Abstract
This paper presents our submission to ML4CO, a branching rule that exploits the structure within a problem class via a graph neural network that is trained using a distributional reinforcement learning algorithm. The learned branching rule participated in the dual bound challenge, ranking third among 23 entries. Experiments on the ML4CO instances show that it performs between 16% and 34% better on average than the default reliability pseudocost branching in the state-of-the-art academic solver SCIP.
| Original language | English |
|---|---|
| Title of host publication | Machine Learning for Combinatorial Optimization |
| Subtitle of host publication | NeurIPS 2021 Competition Workshop |
| Place of Publication | San Diego CA USA |
| Publisher | Neural Information Processing Systems (NIPS) |
| Number of pages | 4 |
| Publication status | Published - 2021 |
| Event | Advances in Neural Information Processing Systems 2021 - Online, United States of America Duration: 7 Dec 2021 → 10 Dec 2021 Conference number: 35th https://papers.nips.cc/paper/2021 (Proceedings) https://nips.cc/Conferences/2021 (Website) |
Conference
| Conference | Advances in Neural Information Processing Systems 2021 |
|---|---|
| Abbreviated title | NeurIPS 2021 |
| Country/Territory | United States of America |
| City | Online |
| Period | 7/12/21 → 10/12/21 |
| Internet address |
|