MIP-GNN: A Data-Driven Framework for Guiding Combinatorial Solvers

Authors: Elias B. Khalil, Christopher Morris, Andrea Lodi10219-10227

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We integrate MIP-GNN into a state-of-the-art MIP solver, applying it to tasks such as node selection and warm-starting, showing significant improvements compared to the default setting of the solver on two classes of challenging binary MILPs. Our code and appendix are publicly available at https://github.com/lyeskhalil/mip GNN.
Researcher Affiliation Academia Elias B. Khalil,*1, 2 Christopher Morris,*3 Andrea Lodi4 1Department of Mechanical & Industrial Engineering, University of Toronto 2Scale AI Research Chair in Data-Driven Algorithms for Modern Supply Chains 3Mila Quebec AI Institute and Mc Gill University 4CERC, Polytechnique Montr eal and Jacobs Technion-Cornell Institute, Cornell Tech and Technion IIT khalil@mie.utoronto.ca, chris@christophermorris.info, andrea.lodi@cornell.edu
Pseudocode No The paper describes the architecture and passes using text and mathematical formulations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code and appendix are publicly available at https://github.com/lyeskhalil/mip GNN.
Open Datasets Yes The generalized independent set problem (GISP) (Colombi, Mansini, and Savelsbergh 2017) and fixed-charge multi-commodity network flow problem (FCMNF) (Hewitt, Nemhauser, and Savelsbergh 2010) have been used to model a variety of applications including forest harvesting (Hochbaum and Pathria 1997) and supply chain planning, respectively.
Dataset Splits Yes During training, 20% of the training instances were used as a validation set for early stopping.
Hardware Specification No Training is done on GPUs whereas evaluation (including making predictions with trained models and solving MILPs with CPLEX) is done on CPUs with a single thread. Appendix section CPU/GPU specifications provides additional details.
Software Dependencies Yes We use CPLEX 12.10.0 as a backend for data collection and BLP solving.
Experiment Setup Yes The training algorithm is ADAM (Kingma and Ba 2015), which ran for 30 epochs with an initial learning rate of 0.001 and exponential learning rate decay with a patience of 10.