Research map / Learning theory and algorithms / Combinatorial and search optimization
CombOptNet: Fit the Right NP-Hard Problem by Learning Integer Programming Constraints
Anselm Paulus, Michal Rolinek, Vit Musil, Brandon Amos, Georg Martius
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CombOptNet instead of BB-GM with unary costs only | keypoint matching · Keypoint matching, 4×4 through 7×7 | per-variable accuracy (%): 84.3 -> 83.1; 81.6 -> 80.7; 79.0 -> 78.6; 76.5 -> 76.1 | Hurt |
| Use CombOptNet instead of BB-GM with unary and quadratic costs | keypoint matching · Keypoint matching, 4×4 through 7×7 | per-variable accuracy (%): 84.3 -> 83.1; 82.9 -> 80.7; 80.5 -> 78.6; 79.8 -> 76.1 | Hurt |
| Replace the Δ basis with the canonical basis (ablation) | CombOptNet · Random Constraints, binary, 4 ground-truth constraints | evaluation accuracy (%): 77.4 + 13.5 -> 14.4 + 3.2 | Hurt |
| Replace the Δ basis with the canonical basis (ablation) | CombOptNet · Random Constraints, dense, 4 ground-truth constraints | evaluation accuracy (%): 29.6 + 10.4 -> 26.8 + 4.1 | No clear effect |
| Replace the Δ basis with the canonical basis (ablation) | CombOptNet · Weighted Set Covering, 8 ground-truth constraints | evaluation accuracy (%): 79.7 + 12.1 -> 0.2 + 0.3 | Hurt |
| Replace the softened minimum with a hard minimum (ablation) | CombOptNet · Random Constraints, binary, 4 ground-truth constraints | evaluation accuracy (%): 77.4 + 13.5 -> 55.4 + 18.9 | Hurt |
| Replace the softened minimum with a hard minimum (ablation) | CombOptNet · Random Constraints, dense, 4 ground-truth constraints | evaluation accuracy (%): 29.6 + 10.4 -> 21.4 + 10.7 | Hurt |
| Replace the softened minimum with a hard minimum (ablation) | CombOptNet · Weighted Set Covering, 8 ground-truth constraints | evaluation accuracy (%): 79.7 + 12.1 -> 45.1 + 14.1 | Hurt |
| Increase the soft-minimum temperature from τ=0.5 to τ=1.0 (ablation) | CombOptNet · Weighted Set Covering, 8 ground-truth constraints | evaluation accuracy (%): 79.7 + 12.1 -> 70.3 + 15.5 | Hurt |
| Replace the MSE loss with L0 loss (ablation) | CombOptNet · Random Constraints, dense, 4 ground-truth constraints | evaluation accuracy (%): 29.6 + 10.4 -> 15.3 + 4.3 | Hurt |
| Increase the number of learnable constraints from 1×m to 4×m (ablation) | CombOptNet · Random Constraints, binary, 8 ground-truth constraints | evaluation accuracy (%): 46.5 + 12.4 -> 77.7 + 3.2 | Improved |
| Increase the number of learnable constraints from 1 to 2 (ablation) | CombOptNet · Weighted Set Covering, 8 ground-truth constraints | evaluation accuracy (%): 79.7 + 12.1 -> 99.3 + 0.8 | Improved |
Later papers that compared with this method
2 later accepted papers report CombOptNet as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Two-Stage Predict+Optimize for MILPs with Unknown Parameters in Constraints (NeurIPS 2023 · better on 3 of 3 shared results)
- A Solver-free Framework for Scalable Learning in Neural ILP Architectures (NeurIPS 2022 · better on 10 of 10 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MIP-GNN: A Data-Driven Framework for Guiding Combinatorial Solvers (AAAI 2022 · similarity 0.87)
- Geometric Algorithms for Neural Combinatorial Optimization with Constraints (NeurIPS 2025 · similarity 0.86)
- BIPNN: Learning to Solve Binary Integer Programming via Hypergraph Neural Networks (NeurIPS 2025 · similarity 0.86)
- Neural Solver Selection for Combinatorial Optimization (ICML 2025 · similarity 0.85)
- GOAL: A Generalist Combinatorial Optimization Agent Learner (ICLR 2025 · similarity 0.85)
- Are Graph Neural Networks Optimal Approximation Algorithms? (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
