Boot camp on CO and graph learning
summary
Claims that graph neural networks speed up mixed integer linear programming by imitating strong branching on variables using a bipartite graph representation of constraints and variables, and that supervised GNNs trained on a problem family (set covering, facility location) beat software heuristics in experiments. Implementation is root-node or hybrid (root GNN then cheaper CPU policies), not yet mainstream because of GPU/CPU transfer overhead and poor cross‑family generalization; preprocessing ML choices are already shipping in solvers.
AI-written summary · about
chapters (10)
- 0:00Workshop introduction
- 1:50Machine learning for CO
- 9:07MILP success story
- 16:28MILP solver building blocks
- 36:16ML in MILP algorithms
- 51:59Representing MILPs for ML
- 1:03:31Branching with graph neural networks
- 1:19:37Reinforcement learning in IP
- 1:26:15Preprocessing and configuration with machine learning
- 1:31:02Challenges and future
related talks
16mPreview: Graph NNs
17mBGL: GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
31mGraph-Based Machine Learning with Python
39mUsing Graph Neural Networks to Embrace the Dependency Within Your Data
13mGNNAdvisor: An Adaptive and Efficient Runtime System for GNN Acceleration on GPUs
41mBetter Than Deep Learning: Gradient Boosting Machines (GBM)
18mTC-GNN: Bridging Sparse GNN Computation and Dense Tensor Cores on GPUs
19mNeuGraph: Parallel Deep Neural Network Computation on Large Graphs