Reinforcement Learning for Branch-and-Bound Optimisation Using Retrospective Trajectories

Authors: Christopher W. F. Parsonson, Alexandre Laterre, Thomas D. Barrett

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments on four combinatorial tasks, our approach enables learning-to-branch without any expert guidance or pre-training. We outperform the current stateof-the-art RL branching algorithm by 3-5 and come within 20% of the best IL method s performance on MILPs with 500 constraints and 1000 variables, with ablations verifying that our retrospectively constructed trajectories are essential to achieving these results.
Researcher Affiliation Collaboration Christopher W. F. Parsonson1*, Alexandre Laterre2, Thomas D. Barrett2 1UCL 2Insta Deep
Pseudocode No The paper mentions 'corresponding algorithms and hyperparameters used' in Appendix A.1, but the appendix content, which might contain pseudocode or algorithm blocks, is not provided in the main paper text.
Open Source Code Yes All code for reproducing the experiments and links to the generated data sets are provided at https://github.com/cwfparsonson/retro_branching.
Open Datasets Yes MILP Problem classes. In total, we considered four NP-hard problem benchmarks: set covering (Balas, Ho, and Center 2018), combinatorial auction (Leyton-Brown, Pearson, and Shoham 2000), capacitated facility location (Litvinchev and Ozuna Espinosa 2012), and maximum independent set (Bergman et al. 2016).
Dataset Splits Yes To evaluate the quality of the agents branching decisions, we used 100 validation instances (see Appendix C for an analysis of this data set size) which were unseen during training, reporting the total number of tree nodes and LP iterations as key metrics to be minimised.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running experiments were mentioned in the paper.
Software Dependencies Yes We used the open-source Ecole (Prouvost et al. 2020) and Py SCIPOpt (Maher et al. 2016) libraries with SCIP 7.0.1 (SCIP 2022) as the backend solver to do instance generation and testing.
Experiment Setup No The paper states 'see Appendix A.1 for a detailed description of our RL approach and the corresponding algorithms and hyperparameters used,' indicating that hyperparameter details are not present in the main text.