Repo for GEAR: A GPU-Accelerated Global Solver for Bounded Continuous Nonlinear Programs via Linear Bound Propagation
Project page: https://verified-intelligence.github.io/GEAR/ (results, analysis and how to reproduce every number), with the gap atlas at
https://verified-intelligence.github.io/GEAR/atlas.html (all 505 instances for GEAR and the 12 AMPL solvers, any subset selectable:
certification under the paper's scoring, gap-vs-time curves, formulations, commands and logs). Source of both in site/.
| Path | Content |
|---|---|
solver/ |
the GEAR solver: command line (complete_verifier/nlp_example.py) and Python API (complete_verifier/gear/); BSD 3-Clause, see solver/NOTICE |
benchmarks/ |
the 505 benchmark instances in the format GEAR reads, with manifest and checksums; see benchmarks/README.md |
logs/ |
solver logs and per-instance results behind the paper's tables |
baselines/ |
the 12 AMPL baseline solvers: models, configurations, harness and results |
site/ |
source of the project page |
Installation, a quick start, the Python API and how to reproduce the paper's run are in the
wiki and in solver/README.md.
logs/ holds the solver output behind the reported results: one archive per solver, one log per instance (505 each).
| Path | Content |
|---|---|
logs/<solver>.tar.xz |
final 180 s runs of the 12 AMPL baselines in baselines/results/ (SCIP: final_scip_pow1) |
logs/gear.tar.xz |
GEAR, the final version evaluated in the paper, 180 s, one NVIDIA RTX 5090 |
logs/gear_results.tsv |
GEAR per instance: certified, status, P and D in the model's sense, relative gap, time |
logs/configs/int3p_crownmerge.yaml |
the GEAR configuration of these runs |
logs/MANIFEST.tsv |
solver, family, instance, bytes and sha256 of every log |
logs/SHA256SUMS |
checksums of the archives, the configuration, gear_results.tsv and the manifest |
- Archives unpack to
<solver>/<family>/<instance>.log. - Local paths are shortened:
benchmarks/is the benchmark root,runs/the baseline run directory,<gear>/the GEAR source tree,<env>/the Python environment,<root>the working directory,~the home directory,<host>the machine name. - The
Configurations:block at the top of a GEAR log lists the argument defaults, not the settings in effect; those are inlogs/configs/int3p_crownmerge.yaml.
A run is solved when its returned point satisfies every constraint and bound of the original model to absolute tolerance 1e-6 and its
primal value P and certified lower bound D satisfy (P - D) / max(1, |P|, |D|) <= 1e-4. The AMPL baselines are scored by
baselines/harness/report.py (see baselines/README.md); GEAR's points are replayed on the original model before they are returned.
# Table 1 of the paper from the scored results
awk -F'\t' 'NR>1 && (($2=="global") ? $6 : $10)==1 {n[$1]++} END {for (s in n) print s, n[s]}' baselines/results/per_instance_long.tsv
awk -F'\t' 'NR>1 && $4==1 {n++} END {print "gear", n}' logs/gear_results.tsv