv_id":null,"evidence_quote":"Introduces the prediction-then-optimize rebalancing framework with mean-and-standard-deviation uncertainty sets that DURO automates; serves as a conceptual baseline."},{"cited_title":"Data-driven robust taxi dispatch under demand uncertainties,","cited_arxiv_id":null,"evidence_quote":"Establishes data-driven uncertainty-set construction for robust taxi dispatch, representing the robust optimization baseline DURO benchmarks against."},{"cited_title":"Data-driven distributionally robust vehicle balancing using dynamic region partitions,","cited_arxiv_id":null,"evidence_quote":"Defines a data-driven distributionally robust vehicle-balancing approach with ambiguity sets; serves as the main performance and efficiency counterpart."},{"cited_title":"Distributionally robust optimization under mo- ment uncertainty with application to data-driven problems,","cited_arxiv_id":null,"evidence_quote":"Provides the moment-based ambiguity set formulation used to construct the DRO engine in the comparison."},{"cited_title":"Lower upper bound estimation method for construction of neural network-based prediction intervals,","cited_arxiv_id":null,"evidence_quote":"Defines the mean prediction interval width metric used to evaluate the quality of the uncertainty intervals produced by the prediction module."}],"review_version":1}