Impact of Scaling and Rounding on Metaheuristic Performance for the Vehicle Routing Problem with Time Windows: Raw Experiment Data
Year: 2026
Type: dataset
Abstract: Raw experiment data of the study "Impact of Scaling and Rounding on Metaheuristic Performance for the Vehicle Routing Problem with Time Windows" (Florian Rascoussier, Romain Billot, Lina Fahed, Christine Solnon), presented at the 16th Metaheuristics International Conference (MIC 2026, Ischia, Italy) and deposited on HAL as hal-05646952. The dataset contains the 9,504 solver runs behind every table and every figure of the paper, together with the small aggregate files the paper's scripts read. It contains no analysis code and no derived conclusions: the paper carries those, and everything in it can be recomputed from what is here. Design. 176 classical Euclidean VRPTW instances with floating-point arc costs: the 56 Solomon instances with 100 customers and the Gehring-Homberger instances with 600 and 1,000 customers (60 each). Each instance is scaled by a factor P in {1, 10, 1000} and rounded to integers under three schemes: pessimistic (costs and service times rounded up, time windows tightened), optimistic (rounded down, time windows loosened) and nearest integer. Two open-source solvers, PyVRP 0.12.2 in its Hybrid Genetic Search configuration and OR-Tools, solve every scaled instance under the hierarchical objective (fleet size first, then cost), with three seeds (42, 123, 456) and a wall-clock limit of 120 s at 100 customers, 900 s at 600 and 1,800 s at 1,000. The grid is 176 x 3 x 3 x 2 x 3 = 9,504 runs. Every incumbent found on the integer instance is mapped back to the original floating-point instance and re-evaluated there, for its true cost and for its feasibility, which is how the study measures what scaling and rounding cost in the solution space that matters. Contents. data/ holds one JSON record per run, named <instance>_<n>_ _<rounding>_<seed>_<solver>.json, where the rounding field reads ceil, floor or round for the pessimistic, optimistic and nearest schemes. Each record carries the configuration, the solver-reported integer cost and route count, the re-evaluated float cost and feasibility verdict, the full timestamped incumbent trajectory with its routes, the area-under-curve values, the final routes and the machine description. The run_status field takes two values across the deposit: FINISHED (7,343 runs) and NO_FLOAT_FEASIBLE (2,161 runs whose final integer solution violates a time window once mapped back to the float instance, the phenomenon the paper's feasibility analysis quantifies). scaling_rounding_curves.json is the compact intermediate the aggregation stage derives from data/ (the trajectories without their route lists), scaling_rounding_raw_results.csv is one row per run, gap_goal_run_summary.csv and gap_goal_float_feasible_incumbents.csv serve the gap-goal and convergence analyses, feasibility_incumbent_summary.csv the feasibility heatmaps, stats/ the pairwise and summary statistics, tex/ the LaTeX table sources, img/ and img_v1/ the figures as the pipeline exported them, and temp/ the Grid'5000 job configuration and log. Reference values. Every gap and every anytime score in the paper is measured against the frozen best-known solutions of the SINTEF benchmark as curated in the MAMUT-routing platform (Sintef2008 family, version 0.1.0), lifted into the penalized fleet-then-cost space. The bks_cost, bks_num_routes and gap_to_bks fields baked into the records at run time predate that convention and are superseded: gaps should be recomputed from the frozen reference rather than read from those fields. The bundled README.md states this and decodes the layout and the record fields. Environment. All runs executed on the neowise cluster of the Grid'5000 testbed at the Lyon site (dual-socket AMD EPYC 7642 nodes, 48 physical cores), one solver process per core slot, under CPython 3.13.11, within a single scheduler job. The records retain their original cluster provenance. The instances themselves are not duplicated here: they are distributed as checksummed artifacts through the MAMUT-routing benchmark platform and carry their own licenses through it.
Institutions: Laboratoire des Sciences et Techniques de l’Information de la Communication et de la Connaissance, IMT Atlantique, EMERAUDE: Systèmes Embarqués audio programmables, Center of Innovation in Telecommunications and Integration of Service, Institut National des Sciences Appliquées de Lyon, Centre Inria de Lyon
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
Funder: Agence Nationale de la Recherche
OpenAlex ID: W7211908419