Research map / Learning theory and algorithms / Combinatorial and search optimization
Unsupervised Learning for Combinatorial Optimization Needs Meta Learning
Haoyu Peter Wang, Pan Li
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 |
|---|---|---|---|
| Train with a meta-learning objective for instance-wise adaptation instead of minimizing average instance loss | EGN · MC, Twitter | ApR: 0.926 ± 0.113(0.17) -> 0.976 ± 0.048(0.17) | Improved |
| Train with a meta-learning objective | EGN · MC, RB 200 | ApR: 0.820 ± 0.188 (0.26) -> 0.834 ± 0.178 (0.26) | No clear effect |
| Train with a meta-learning objective | EGN · MC, Twitter to RB2000 | ApR: 0.594±0.210(0.07) -> 0.690±0.201(0.07) | Improved |
| Train with a meta-learning objective | EGN · MVC, RB 500 | ApR: 1.021 ± 0.002 (0.48) -> 1.016 ± 0.002 (0.48) | Improved |
| Add one-step fine-tuning at test time (ablation) | Meta-EGN · MC, RB 500 | ApR: 0.834 ± 0.198 (0.29) -> 0.878 ± 0.181 (0.89) | Improved |
| Add one-step fine-tuning at test time (ablation) | Meta-EGN · MVC, Twitter | ApR: 1.019 ± 0.017(0.29) -> 1.017 ± 0.017(0.80) | Improved |
| Use Meta-EGN to modify DGA solutions | DGA · MIS, RRGs, 10^5 nodes | ApR improvement: 0.0032; 0.0045; 0.0051; 0.0050 | Improved |
| Use Meta-EGN to modify RGA solutions | RGA · MIS, RRGs, 10^5 nodes | ApR improvement: 0.0871; 0.1045; 0.1083; 0.0969 | Improved |
| Reduce training data from 750 instances to 64 (ablation) | Meta-EGN · MC, Twitter | ApR: 0.976±0.048 -> 0.970±0.058 | Hurt |
| Reduce training data from 750 instances to 64 (ablation) | EGN · MC, Twitter | ApR: 0.926±0.113 -> 0.909±0.122 | Hurt |
| Train only on degree-3 RRGs rather than the mixed-degree training set (ablation) | Meta-EGN · MIS, RRGs at other degrees | ApR | Hurt |
| Train only on degree-20 RRGs rather than the mixed-degree training set (ablation) | Meta-EGN · MIS, RRGs at other degrees | ApR | Mixed |
Closest papers by meaning
- Unsupervised Learning for Combinatorial Optimization with Principled Objective Relaxation (NeurIPS 2022 · similarity 0.91)
- Tackling Prevalent Conditions in Unsupervised Combinatorial Optimization: Cardinality, Minimum, Covering, and More (ICML 2024 · similarity 0.88)
- GOAL: A Generalist Combinatorial Optimization Agent Learner (ICLR 2025 · similarity 0.86)
- Geometric Algorithms for Neural Combinatorial Optimization with Constraints (NeurIPS 2025 · similarity 0.86)
- Controlling Continuous Relaxation for Combinatorial Optimization (NeurIPS 2024 · similarity 0.86)
- Memory-Reduced Meta-Learning with Guaranteed Convergence (AAAI 2025 · similarity 0.86)
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.
