INVAX: an invariance benchmarking protocol for metaheuristics — preregistered code and manifest

OpenAlex · English

Nature-inspired metaheuristics are usually compared on how well they optimise. This study compares them on a property prior to performance: whether their behaviour survives transformations under which the optimisation problem is unchanged. Four axes are measured — translation and scaling of the search space, rotation, and monotone transformation of the objective — over a pool of 21 published methods plus controls, with component-level ablations that attribute each measured loss to a specific published equation. The objective-space axis and the absolute-coordinate form of the translation axis are the study's contributions; existing work on this question is confined to the search space. The full experiment is enumerated in advance as a 2,079,000-row manifest whose SHA-256 digest is registered here, so that the plan cannot change after this timestamp without the change being detectable. Registered manifest digest (SHA-256 over 2,079,000 rows): c9db87df53ffc77d97e84db75d846ec9fa021fbe382bdfac46f64177681cbe2b Licence: MIT for the code; the prose documents are additionally available under CC-BY-4.0. Verify that this deposit produces exactly the registered experiment: cd code && pip install -e ".[dev]" && python -m invax.runner.build --check1.0.1 adds the implementations of the Spearman, Kendall tau-b and clustering tests specified in the analysis plan. The manifest digest is unchanged.