Learning to Schedule Heuristics in Branch and Bound

Authors: Antonia Chmiela, Elias Khalil, Ambros Gleixner, Andrea Lodi, Sebastian Pokutta

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform extensive computational experiments on a class of challenging instances and demonstrate the benefits of our approach (Section 6).
Researcher Affiliation Academia Antonia Chmiela Zuse Institute Berlin, Germany chmiela@zib.de Elias B. Khalil University of Toronto, Canada khalil@mie.utoronto.ca Ambros Gleixner Zuse Institute Berlin, Germany HTW Berlin, Germany gleixner@zib.de Andrea Lodi CERC, Polytechnique Montréal, Canada andrea.lodi@polymtl.ca Sebastian Pokutta Zuse Institute Berlin, Germany Technische Universität Berlin, Germany pokutta@zib.de
Pseudocode Yes The final algorithm can be found in Appendix C.
Open Source Code Yes The code we use for data collection and scheduling is publicly available.1 [1https://github.com/antoniach/heuristic-scheduling]
Open Datasets Yes The Generalized Independent Set Problem (GISP) (Hochbaum and Pathria, 1997; Colombi et al., 2017) and the Fixed-Charge Multicommodity Network Flow Problem (FCMNF) (Hewitt et al., 2010). For GISP, we generate two types of instances: The first one takes graphs from the 1993 DIMACS Challange which is also used by Khalil et al. (2017) and Colombi et al. (2017)
Dataset Splits No The paper provides training and testing set sizes (e.g., “120 for training and testing,” “25 for training and 10 for testing,” “20 for training and 120 for testing”) but does not specify a separate validation set or detailed split methodology (e.g., percentages, random seed, or cross-validation setup) that would allow for exact reproduction of data partitioning.
Hardware Specification Yes For our experiments, we used a Linux cluster of Intel Xeon CPU E5-2660 v3 2.60GHz with 25MB cache and 128GB main memory.
Software Dependencies Yes used the state-of-the-art solver SCIP 7.0 (Gamrath et al., 2020) with CPLEX 12.10.0.0 as the underlying LP solver.
Experiment Setup Yes The time limit in all experiments was set to two hours; for data collection to four hours. [...] we implemented an exhaustive testing framework that uses four random seeds [...] We used the primal integral as a performance metric. [...] We set the frequency offset to 0 for all diving heuristics.