Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
A Markov Decision Process for Variable Selection in Branch & Bound
Authors: Paul STRANG, Zacharie ALES, Côme Bissuel, Olivier JUAN, Safia Kedad-Sidhoum, Emmanuel Rachelson
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Computational experiments validate our model empirically, as our branching agent outperforms prior state-of-the-art RL agents on four standard MILP benchmarks. |
| Researcher Affiliation | Collaboration | Paul Strang EDF R&D, France CNAM Paris, France EMAIL Al es ENSTA IP Paris, France CNAM Paris, France zacharie.ales@ensta.frCôme Bissuel EDF R&D, France EMAIL Juan EDF R&D, France EMAIL Kedad-Sidhoum CNAM Paris, France safia.kedad EMAIL Rachelson ISAE-SUPAERO, France EMAIL |
| Pseudocode | Yes | Algorithm 1 DQN-BBMDP |
| Open Source Code | Yes | To foster reproducibility, our implementation and pretrained models are made publicly available at https://github.com/abfariah/bbmdp. |
| Open Datasets | Yes | We consider the usual standard MILP benchmarks for learning branching strategies: set covering, combinatorial auctions, maximum independent set and multiple knapsack problems. ... We train and test on instances of same dimensions as Scavuzzo et al. [2022] and Parsonson et al. [2022], see Appendix F. ... Instance datasets used for training and evaluation are decribed in Table 4. ... Combinatorial auction Leyton-Brown et al. [2000] ... Set covering Balas and Ho [1980] ... Maximum independent set Bergman et al. [2016] ... Multiple knapsack Fukunaga [2011] |
| Dataset Splits | Yes | Models are trained on instances of each benchmark separately, and evaluated on test instances and transfer instances. ... For evaluation, we report the node and time performance over 100 test instances unseen during training, as well as on 100 transfer instances of higher dimensions (see Table 4 in Appendix F).Instance datasets used for training and evaluation are decribed in Table 4. |
| Hardware Specification | Yes | All experiments were conducted on an NVIDIA DGX A100 system equipped with 8 A100 40GB GPUs, 2 AMD EPYC 7742 64-core CPUs (128 threads total), and 1 TB of DDR4 RAM. |
| Software Dependencies | Yes | For our experiments, we use the open-source solver SCIP 8.0.3 [Bestuzheva et al., 2021] as backend MILP solver, along with the Ecole library [Prouvost et al., 2020] both for instance generation and environment simulation. |
| Experiment Setup | Yes | As to SCIP configuration, as in previous work, we set the time limit to one hour, disable restart, and deactivate cut generation beyond root node. All the other parameters are left at their default value. ... Table 5: Training parameters for all DQN branching agents. |