Skip to content

Vu2020 (TDVRP)

The TDVRP layer of Vu2020 is the same curated 168-instance subset with the time windows removed, sharing the arrival-time-function sidecars bit-for-bit with the TDVRPTW layer under the same Duration checker contract. At these sizes (59--99 customers) the variant provides the largest capacitated time-dependent instances of the current TD collection without time-window pruning. No previously published solutions exist for this variant; all BKS come from KAYROS heuristic runs on Grid5000.

Licensing note: as for the TDVRPTW layer, MAMUT-authored curation artifacts are under the MIT License, while underlying third-party material is not relicensed.

The family currently contains 168 curated instances over 3 sizes, each with a Duration BKS.

Data

Instances and best-known solutions: /benchmarks/tdvrp/vu2020/ on the website. Repository: https://github.com/ANR-MAMUT/MAMUT-routing-TDVRP-Vu2020.

Repository README

Vu2020 TDVRP benchmark family (Arigliano-generator TD travel times, larger sizes, curated VRP variant)

Satellite benchmark repository of MAMUT-routing (ANR MAMUT, ANR-22-CE22-0016), mounted there as benchmarks/TDVRP/Vu2020. Self-contained: instances, canonical arrival-time-function (ATF) sidecars and best-known solutions ship together.

What this family is — and what it is not

The raw data extends the Arigliano et al. TD-TSPTW generator (same IGP model: 73 speed zones, 3 shared speed profiles) to larger sizes, as used by Vu et al. (2020): sizes 59/79/99 customers, congestion depth Delta ∈ {70, 80, 90, 98}, traffic patterns A/B, absolute TW widths ∈ {40, 60, 80, 100, 120, 150, 180}, five instances each — an 840-instance full factorial with no demands, capacities, fleet sizes or service times. Instance names keep the Vu- prefix; the recorded instance_origin is Ari2018 (same generator).

This canonical family is the VRP variant curated by Onyr (2024–2026): demands, vehicle capacities, fleet sizes and gaussian service times were generated once by the legacy TDVRPTW-benchmarks pipeline and are pinned here verbatim (recorded per instance in metadata.legacy_attribute_source); time windows removed (TDVRP twin). It is therefore not comparable with published TD-TSPTW results on the underlying raw files.

Curated subset (deterministic)

The family ships 168 instances, not the full 840: one per structural cell (n, Delta, pattern, width), candidates holding a legacy best-known solution first, remaining ties by ascending SHA-256 of the canonical instance name. Names Vu-A<id>-p<Pattern>-d<Delta>-w<Width> under n=<customers> carry all structural parameters. The full raw trees remain available upstream.

Format

  • <Name>.vrp.json — MAMUT shape: depot 0, mandatory display coordinates (deterministic stress-layout embedding of the symmetric raw distance matrix; near-exact here, quality in metadata.coordinates_method), horizon [0, depot due date], td block with the sidecar's storage-independent sha256.
  • <Name>.atf.json.gz — canonical ground truth: one arrival-time NDCPWLF per arc, exact IGP consolidation (breakpoints at every speed-zone boundary crossing, exact forward Ichoua re-evaluation; last zone extended beyond the horizon). All sidecars are gzipped in this satellite.
  • <Name>.bks.Duration.json — best known solution under the Duration objective; costs are always the canonical checker's output.

Population pipeline: populate_td_ari_vu v1 (curation tooling maintained outside this repository; links will be published with the benchmark paper).

Provenance and licensing

Underlying raw data: the Arigliano-generator TD-TSPTW distribution at Vu et al. (2020) sizes. MAMUT-authored curation artifacts under the MIT license; this notice does not relicense the underlying third-party benchmark definitions.