{"id":"92912e68-f4d1-4a19-9312-80c01dae2de6","arxiv_id":"2507.15307","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A CNN that predicts binary routing and charging variables, plus a zero-padding trick for varying fleet sizes, reduces solve time for an EV routing-scheduling MIP by about 98% on synthetic test instances.","lead":"Researchers trained a convolutional neural network to guess the hard binary decisions in an electric-vehicle routing and charging optimization problem, letting the Gurobi solver finish much faster. On a 33-bus grid with 20 to 100 EVs, the best model cut solve time by 97.8% while staying close to the solver's own solution quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed <0.01% optimality gap is unsubstantiated: baseline Gurobi stops at 0.1% MIP gap, so the reported -0.01% loss only shows parity with an approximate reference.","rationale":"The reader's CONDITIONAL verdict is appropriate, and my concern does not move it, so I keep UNCHANGED. The paper has real strengths: the padding mechanism is a concrete contribution, the authors report all four CNN variants rather than hiding the best-model selection, and the limitation section is honest. However, the headline optimality claim needs a high-accuracy benchmark; with a 0.1%-gap reference, 'less than 0.01% from optimal' is not verifiable from Table I. I partially disagree with the reader's stated weakest_assumption: the deterministic-to-stochastic transfer is not an unjustified assumption, because the stochastic MIP is separable by scenario (all constraints and the objective carry a per-scenario index with no coupling), so per-scenario prediction in Algorithm 1 is exact. The same separability, though, means the 'stochastic' formulation is weaker than a true here-and-now day-ahead decision model. The proposed concrete test would settle whether the central optimality number actually holds.","tokens_in":920,"tokens_out":2157,"duration_ms":101453,"concrete_test":"Take 20-30 test instances spanning the EV-count range 20-100, including values outside the training multiples. Solve each with Gurobi from scratch with MIPGap=0.01% (or 0.001%) and no artificial time limit to obtain a near-optimal benchmark. Run the CNN5-assisted pipeline on the same instances and compute (assisted_obj - benchmark_obj)/benchmark_obj for each. If any instance exceeds 0.01%, or the mean exceeds 0.01%, the Abstract's optimality claim is not supported. Also record the actual final MIPGap for the unassisted baseline: if it terminated on the 120-min timeout with a gap above 0.1%, the reported l-bar and the 97.8% runtime reduction are also affected (unassisted time is capped).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the optimality claim in the Abstract and Table I. Section IV-A sets MIPGap=0.1% (or a 120-minute timeout) when labelling the datasets, and the unassisted Gurobi baseline used for comparison appears to be subject to the same termination criterion. Table I reports l-bar = -0.01% for CNN5, meaning the assisted solution is about 0.01% better than the reference solution. Since the reference is itself only certified to within 0.1% of the true optimum, the abstract's statement 'deviating less than 0.01% from the optimal solution' is not established: the comparison shows near-equality with a 0.1%-approximate baseline, not 0.01% optimality. Furthermore, fixing CNN-predicted binaries restricts the feasible region, so even a 0.1%-gap solve of the reduced MIP can be more than 0.1% from the original optimum; the reported metric does not bound this. The reader's nominated weakness (deterministic-to-stochastic transfer) is not actually load-bearing: objective (1) and constraints (2)-(20) are all indexed by sc and contain no cross-scenario or non-anticipativity constraints, so the stochastic MIP is scenario-separable and per-scenario prediction in Algorithm 1 is exact. What remains is that the headline 'near-optimal' number is anchored to a loose baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a deep learning-assisted approach to accelerate the solution of a day-ahead electric vehicle (EV) joint routing and scheduling (JRS) problem. The problem is formulated as a stochastic mixed-integer program (MIP) with scenario-based solar uncertainty, and the routing component uses a time-space network extended with virtual congestion nodes (VCNs) to model time-varying travel times. A convolutional neural network (CNN) is trained to predict the binary routing and charging variables, using a padding mechanism so that a single model can handle varying fleet sizes without retraining. The predicted binaries are thresholded and then fixed, and Gurobi solves the reduced problem. In a case study on the IEEE 33-bus distribution network coupled with the Nguyen-Dupuis transportation network, the best model (CNN5) achieves a 97.83% average runtime reduction, a 99.5% feasibility rate, and an average optimality loss of -0.01% relative to a Gurobi baseline.","tokens_in":12838,"tokens_out":7227,"duration_ms":75703,"significance":"The paper addresses a timely and practically important problem—fast, near-optimal scheduling of EV fleets for grid support. The proposed CNN-assisted solution with padding for variable EV counts is a sensible approach to a real obstacle in applying learning-based acceleration to problems with changing dimensions. The VCN modification to the time-space network is a useful modeling contribution. The experimental study is comprehensive in terms of model variants (four training intervals) and includes an ablation of the padding mechanism. However, the headline near-optimality claim is not supported by the experimental design, because the reference 'optimal' solutions are themselves only approximated to a 0.1% MIP gap. This limits the strength of the paper's central contribution until the claim is either re-scoped or replaced with a certified optimality evaluation.","major_comments":[{"comment":"The abstract claims that the approach deviates 'less than 0.01% from the optimal solution,' but this is not established by the experiments. Section IV-A states that Gurobi labels the datasets with a MIPGap of 0.1% or a 120-minute timeout, so the reference solutions are approximate to within 0.1% of the true optimum. Table I reports an optimality loss l-bar of -0.01% for CNN5, which means the assisted solution is 0.01% better than this approximate reference, not that it is within 0.01% of the true optimum. Moreover, fixing CNN-predicted binaries restricts the feasible region, so even solving the reduced MIP to a 0.1% gap does not bound the gap to the original optimum. Please either solve a subset of test instances to certified optimality (or with a much tighter MIPGap) and report the resulting suboptimality, or revise the abstract and conclusion to state that the solution is 'within 0.1% of the solver reference' rather than 'within 0.01% of the optimal solution.'","section":"Abstract; Section IV-A; Table I"},{"comment":"The runtime comparison is incompletely specified. The paper does not state the termination criterion used for the unassisted Gurobi baseline in the runtime experiments. If the baseline is not given the same MIPGap=0.1% and 120-minute time limit used for labelling, the reported 97.8% speedup could be inflated by instances where the baseline hits the time limit. If the baseline does use the same limits, then the comparison is between two approximate solves and should be described accordingly. Please specify the exact solver parameters for the baseline and, ideally, report time-to-given-optimality-gap curves for both the assisted and unassisted solvers.","section":"Section IV-C; Fig. 6"}],"minor_comments":[{"comment":"The abstract reports a 99.5% feasibility rate, but this is achieved by only one of the four trained models (CNN5); Table I shows CNN10, CNN15, and CNN20 have feasibility rates of 97.5%, 95.5%, and 77.0%, respectively. The abstract should either qualify this as the best-case result or acknowledge the sensitivity of feasibility to the training interval.","section":"Abstract"},{"comment":"The thresholding logic in Algorithm 1 is unclear. The condition 'if p_t0 >= y_hat >= p_t1 then Dispose' uses scalar threshold values p_t0 and p_t1, but the text describes 'extracting predictions with probabilities above or below a threshold.' With p_t0≈0.99 and p_t1≈0.70, the condition disposes predictions in [0.70, 0.99], which would include many high-confidence positive predictions, while predictions below 0.70 are rounded to 0. Please clarify whether the intended comparison is y_hat > p_t1 for rounding to 1 and (1 - y_hat) > p_t0 for rounding to 0, and correct the pseudocode accordingly.","section":"Section III-C; Algorithm 1"},{"comment":"In equation (11), the summation range is written as 't = 2, ..., t', which should be 't = 2, ..., |T|'.","section":"Section II-C, Eq. (11)"},{"comment":"The description of virtual congestion nodes (VCNs) would benefit from a small illustrative example or figure showing how a VCN extends a trip time by one or more timespans. As written, the distinction between 'destination nodes', 'virtual nodes', and 'VCNs' is easy to lose.","section":"Section II-A"},{"comment":"The paper states that 'Code for reproducing experiments will be available at' a GitHub URL. The repository was not available at the time of review. Please ensure the code is released before publication to support reproducibility.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim of 'less than 0.01% optimality loss' is not supported because the baseline is only approximate to 0.1%. This is fixable by adding a certified optimality evaluation on a subset of instances or by re-scoping the claim. The runtime comparison also needs clarification of the baseline's termination criteria. The methodological core—CNN-assisted binary prediction with padding for variable fleet sizes—is sound and within the journal's scope. No concerns about citation practices or novelty disclosure beyond the need to fully specify the experimental setup."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know up front. The padding mechanism that lets one CNN handle 20-100 EVs without retraining is a real, practical extension of the ML-warm-start-MIP program that has been applied to unit commitment; that part is the genuine new chip on the table. But the abstract's 'deviating less than 0.01% from the optimal solution' is not actually established, because the reference solutions used for comparison were computed by Gurobi with a 0.1% MIP gap or 120-minute timeout. A -0.01% loss against that baseline just says the assisted solve matches an approximate reference. Fixing predicted binaries also restricts the feasible region, so a 0.1%-gap solve of the reduced problem is not a 0.1% bound on the original MIP. The headline number overstates what Table I really shows.\n\nWhat the paper does well: the scenario-separability point—the reader's worry about deterministic-to-stochastic transfer—turns out to be a non-issue. Looking at constraints (2)-(20), everything is indexed by sc and there are no non-anticipativity or cross-scenario constraints, so per-scenario binary prediction and concatenation is structurally exact. I checked that after the stress-test note and the stress-test is right. The VCN modification to the time-space network is a small but sensible modeling addition that lets travel times vary by time period. The paper also reports results for four different training intervals, which is a decent generalization study, and the limitation section is honest about redistributing computational burden and about the lack of feasibility guarantees.\n\nSoft spots, in proportion. The optimality claim is the load-bearing one and it needs reworking; either solve references to a tighter gap (0.01% or 0%) on a subset, or rephrase the claim to 'matches the solver's 0.1%-gap solution.' The other issues are minor: no code or data actually available at the time of writing, no variance over repeated runs of the same test set, and the padding ablation uses estimated labelling times from a sample of 10 instances per EV count. CNN20's 77% feasibility shows the approach degrades with wider training intervals, and the paper says so—credit where due—but the abstract's selected best-case numbers hide that.\n\nWho should read this: anyone working on ML-assisted MIP for EV routing-scheduling or similar day-ahead problems. It is an incremental engineering contribution, not a new capability. With a corrected optimality metric and released code, it would be a solid journal paper. I'd send it to review; a serious referee can fix the overclaim without killing the contribution.","headline":"Solid ML-for-MIP acceleration with a genuinely useful padding trick, but the near-optimality claim rests on a 0.1%-gap baseline and should be restated or re-benchmarked.","tokens_in":13350,"tokens_out":3272,"would_cite":true,"duration_ms":36733,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A CNN that predicts the binary variables of an EV fleet's day-ahead routing-and-scheduling problem lets a MIP solver finish 97.8% faster, with 99.5% feasibility, under 0.01% optimality loss, and no retraining for fleet sizes between 20…","keywords":["Electric Vehicles","Power system simulation","Optimization","Neural networks","joint routing and scheduling","mixed-integer programming","variable fleet size"],"falsifier":"Generate test instances where the optimal binary routing decision in one solar scenario conflicts with another scenario's optimal decision at the same timespan, then run the paper's training and inference pipeline; if feasibility falls well below 99.5% or the optimality loss rises above 0.01%, the concatenated-prediction assumption is the bottleneck.","tokens_in":12319,"feed_emoji":"⚡","tokens_out":10173,"duration_ms":99818,"temperature":0.7,"pith_summary":"This paper claims that a deep convolutional network can take over the hard binary decisions in a day-ahead electric-vehicle routing-and-scheduling problem, converting a mixed-integer program that is slow into one a solver can handle quickly. The network is trained on single-scenario problems but applied per scenario at run time, with a zero-padding scheme that lets it accept any fleet size between 20 and 100 without retraining. If correct, this gives grid operators a way to coordinate EV charging, discharging, and routing with renewable generation inside the clearing window of a day-ahead energy market. The paper reports a 97.8% runtime reduction against an unassisted MIP solver while keeping 99.5% of test instances feasible and deviating less than 0.01% from the optimal cost.","feed_headline":"Neural net predicts EV routing binaries, cutting solve time 97.8%","feed_subtitle":"A single model handles fleets of 20 to 100 EVs with 99.5% feasibility and under 0.01% optimality loss.","key_machinery":"The load-bearing object is a convolutional neural network that takes the solar, load, and job-schedule time series as input and outputs probabilities for every binary variable of the MIP. A thresholding step keeps only variables the network predicts with high confidence, passing the rest to the solver, while a padding mechanism sizes the network for a maximum fleet size $e_{\\max}$ and zero-fills missing EV rows so the same weights serve any fleet size from 20 to 100. This combination lets one trained model serve the tested fleet range and hands the solver only the genuinely ambiguous binaries. The formulation side is carried by a modified time-space network with virtual congestion nodes (VCNs) that allow arc travel times to lengthen or shorten with traffic conditions, so the routing variables the CNN predicts are time-dependent.","core_discovery":"The central claim is that a CNN trained to predict the binary variables of the EV joint routing and scheduling (JRS) MIP can be used as a warm start for the solver, shrinking the search space so that the remaining variables are found quickly. The paper introduces a deterministic-training, stochastic-inference workflow: because labelling full stochastic instances is prohibitively expensive, the CNN is trained on deterministic versions and then applied to each of the five solar scenarios independently, concatenating the predicted binaries into a single solution. A padding mechanism zero-fills the input and output rows of missing EVs, so the same network handles fleets of size 20 to 100 with no retraining. On the IEEE 33-bus distribution system combined with the Nguyen-Dupuis transportation network, the assisted solver achieves 97.8% runtime reduction, 99.5% feasibility, and a mean optimality deviation of -0.01% relative to the unassisted MIP baseline.","pith_inferences":["The deterministic-to-stochastic transfer is the most fragile link; comparing against a model trained on a handful of full stochastic labels would reveal how much of the reported 0.01% gap is transfer loss rather than network capacity.","The zero-padding scheme is generic: any MIP with a variable number of homogeneous agents whose binary variables scale linearly with the agent count could reuse it, so the method may carry over to unit commitment with flexible generator counts or to multi-robot task allocation.","The paper states that the computational burden shifts to labelling; that suggests reinforcement learning or self-supervised pretraining on MIP dual information could replace some supervised labelling and improve scalability.","Because the optimality loss is so small, operators might skip the feasibility-retry loop and accept the CNN's binaries outright, trading a tiny cost penalty for predictable, near-constant solve times."],"forward_implications":["Runtimes for day-ahead EV coordination drop from hours to minutes, putting the problem inside the typical 3-4 hour clearing window of day-ahead energy markets.","A single trained model covers EV fleet sizes from 20 to 100, so operators avoid generating a fresh labelled dataset and retraining for each new fleet count.","The 99.5% feasibility rate, together with the threshold-raising retry loop in the paper, gives a practical fallback when the predicted binary solution is infeasible.","The near-zero optimality loss indicates that the binary structure of the JRS problem can be learned from local input features, which is a prerequisite for using learned warm starts in other power-system MIPs."],"supporting_citations":[{"why":"Supplies the joint routing and scheduling MIP formulation that this work extends with time-varying travel times and stochastic solar scenarios.","marker":"[11]"},{"why":"Defines the Nguyen-Dupuis transportation network used as the test transportation system in the case study.","marker":"[22]"},{"why":"The commercial MIP solver used both to label training instances and as the unassisted baseline for runtime and optimality comparisons.","marker":"[23]"},{"why":"Provides the time-space network formulation that the paper adapts by adding virtual congestion nodes to model dynamic travel times.","marker":"[24]"},{"why":"Supplies the LinDistFlow linearized power-flow model that couples EV charging/discharging to distribution-network voltages and line flows.","marker":"[25]"},{"why":"Gives the class-probability formula used by the thresholding filter to decide which predicted binary variables are retained for the solver.","marker":"[27]"}],"fun_headline_variants":["CNN predicts binaries, speeding EV routing solver by 97.8%","Deep learning cuts EV scheduling solve time 97.8% with 99.5% feasibility","One neural net handles EV fleets 20-100, near-optimal, 97.8% faster","AI warm-start shrinks search space for EV routing and scheduling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole speed-up rests on the assumption that a network trained on simplified single-scenario problems can still propose binary decisions that, when pasted together across five solar scenarios, remain jointly feasible and near-optimal for the full stochastic problem, without any explicit justification or ablation in the paper.","fun_headline_variants_meta":{"raw":{"variants":["CNN predicts binaries, speeding EV routing solver by 97.8%","Deep learning cuts EV scheduling solve time 97.8% with 99.5% feasibility","One neural net handles EV fleets 20-100, near-optimal, 97.8% faster","AI warm-start shrinks search space for EV routing and scheduling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1404,"prompt_tokens":974,"completion_tokens":430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":340}},"tokens_in":590,"tokens_out":430,"duration_ms":4674,"temperature":1.0,"reasoning_tokens":340,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:35:21.450806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate test instances where the optimal binary routing decision in one solar scenario conflicts with another scenario's optimal decision at the same timespan, then run the paper's training and inference pipeline; if feasibility falls well below 99.5% or the optimality loss rises above 0.01%, the concatenated-prediction assumption is the bottleneck.","supporting_citations":[],"review_version":1}