PHASE-MOO: reproducibility data and code for calibrated geometric population optimization
Abstract
Data and source code supporting PHASE-MOO: population search over competing local geometries for Pareto front approximation (version 2.0.0, revised implementation and experiments). This version replaces the experimental evidence of version 1.0.0, which remains available as the historical record. It contains the experiments of the revised manuscript, all run with the third revision of the PHASE-MOO implementation (adaptive initial step, normalized merit, projected sufficient-decrease test, output of all nondominated evaluated points, and de-duplicated geometry portfolio): a racing-based calibration with 2,172 runs (64 Latin-hypercube configurations and a control per problem, seeds 1101-1110); a component ablation with 480 runs (five variants, seeds 6101-6120); six high-budget case studies (seed 6106; 1,000,000 evaluations, 2,000,000 on HD200), with stored evaluations, initial and final vectors, and the recovered seed of origin of every nondominated point; a comparison with eight methods (NSGA-II, MOEA/D, SMS-EMOA, NSMA, FD-BB, FD-SD, COMO-CMA-ES, UHV hybrid) with 519 recorded runs on seeds 6101-6110, including the protocol, declared adaptations, accounting audits, statistical analysis, and an equal-size (100-point) check; one FD-BB run on HD200 failed and is missing. Contents include per-run records and indicators, nondominated fronts, frozen reference fronts, setups and protocols, source-code snapshots of the PHASE-MOO kernel (with the copies used in each experiment), own adapters for the comparators, analysis and plotting scripts, figures, a file manifest, and SHA-256 checksums with a verification script. Straight initial-to-final connections in the high-budget figures are not optimization trajectories. Trajectory databases, software environments, third-party source code, and compiled binaries are not included; third-party methods must be obtained from their original repositories at the commits listed in the protocol. The manuscript is not part of this archive. Original code is licensed under MIT. Generated data, figures, and documentation are licensed under CC BY 4.0. License scope and third-party exclusions are specified in the archive.
