Research map / Learning theory and algorithms / Combinatorial and search optimization
A Surrogate Objective Framework for Prediction+Programming with Soft Constraints
Kai Yan, Jie Yan, Chuan Luo, Liting Chen, Qingwei Lin, Dongmei Zhang
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 the proposed surrogate soft-constraint objective instead of L1 two-stage | prediction network · Synthetic LP, N=100, size (40, 40, 0) | regret: 2.454 -> 2.258 | Improved |
| use the proposed objective instead of DF | prediction network · Synthetic LP, N=1000, size (80, 80, 40) | regret: 3.351 -> 2.781 | Improved |
| use SPO+ instead of two-stage L1/L2 losses | prediction network · Synthetic LP, N=100, size (80, 80, 40) | regret: 4.786 -> 4.939 | Hurt |
| use DF instead of two-stage losses | prediction network · Synthetic LP, N=5000, size (80, 80, 40) | regret: 2.224 -> 2.748 | Hurt |
| increase training set size from 100 to 5000 | all methods · Synthetic LP, size (80, 80, 0) | regret: 5.200 -> 2.845 | Improved |
| increase training set size from 100 to 5000, gap to two-stage | Ours vs L1 · Synthetic LP, size (40, 40, 0) | regret: 2.454 -> 1.077 | Mixed |
| add 20 soft constraints to the LP | Ours · Synthetic LP, N=1000, size 40 variables | regret: 1.346 -> 1.506 | Hurt |
| use the proposed objective instead of DF | prediction network · Portfolio optimization, 50 equities | regret (%): 4.016 -> 3.662 | Improved |
| use DF instead of L1/L2 two-stage | prediction network · Portfolio optimization, 100 equities | regret (%): 4.262 -> 3.500 | Improved |
| increase number of equities from 50 to 250 | Ours · Portfolio optimization | regret (%): 3.662 -> 3.212 | Improved |
| use the proposed objective instead of L1 | prediction network · Resource provisioning, alpha1/alpha2=100 | regret: 105.061 -> 20.829 | Improved |
| use the proposed objective instead of L1 when alpha1/alpha2=1 | prediction network · Resource provisioning, alpha1/alpha2=1 | regret: 4.267 -> 5.839 | Hurt |
Closest papers by meaning
- Risk Bounds and Calibration for a Smart Predict-then-Optimize Method (NeurIPS 2021 · similarity 0.86)
- Smart Surrogate Losses for Contextual Stochastic Linear Optimization with Robust Constraints (NeurIPS 2025 · similarity 0.85)
- Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · similarity 0.83)
- Learning Valid Dual Bounds in Constraint Programming: Boosted Lagrangian Decomposition with Self-Supervised Learning (AAAI 2025 · similarity 0.83)
- Feasibility-Aware Decision-Focused Learning for Predicting Parameters in the Constraints (NeurIPS 2025 · similarity 0.83)
- Solver-Free Decision-Focused Learning for Linear Optimization Problems (NeurIPS 2025 · similarity 0.83)
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.
