Skip to content

feat: catalogue traffic assignment, independent set, dominating set and Steiner tree benchmarks - #323

Merged
Mearman merged 18 commits into
mainfrom
data/optimisation-benchmark-networks
Nov 11, 2026
Merged

Mearman merged 18 commits into
mainfrom
data/optimisation-benchmark-networks

Conversation

@Mearman

@Mearman Mearman commented Nov 11, 2026 •

Copy link
Copy Markdown
Owner

Adds optimisation benchmark networks to the catalogue, with four new graph/v1 parsers to read them.

Added

Transportation Networks for Research (https://github.com/bstabler/TransportationNetworks, read from the repository zip at commit d1639b4). New tntp parser reads a network's _net.tntp file and trip table together from a parts archive into one graph: zones and nodes, each link a directed edge carrying the columns its harmonised header names (capacity, length, free-flow time, B, power, speed, toll, link type; Gold Coast has critical speed and lanes instead), and each origin-destination pair a commodity node with its demand joined to its two zones, as the commodity convention does. A parts document now keeps the meta its parser gives, as a single-file document already did. The repository has no licence file; its README states terms of use (academic research purposes only, indicate the source in any publication), recorded as a custom licence pointing at that section.

  • Fetchable and round-tripped with counts matching: tntp-anaheim, tntp-austin, tntp-barcelona, tntp-berlin-center, tntp-berlin-friedrichshain, tntp-berlin-mitte-center, tntp-berlin-mitte-prenzlauerberg-friedrichshain-center, tntp-berlin-prenzlauerberg-center, tntp-berlin-tiergarten, tntp-birmingham, tntp-braess, tntp-chicago-sketch, tntp-eastern-massachusetts, tntp-gold-coast, tntp-philadelphia, tntp-sioux-falls, tntp-terrassa-asymmetric, tntp-winnipeg, tntp-winnipeg-asymmetric.
  • Recorded as not fetchable, with the reason in the entry: tntp-chicago-regional (6,311,900 edges; the round trip ran out of the default Node heap, waits on Stream graph/v1 results #305), tntp-hessen-asymmetric (the header names nine columns, every link line has ten), tntp-munich (no first through node, zone numbers outside 1 to n), tntp-sydney (link lines lack the ; terminator and the trip table ships only as a zip inside the repository).

FrontierCO (Hugging Face CO-Bench/FrontierCO at revision e92b9d6, MIT per the dataset card, as the existing CFLP entry records it). Each instance file is a source of a collection; each entry names the track's config.py, which holds the reference values.

  • huggingface-frontierco-mis: 73 graphs (complements of the DIMACS clique benchmark graphs, BHOSLIB frb, generated RB), read with the new dimacs-graph parser, which keeps isolated vertices from the problem line. Round-tripped whole.
  • huggingface-frontierco-stp: 88 SteinLib instances, read with the new steinlib parser (terminals typed, weighted edges, Comment and Presolve values in meta). Round-tripped whole.
  • huggingface-frontierco-mds: 60 PACE 2025 style graphs, not fetchable: the hard set holds 164,677,007 edges by its problem lines, so it waits on Stream graph/v1 results #305.

PGLib-OPF v23.07 (https://github.com/power-grid-lib/pglib-opf, read from the repository zip at commit dc6be4b, CC BY 4.0 per its LICENSE and README). New matpower parser reads MATPOWER version 2 cases: buses and generators as nodes with their tables' columns and quadratic costs, generators joined to their buses, branches as edges. pglib-opf-typical, pglib-opf-api and pglib-opf-sad, 66 cases each, all round-tripped; each names BASELINE.md for the reference objective values.

Skipped

  • FrontierCO TSP, CVRP and CPMP tracks: they are point sets with coordinates, demands and capacities, not graphs, and the tsplib parser reads only the orienteering types into tasks/v1. They would need that parser extended for TSP and CVRP and a new one for the capacitated p-median files.
  • LEMON minimum-cost flow data (NETGEN, GRIDGEN and the rest): the data host lime.cs.elte.hu still does not answer and the Wayback Machine holds no instance file, as Catalogue the Kovacs minimum-cost flow benchmark once its data host answers #300 already records.
  • The Zitzler and Thiele multi-objective knapsack problems: they are item lists, not graphs, and no category describes them.

Checks

Each new entry passes keepset validate; each fetchable one round-trips with its expected counts (Sioux Falls through the CLI from the pinned snapshot, the rest with the same bytes served locally). Wayback pins and checksums are recorded for the TNTP sources. The FrontierCO and PGLib sources have no pins yet: the archive script's captures of the FrontierCO files were refused because the raw snapshots answered 404 when read back, so archive:check lists them as unaccounted for, beside the sources it already lists on main. pnpm run prepublishOnly passes locally (141 test files, 1,496 tests).

Add a tntp parser that reads a Transportation Networks for Research
dataset from a parts archive of its network file and its trip table
into one graph: zones and nodes typed as such, each link a directed
edge carrying the columns its harmonised header names, and each
origin and destination pair a commodity node with its demand, joined
to its two zones as the graph/v1 commodity convention does.

A parts document now keeps the meta its parser gives beneath the
entry's own, as a single-file document already did.
… assignment networks

Catalogue each TNTP network of the bstabler/TransportationNetworks
repository, read from the repository's zip at a pinned commit with the
tntp parser, under the repository's stated terms of use. Each entry
names the source's best known link flows where it publishes them.

The networks whose files do not follow the format (Hessen asymmetric,
Munich, Sydney) and Chicago Regional, whose trip table makes a graph
larger than a held result reads, are recorded without a recipe and say
why.
Add a dimacs-graph parser for undirected graphs written with a problem
line and one edge per line, either as the DIMACS challenge graph format
(p edge, then e lines) or as the PACE challenge format (p and the
problem's name, then bare pairs). Nodes are 1 to n from the problem
line, so a vertex no edge touches is kept, which independent set and
dominating set instances depend on.
Add a steinlib parser for the STP format: the nodes with the terminals
typed as such, each weighted undirected edge, the Comment section's
values and, for a presolved instance, the Presolve section's fixed
value, bounds and original counts in the meta. The original instance's
mapping lines, terminals and coordinates are not read, and directed,
rooted, prize-collecting and obstacle variants are refused.
Add a matpower parser for the MATPOWER case format, version 2: buses
and generators as nodes carrying their tables' columns and each
generator's quadratic cost, an edge joining each generator to its bus,
and the branches as edges carrying theirs. The case's name, version,
base MVA and legacy area table go in the meta.

The pattern for a decimal with an exponent is shared with the tntp
parser, which reads the same notation.
…nd Steiner tree tracks

Catalogue three more graph tracks of the FrontierCO benchmark on
Hugging Face at a pinned revision, each instance file a source of a
collection: the maximum independent set track read with the
dimacs-graph parser, the Steiner tree track with the steinlib parser,
and the minimum dominating set track, whose hard instances make a
graph larger than a held result reads, recorded without a recipe.
Each names the track's config.py, which holds FrontierCO's reference
values.
Catalogue the typical, heavily loaded (API) and small angle difference
(SAD) groups of the IEEE PES Power Grid Library's optimal power flow
cases, read from the repository's zip at a pinned commit with the
matpower parser, under the CC BY 4.0 licence the library states for
its case files. Each names BASELINE.md, which holds the reference
objective values.
@Mearman
Mearman marked this pull request as ready for review November 11, 2026 21:56
@Mearman
Mearman merged commit af5fe2a into main Nov 11, 2026
10 checks passed
@Mearman
Mearman deleted the data/optimisation-benchmark-networks branch November 11, 2026 21:56
@github-actions

Copy link
Copy Markdown

🎉 This PR is included in version 0.138.0 🎉

The release is available on:

Your semantic-release bot 📦🚀

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant