Research map / Learning theory and algorithms / Combinatorial and search optimization
Landscape Surrogate: Learning Decision Losses for Mathematical Optimization Under Partial Information
Arman Zharmagambetov, Brandon Amos, Aaron M Ferber, Taoan Huang, Bistra Dilkina, Yuandong Tian
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 |
|---|---|---|---|
| reuse pretrained landscape surrogate M_w without solver access (ablation) | LANCER-zero · Stochastic shortest path 15x15, normal deadline | runtime (s): 62.3 -> 2.7 | Improved |
| reuse pretrained landscape surrogate M_w without solver access (ablation) | LANCER-zero · Stochastic shortest path 15x15, tight deadline | objective: 0.434 -> 0.432 | No clear effect |
| reuse pretrained landscape surrogate M_w without solver access (ablation) | LANCER-zero · Stochastic shortest path 15x15, loose deadline | objective: 0.556 -> 0.556 | No clear effect |
| reduce M_w max iterations from 10 to 5 (lr 0.0005) (ablation) | LANCER · Nonlinear shortest path | objective: 0.4651 -> 0.4590 | Hurt |
| increase M_w max iterations from 10 to 20 (lr 0.001) (ablation) | LANCER · Nonlinear shortest path | objective: 0.4649 -> 0.4612 | Hurt |
| lower c_theta learning rate from 0.001 to 0.0005 (ablation) | LANCER · Nonlinear shortest path | objective: 0.4649 -> 0.4612 | Hurt |
| raise c_theta learning rate from 0.001 to 0.01 (ablation) | LANCER · Nonlinear shortest path | objective: 0.4649 -> 0.4650 | No clear effect |
| raise c_theta max iterations from 10 to 20 (ablation) | LANCER · Nonlinear shortest path | objective: 0.4649 -> 0.4648 | No clear effect |
| vary M_w depth and width from 1x50 to 3x200 (ablation) | LANCER · Nonlinear shortest path | objective: 0.4644 -> 0.4642 | No clear effect |
| use LANCER instead of MDFL, which uses exact gradients | LANCER · Portfolio selection (QP) | normalized test decision loss: 0.52 -> 0.53 | Hurt |
| use LANCER instead of 2-Stage | LANCER · Portfolio selection (QP) | normalized test decision loss: 0.57 -> 0.53 | Improved |
| use LANCER-zero instead of SurCo-zero | LANCER-zero · Nonlinear shortest path | objective | Improved |
Closest papers by meaning
- Solver-Free Decision-Focused Learning for Linear Optimization Problems (NeurIPS 2025 · similarity 0.87)
- Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · similarity 0.86)
- Incorporating Surrogate Gradient Norm to Improve Offline Optimization Techniques (NeurIPS 2024 · similarity 0.86)
- Solving hidden monotone variational inequalities with surrogate losses (ICLR 2025 · similarity 0.86)
- Gradient Descent as Loss Landscape Navigation: a Normative Framework for Deriving Learning Rules (NeurIPS 2025 · similarity 0.86)
- A Generalization Result for Convergence in Learning-to-Optimize (ICML 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.
