Research map / Learning theory and algorithms / Combinatorial and search optimization
MEGO: Learning Mixture-of-Experts for General-Purpose Binary Optimization
Shengcai Liu, Zi Wang, Yew-Soon Ong, Yao Xin, Ke Tang
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 |
|---|---|---|---|
| add decoder fine-tuning and source-target mapping (ablation) | MEGO · 72 test instances (6 classes) | Wilcoxon W-D-L vs no mapping: 67 W, 4 D, 1 L | Improved |
| sample in source space instead of searching the latent space (ablation) | MEGO · 72 test instances (6 classes) | Wilcoxon W-D-L vs latent-space search: 53 W, 16 D, 3 L | Improved |
| use all-one padding instead of zero padding (ablation) | MEGO · Dim=20 instances (18) | Wilcoxon W-D-L: 2-15-1 | No clear effect |
| use random padding instead of zero padding (ablation) | MEGO · Dim=20 instances (18) | Wilcoxon W-D-L: 2-14-2 | No clear effect |
| correlation-based routing instead of random selection of same number of experts (SAMPLE) (ablation) | MEGO · 72 test instances | Wilcoxon W-D-L: 19-45-8 | Improved |
| activate multiple relevant experts instead of one random expert (MEAN) (ablation) | MEGO · 72 test instances | Wilcoxon W-D-L: 71-0-1 | Improved |
| compare MEGO routing against hindsight best single expert (MAX) (ablation) | MEGO · 72 test instances | Wilcoxon W-D-L: 9-15-48 | Hurt |
| use MEGO outputs to warm-start GA, HC and BO | GA/HC/BO · CA class | W-D-L MEGO+X vs X: 4-8-0, 5-7-0, 6-6-0 | Improved |
| use MEGO outputs to warm-start GA, HC and BO | GA/HC/BO · CIM class | W-D-L MEGO+X vs X: 9-3-0, 11-1-0, 12-0-0 | Improved |
| use MEGO outputs to warm-start GA, HC and BO | GA/HC/BO · AS class | W-D-L MEGO+X vs X: 9-3-0, 12-0-0, 11-1-0 | Improved |
| warm-start BO with MEGO against specialized SMARTEST | BO · CA, 12 instances, #FEs=800 | W-D-L MEGO+BO vs SMARTEST: 5-7-0 | Improved |
| warm-start BO with MEGO | BO · CA, 12 instances, #FEs=800 | W-D-L MEGO+BO vs BO: 6-6-0 | Improved |
Closest papers by meaning
- BOPO: Neural Combinatorial Optimization via Best-anchored and Objective-guided Preference Optimization (ICML 2025 · similarity 0.86)
- RoME: Domain-Robust Mixture-of-Experts for MILP Solution Prediction across Domains (NeurIPS 2025 · similarity 0.86)
- Solver-Free Decision-Focused Learning for Linear Optimization Problems (NeurIPS 2025 · similarity 0.83)
- Runtime Analysis of the SMS-EMOA for Many-Objective Optimization (AAAI 2024 · similarity 0.83)
- LLMOPT: Learning to Define and Solve General Optimization Problems from Scratch (ICLR 2025 · similarity 0.83)
- Multi-Objective Bayesian Optimization with Active Preference Learning (AAAI 2024 · similarity 0.82)
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.
