acceptodds
Under review as a conference paper at ICLR 2027

From Real Roads to Depot-Scale Routing: A Dataset and Benchmark for Electric Vehicle Routing with Time Windows

Abstract

Solvers for the Electric Vehicle Routing Problem with Time Windows (EVRP-TW) are commonly evaluated on small synthetic benchmarks. Whether conclusions drawn from such benchmarks generalize to operational workloads on real road networks remains unclear. We introduce EVRP-TW-D/B, a public dataset-and-benchmark suite. EVRP-TW-D provides 71,500 instance views across 11 U.S. cities and 100–2,000 customers. Its two-stage generator builds reusable city environments from directed road networks and instantiates delivery days using joint demand, service-time, and time-window tuples from Amazon records. EVRP-TW-B compares five learning-based policies with exact and metaheuristic baselines using a common task and objective, released implementations, independent verification, and declared budgets. With 100-customer Euclidean training, only two policies achieve lower mean costs than the best evaluated search baseline under a 30-minute budget at 500 and 1,000 customers. Training the same architectures on 100-customer EVRP-TW-D instances yields full feasibility and lower mean costs than this baseline at both scales. Policies trained at 1,000 customers retain this advantage beyond the training geography. These findings motivate evaluating solver competitiveness across operational workloads and geographies, with EVRP-TW-D/B providing a reproducible foundation.

Primary area:
datasets and benchmarks
Keywords:
electric vehicle routing, time windows, road networks, datasets and benchmarks, neural combinatorial optimization, generalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yjwDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a large-scale, road-grounded corpus for EVRP-TW and a benchmark for it.

Critical strengths

The dataset scale of this paper is far larger than that of previous works.

Critical weaknesses
  1. One important difference between the proposed dataset and previous ones is the travel representation. But actually, users themselves can get it through the real road network (e.g., from OSRM) and the coordinates, which has been widely used by many ride-hailing simulators.

  2. As a dataset and benchmark paper, the data seems all collected from websites instead of new ones, and the compared methods are limited without standard deviation provided.

  3. Even though there are some findings through the evaluation, most of them are very intuitive without providing enough new insight.

Questions for authors
  1. What is the reason for choosing these methods in your benchmark?

  2. What are the core differences between this work and previous EVRP-TW dataset works that should make this paper be considered for publication in ICLR?

5 comments
AuthorsOpenReview ↗

Response to Reviewer yjwD (W1/W2/Q2): Dataset Contribution and Construction

Response to Reviewer yjwD

1. W1 / W2 (data) / Q2 — What does the resource contribute beyond public inputs and road matrices?

Thank you for recognizing the scale of the dataset. EVRP-TW-D/B combines a road-grounded task corpus, a reusable delivery-day generator, and a shared evaluation protocol with released reference implementations. This combination supports policy training and comparisons across new days, workload sizes, held-out communities, and an unseen city. Table 1 positions these capabilities relative to existing resources; Sections 4–5 describe their construction and evaluation roles.

We agree that road distances between given coordinates can be computed with tools such as OSRM, and Table 1 lists an existing generator with directed-road travel. The contribution is the routing task built on those roads: which locations are active customers, their joint demand, service-time, and time-window tuples, the charging roster selected for each delivery day, single-customer feasibility certificates, and fixed splits for new days, held-out communities, and an unseen city. The public inputs and construction procedures are documented for reproducibility, and redistribution follows the component-specific licenses (Appendix D.3).

The 71,500 benchmark views span 100–2,000 delivery locations in 11 cities and share 7,500 parent families. Constructing these tasks links residential service candidates and facilities to directed roads, combines reference speeds with road topology, and transfers intact package-volume/service/window tuples under feasibility checks. Appendix C.1 identifies each public source's role; C.3–C.6 specify speed construction, location/facility layers, and instance materialization. C.4 documents the building-source reconciliation and residential-unit assignment used for the benchmark.

The nested views and held-out partitions support training and tests at distinct workload sizes and geographic settings (Appendix C.7). EVRP-TW-B adds a common task, cost objective, budgets, and independent replay so solver outputs can be compared consistently (Section 5; Appendix E). Code, checkpoints, and sample data are provided with the submission; the full corpus will be released at camera-ready. The implementation descriptions distinguish author-code adaptations from paper-guided local implementations.

AuthorsOpenReview ↗

Response to Reviewer yjwD (W2/Q1): Method Selection and Statistical Reporting

2. W2 (methods and statistics) / Q1 — Why these methods, and how reliable are the comparisons?

The evaluated set covers one exact model, four heuristics, and five neural designs, and standard deviations are now reported. We added HMA and RoutingBlocks in the revision.

FamilyMethodsImplementation sourceRole in the comparison
ExactGurobiOur MILP formulation (Appendix A.2)Solves each instance within the budget
HeuristicVNS/TS, ALNSLocal implementations based on the published descriptionsEstablished search designs
Heuristic (added)HMA, RoutingBlocksAuthors' C++ solver (HMA); public Python/C++ framework (RoutingBlocks)Stronger search references
NeuralAM-EVRPTW, TERRANAdapted from released codeOmit explicit pairwise travel inputs
NeuralEVRPTW-RLPaper-guided implementationAggregates travel-time rows per node
NeuralDRL-TS, RRNCOPaper-guided implementation (DRL-TS); adapted from released code (RRNCO)Retain explicit pairwise travel relations

The neural policies also differ in attention/recurrent architecture and training procedure. Appendix E.2 identifies each method's source, adaptations, and restrictions.

The complete comparison uses the same 500 T1 instances at N=1000. Exact and heuristic runs receive 1,800 seconds, including initialization for heuristics; neural policies use fixed D1000 checkpoints and 30 trajectories. VNS/TS and ALNS share identical initial routes, RoutingBlocks uses its own ALNS Init construction, and HMA uses native initialization (Appendix E.3).

Reported statisticWhere
Mean cost with SDTables 3 and 5; Appendix F.2 for the configurations in Tables 2 and 4
Median, P90, P99, and paired winsAppendix F.1
Search progress at 60–1,800 secondsAppendix F.1
DefinitionsAppendix E.5

SD is the sample SD across the 500 test instances for a fixed configuration. Main-text mean cost is reported only for fully feasible cohorts. For example, HMA's mean/SD is 1,945.47/438.08 versus DRL-TS's 1,997.51/430.31; Gurobi returns accepted incumbents on 2/500 tasks within the budget.

Initialization controls address the vehicle count in the submitted VNS/TS result: its 683.340-vehicle, 288,943.07-cost result becomes 17.220 vehicles and 7,369.50 with ALNS Init. RoutingBlocks improves from 454/500 accepted outputs in its original control to 500/500. Appendix F.1 reports these controls, and E.2 records search restrictions. HMA supplies the strongest mean-cost reference in this complete comparison.

AuthorsOpenReview ↗

Response to Reviewer yjwD (W3): Experimental Findings and Implications

3. W3 — What do the experiments reveal beyond expected training or scale effects?

The experiments show three things that are not implied by scale alone: the training corpus changes which architecture is cheapest, a strong heuristic still leads the best neural policy, and feasibility holds where cost does not.

FindingEvidenceWhere
(a) Road training saves mainly through fewer vehiclesMean cost falls by 31.4–65.6% at N=1000 with matched 100-location training; the vehicle cost term accounts for 81.8–86.9% of the savingSection 6.1; Appendix F.2
(a) The training corpus changes the lowest-cost architectureAt N=500 and N=1000, RRNCO has the lowest reported mean cost among the E100 policies, while DRL-TS leads the D100 policiesSection 6.1, Table 2
(a) The training stage matters for checkpoint selectionThe checkpoint matching N=100, 500, or 1000 has the lowest mean cost at that size among the three evaluated stages; all evaluated stages are fully feasible, including at N=2000Section 6.3; Appendix F.2
(b) Strong search leaves room for neural solution qualityHMA 1,945.47 vs. DRL-TS 1,997.51 on the same 500 T1 tasks; 477/500 paired wins; lower mean cost in all ten cities; Gap 2.7%Section 6.2; Appendix F.1
(b) Road-information inputs align with costThe two lowest-cost D1000 policies retain explicit pairwise road information; EVRPTW-RL, which aggregates travel-time rows, costs 19.79% more than TERRAN, which omits explicit pairwise inputTable 3; Appendix E.2
(c) Full feasibility coexists with city-dependent cost sensitivityUsable capacity 100%→50%: every policy stays fully feasible while cohort mean costs rise by 15.42–59.14%; Los Angeles has the largest increase for every policy (43.37–93.22%)Appendix F.6

On (a), earlier checkpoints therefore remain useful for smaller tasks, and Section 6.5 summarizes the cross-experiment findings. On (b), the worked example in Appendix E.2 establishes that row aggregation loses destination identity; the empirical rankings motivate controlled study of road-relation representations. On (c), the capacity study reruns decoding as a separate evaluation, using 30 trajectories per instance at each capacity level and holding the selected D1000 checkpoint fixed within each method; each percentage uses that policy's own full-capacity baseline from this study. These results show why feasibility should be accompanied by operating cost when evaluating deployment conditions.

Reviewer_228FOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper provides a large-scale dataset for the well-celebrated EV routing problem with time windows. The dataset is grounded in real-world road network and delivery contexts. Computational experiments are conducted to evaluate existing methods in various settings, including generalization across geographical contexts and instance sizes.

Critical strengths
  • I think a large-scale dataset (grounded in real-world data) can be a very valuable asset for the operations research and learning to optimize communities. I really appreciate the efforts.
Critical weaknesses
  • However, the writing of this paper really needs improvement. I was looking forward to reading the findings in Sections 6.1–6.2, but found the discussion difficult to follow. Many abbreviations are used without definition, and much of the discussion appears to be a mechanical restatement of the numbers reported in the tables, with limited interpretation. Moreover, what exactly does "training" mean for Gurobi in this context? I encourage the authors to improve the presentation of the paper, as the findings could be of interest to a broad research community.
Questions for authors

See above

3 comments
AuthorsOpenReview ↗

Response to Reviewer 228F (W1): Experimental Clarity and Interpretation

Response to Reviewer 228F

W1 — Experimental clarity, interpretation, and terminology

Thank you for highlighting these presentation problems. We have rewritten Section 6 around four explicit questions, separating the training-source and training-stage analyses previously combined in Section 6.2. Each subsection now explains the comparison and develops its answer from the results. The restructured discussion in Section 6 is marked in blue in the revised manuscript.

Clarifying what is trained. Gurobi has no training stage: it solves the mixed-integer linear programming (MILP) model separately for each test instance under a 1,800-second optimization budget. The heuristics likewise optimize each test instance. Only the neural policies are trained offline, and their parameters remain fixed during evaluation. The E/D labels identify their training corpora: Euclidean instances and EVRP-TW-D, respectively. Table 3 separates neural checkpoints from the exact model and heuristics.

Connecting experiments to their questions.

QuestionComparisonAnswerWhere
Training sourceEuclidean vs. road training at matched training size, within each architecture, on the same road tasksRoad training improves performanceSection 6.1, Table 2
Method comparisonAll five D1000 neural policies vs. the exact model and four heuristics on the same 1,000-location T1 tasksNeural policies have lower mean costs than three of the four heuristics; the newly added hybrid memetic algorithm (HMA) is lowest, with the best neural policy costing 2.7% moreSection 6.2, Table 3
Workload sizeEach retained checkpoint deployed unchanged across 100–2,000 locationsFull feasibility throughout; size-dependent costs explain why earlier checkpoints remain useful for smaller workloadsSection 6.3, Table 4
Geographic generalizationFixed D1000 checkpoints and workload size across T1–T3Feasibility carries to both geographic holdouts; gaps to HMA narrow in held-out communities and widen in the unseen city by architecture-dependent amountsSection 6.4, Table 5

Interpreting the findings. For the four architectures feasible under both training sources, reduced vehicle count accounts for 81.8–86.9% of the mean operating-cost savings at 1,000 locations. The main cost benefit is therefore serving the same deliveries with fewer vehicles, with shorter road travel providing the remainder (Section 6.1; training-source cost accounting in Appendix F.2). The discussion also connects experiments: on the same road tasks, the lowest-cost neural architecture changes with the training corpus. RRNCO has the lowest reported mean cost among the E100 policies at 500 and 1,000 locations, whereas DRL-TS leads the D100 policies (Section 6.1, Table 2).

On the 1,000-location T1 tasks, the two D1000 neural policies closest to HMA both retain explicit pairwise travel relations, motivating further study of how policies represent and use road information (Section 6.2). The revised Summary (Section 6.5) draws these findings together: road-based training helps, a strong heuristic still leads at the operational scale tested, and, for the road-trained policies, feasibility is more stable than solution quality across the tested workloads and geographic splits.

Making notation accessible. Section 6 distinguishes the method families, introduces training configurations and T1–T3 test splits, and states the reported metrics. Training-label subscripts specify the training workload size, while N denotes the number of delivery locations in a test instance. Table 2 explicitly identifies the training sources and left/right result order. The heuristic abbreviations are expanded in Related Work. Appendix E.1, “Abbreviations and Experimental Notation,” collects method terminology, training/test labels, and metrics for quick reference.

I'd like to clarify that my comment on clarity and presentation applies to the entire paper, not just the experiments. I really think the paper needs substantial changes to improve the presentation to make it meet the bar for ICLR, and more importantly, make it understandable by humans. For this reason, I keep my rating.

Thank you for the clarification. Since the original review identified Sections 6.1–6.2 as difficult to follow, we restructured Section 6 around four explicit questions. We also rewrote the abstract, introduction, and Sections 3–5; all changes are marked in blue in the revised manuscript. We would be grateful if the reviewer could point to the specific sections or passages that remain hardest to follow, so that we can address them directly.

Reviewer_YLQPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces EVRP-TW-D/B, a dataset and benchmark for electric vehicle routing with time windows. It combines directed road networks with Amazon delivery attributes, providing 71,500 instance views across 11 US cities with 100–2,000 customers. It also provides a reusable generator, splits for scale and geographic generalization, and independent route verification. Experiments compare five learned policies with Gurobi and two metaheuristics. Under the stated cost model and budgets, road training improves performance, and road-trained policies outperform the evaluated search baselines at 500 and 1,000 customers. Results at 1,000 customers also show transfer to held-out regions.

Critical strengths
  1. The paper brings directed roads, observed delivery-attribute tuples, and workloads of up to 2,000 customers into one dataset. The generator and splits support separate evaluations on new delivery days, larger workloads, and unseen regions. This enables studies of scale and geographic generalization beyond what small, fixed synthetic benchmarks support.
  2. Independent route replay checks feasibility and recomputes cost. The paper reports feasibility separately from solution quality, specifies computational budgets, and distinguishes best-known solutions from proven optima. It also documents implementation changes and restrictions, making the scope of each comparison explicit.
Critical weaknesses
  1. The paper documents its input sources in detail but provides limited quantitative evidence about the resulting workloads. Transferring intact delivery tuples preserves within-stop attribute combinations, but assigning them to new locations and filtering for feasibility can change their distribution and spatial relationships. Appendix C describes held-out source records and diagnostic summaries without showing enough of the comparisons to assess these changes. This makes it difficult to judge which real delivery conditions the dataset represents, a central question for its main contribution.
  2. ALNS and VNS/TS are modified reimplementations using single-customer initialization and restricted search. At 1,000 customers, their mean fleet sizes are 16.76 and 683.34, compared with 4.43 for Road-1000 DRL-TS. The paper reports consolidation difficulties and anytime results, but I cannot tell how much of this gap comes from the workload and budget versus implementation choices. Route verification checks the returned solutions; it does not establish that the solvers are competitive. This limits the comparison with search methods, while the improvements from changing training data within each learned method remain informative.
Questions for authors
  1. Could you report the main differences between observed and generated workloads using the held-out records and diagnostics in Appendix C? Relevant comparisons include joint delivery-attribute distributions, customer travel times and spatial clustering, these results would help establish which operational properties the generator preserves.

  2. What performance checks support the use of these ALNS and VNS/TS implementations as competitive baselines? Results on compatible established benchmarks, or a limited comparison of initialization and search restrictions, would help explain the large fleet counts. In particular, does the reported advantage over search methods persist when these implementation choices are varied?

3 comments
AuthorsOpenReview ↗

Response to Reviewer YLQP (W1/Q1): Workload Validation

1. W1 / Q1 — Which delivery conditions do the generated workloads retain?

Thank you for requesting validation of the resulting workloads. In the two shared cities, generated Train closely matches its actual input pools in the shares of stops with at least four packages, long service, and specified windows. It also retains the ordering of service and window prevalence across package groups observed in the independent holdouts. Travel and spatial statistics differ from those holdouts.

The diagnostics use all 5,000 Train parents at 1,000 locations, 500 per training city. Independent references are DLA9 in Los Angeles (39 station-days, 98,177 stops) and DCH2 in Chicago (29 station-days, 23,751 stops), both excluded from generation. Statistics retain the stated within-city weekday/weekend weighting of 5:2. Appendix C.9 documents the cohorts, definitions, ten-city training summaries, and 11-city T1/T3 evaluation summaries.

(a) Joint delivery attributes

Entries show Amazon holdout → Train. Long service means more than 104 seconds. The final three columns are conditional percentages within each package group; the last reports both long service and a specified window (Appendix C.9.2, Table 16).

CityPackages/stopShare of stops (%)Long service (%)Specified window (%)Service & window (%)
LA159.8 → 66.220.9 → 19.05.7 → 5.12.6 → 2.1
LA2–331.9 → 28.154.7 → 51.97.2 → 7.15.3 → 4.9
LA4+8.3 → 5.792.6 → 90.715.5 → 18.015.0 → 17.2
Chicago160.9 → 65.818.5 → 18.77.1 → 4.92.6 → 2.0
Chicago2–331.0 → 28.453.9 → 51.48.8 → 6.86.2 → 4.7
Chicago4+8.1 → 5.992.5 → 90.716.2 → 17.715.5 → 17.0

Larger package groups have more long service and specified windows in both populations. Here, the actual input pool comprises all cleaned tuples from each selected order-source day, retaining that day's reuse frequency. The four-or-more-package shares closely follow these input pools: 5.67%→5.68% in LA and 5.87%→5.86% in Chicago. Across both cities, the three reported input-to-Train rates differ by at most 0.1 percentage points. Most differences in these three marginal rates relative to the independent holdouts are already present in these inputs (Appendix C.9.2, Table 17).

(b) Generated tasks differ from the holdouts in travel conditions

All times are in minutes; spatial units are Census Block Groups. Amazon statistics use available within-route travel-time matrix entries, while generated statistics use full-instance road-time matrices. Pairs from different Block Groups are restricted to a straight-line separation of 1–5 km. Appendix C.9.3 (Table 18) reports these statistics, including the combined 1–5 km row and the finer distance bands.

Travel betweenStatisticLA: Amazon → TrainChicago: Amazon → Train
Within Block GroupMedian2.50 → 1.541.88 → 1.67
Between Block Groups, 1–5 kmMean7.03 → 8.997.22 → 9.82
Depot→stopMedian25.20 → 31.0720.41 → 30.45

Both cohorts combine short travel within Block Groups with longer connections between Block Groups and from the depot. In both cities, generated tasks have higher depot-to-stop travel-time medians and higher mean travel times between Block Groups in the combined 1–5 km band. Their medians for travel within Block Groups are lower. The generated tasks retain source attribute combinations, while their locations exhibit a different travel regime.

(c) Spatial diagnostics

The 1-km nearby-pair share is 4.05%→3.63% in LA and 14.83%→4.59% in Chicago. The difference is larger in Chicago, where the mean number of active Block Groups per workload also increases from 87.77 to 157.87. The generated Chicago workloads are therefore more dispersed than the DCH2 holdout. Appendix C.9.4 (Tables 19–20) reports these spatial differences, workload sizes, and Block Group occupancy.

AuthorsOpenReview ↗

Response to Reviewer YLQP (W2/Q2): Initialization Controls and Stronger Baselines

2. W2 / Q2 — Do the search comparisons survive better initialization and stronger baselines?

The initialization controls explain the large vehicle count in the submitted VNS/TS result, while the added HMA changes the quality reference: HMA has the lowest mean cost in the complete T1 comparison at 1,000 locations.

The initialization controls and 1,000-location method comparison use the same 500 T1 instances and cost model. Heuristics receive 1,800 seconds including initialization; neural policies use the fixed D1000 checkpoints and 30 trajectories. ALNS Init is ALNS's construction procedure. VNS/TS and ALNS receive identical ordered starting routes; RoutingBlocks constructs its own ALNS Init routes, and HMA uses native initialization (Appendices F.1 and E.2–E.3).

(a) Initialization controls

The table reports accepted outputs after 1,800 seconds. Partially feasible controls use only their solved instances for the displayed mean cost and vehicle count.

MethodInitializationFeasibleFinal mean cost (USD)Final mean vehicle countAveraging set
VNS/TSSchneider-style, no fallback0/500———
VNS/TSSingle-customer certificates (submitted fallback)500/500288,943.07683.340All 500
VNS/TSALNS Init500/5007,369.5017.220All 500
ALNSALNS Init500/5007,089.9116.550All 500
RoutingBlocksOriginal random initialization454/500268,251.26631.559Its 454 feasible instances
RoutingBlocksALNS Init500/5003,937.909.136All 500

The mean vehicle count of 16.76 cited in the review refers to the submitted ALNS run. The revised ALNS row comes from the paired shared-initialization evaluation described above.

The submitted certificate fallback starts with a mean vehicle count of 705.838 and finishes with 683.340. ALNS Init instead starts with 19.054 and VNS/TS finishes with 17.220, reducing its mean cost by 97.45%. From the identical initial mean cost of 8,168.15, VNS/TS and ALNS reduce cost by 9.78% and 13.20%; ALNS wins 240 paired instances and VNS/TS wins 260. These controls separate initialization quality from subsequent search. The revision also uses ALNS Init for VNS/TS on T2 and T3 at 1,000 locations, where its mean vehicle counts are 19.00 and 21.21 (Appendix F.4). The T1 cohorts at 100 and 500 locations retain the certificate-based configuration and are marked accordingly in Appendix F.

Both RoutingBlocks configurations include independent verification within the 1,800-second budget. On their common 454 feasible instances, mean cost falls from 268,251.26 to 3,926.68, with lower cost on all pairs. Appendix F.1 summarizes the controls and search progress; this response additionally reports means on partially feasible and matched subsets, the certificate-based starting vehicle count, and ALNS–VNS/TS paired wins.

The evaluated implementations retain VNS/TS's nearest-candidate restrictions and ALNS's scale-dependent candidate truncation (Appendix E.2). We have not separately measured the effects of these restrictions; the comparisons therefore concern the evaluated configurations.

(b) Comparison with stronger search

We broadened the search comparison with HMA adapted from the authors' C++ solver and RoutingBlocks built from its public Python/C++ framework and EVRP-TW ALNS example (Appendix E.2). Section 6.2 (Table 3) reports every method on the complete 1,000-location cohort. The neural policies have lower mean costs than the evaluated VNS/TS, ALNS, and RoutingBlocks configurations, while HMA attains 1,945.47 versus DRL-TS's 1,997.51. HMA wins 477/500 pairs and has lower mean cost in all ten cities. DRL-TS's Gap to HMA is 2.7%. Appendix F.1 provides distributions, paired comparisons, and the city breakdown.

At N=500, HMA also has lower mean cost than all evaluated road-trained neural checkpoints (USD 1,173.02; Section 6.3, Table 4).

The revised comparison therefore distinguishes initialization sensitivity from the remaining quality gap to HMA. Section 6.5 summarizes these findings; Appendix E specifies execution and reporting conventions.

Thank you for the detailed response and additional experiments. Diagnostic comparison with held out records and the initialization controls with stronger baselines like HMA answered my questions, and made the scope of the benchmark more clear. I will maintain my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Benchmark Contributions and Practical Value

Thank you for the constructive feedback. We have updated the manuscript and posted detailed responses to each reviewer. We would like to highlight three contributions:

  • A large-scale, reusable benchmarking framework. To our knowledge, EVRP-TW-D/B is the first EVRP-TW framework to combine reusable generation at depot scale, directed road networks, and joint delivery attributes from operational records, with predefined evaluation splits. It provides 71,500 benchmark views spanning 100–2,000 delivery locations across 11 cities. The same generator can produce additional delivery-day instances and construct environments for further U.S. cities covered by the same public sources.

  • Open implementations for reproducible comparisons. We evaluate one exact model, four heuristics, and five neural policies under a shared cost objective, declared computational budgets, and independent route verification. Baseline code and neural checkpoints are provided with the submission, with method sources and adaptations documented.

  • Practical value for algorithm evaluation. At a common training size, changing the training corpus changes which neural architecture has the lowest reported mean cost on the same road tasks. On the 1,000-location road tasks, the newly added hybrid memetic algorithm (HMA) achieves the lowest mean cost, with the best neural policy costing 2.7% more. In the geographic tests, the neural policies remain fully feasible while their gaps to HMA widen in the unseen city, showing why operating cost needs to be assessed alongside feasibility.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.