Reproducibility Limits of Mixed-Integer Linear Programming-based methods
Résumé
``Do not trust computers,'' this sentence has probably been uttered by most, if not all, computer scientists at some point. At the same time, computer scientists are building version control tools, databases and many other tools to help them reproduce certain behavior of a software library, with the ultimate goal of trusting the results of computations. In this paper, we propose to look at issues that can arise when trying to reproduce the results of Mixed-Integer Linear Programming (MILP)-based methods, from the user perspective. These methods rely on external tools which introduce new parameters when dealing with reproducibility issues. In particular, the phenomenon of performance variability has a significant impact on execution times, which is currently underestimated. In this work, we question the meaning of reproducibility in the context of the MILP approaches and we propose ways to reduce output dependence on solver uncertainties.
Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
Licence |