Research map / Learning theory and algorithms / Combinatorial and search optimization
Feasibility-Aware Decision-Focused Learning for Predicting Parameters in the Constraints
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Odece, a decision-focused learning framework for predicting uncertain constraint parameters in generic (non-LP) constrained optimization, combining two MLE-derived losses (IPL for infeasibility, OPL for suboptimality) via a tunable infeasibility-aversion coefficient alpha.
Problem
Predict-then-optimize when parameters appearing in constraints are uncertain; predicted constraint parameters can yield solutions infeasible under true parameters, while prior DFL methods focus on objective parameters or are limited to LP/ILP settings.
Approach
Two MLE-based losses: an Infeasibility Penalty Loss that penalizes predicted parameters whose solution violates true constraints, and an Optimality-Preserving Loss that penalizes predicted parameters that make the known true optimal solution infeasible; a single weight alpha trades off the two, with no second-stage corrective action and no LP assumption.
Key results
- On MDKP weight prediction, MSE has 33.5% infeasibility; Odece(0.8) reduces it to 6.5% (per Table) with regret rising to 0.127 from MSE's 0.063.
- On Brass alloy production, MSE has 53.2% infeasibility while Odece(0.8) reaches 17.4%, at normalized regret 0.486 vs MSE 0.169.
- Paired t-tests show Odece has significantly lower infeasibility than MSE for alpha > 0.5 and lower regret for alpha < 0.3 across problems.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase α from 0.2 to 0.8 | Odece · MDKP weight prediction test | infeasibility: 0.952 -> 0.065 | Helped |
| Increase α from 0.2 to 0.8 | Odece · MDKP weight prediction test | regret: 0.017 -> 0.127 | Hurt |
| Increase α from 0.2 to 0.8 | Odece · MDKP capacity prediction test | infeasibility: 0.791 -> 0.038 | Helped |
| Increase α from 0.2 to 0.8 | Odece · MDKP capacity prediction test | regret: 0.123 -> 0.428 | Hurt |
| Increase α from 0.2 to 0.8 | Odece · Brass alloy production test | infeasibility: 0.790 -> 0.174 | Helped |
| Increase α from 0.2 to 0.8 | Odece · Brass alloy production test | regret: 0.081 -> 0.486 | Hurt |
| Use Odece(0.8) instead of MSE | Constraint parameter prediction · MDKP weight prediction test | infeasibility: 0.335 -> 0.065 | Helped |
| Use Odece(0.8) instead of MSE | Constraint parameter prediction · MDKP capacity prediction test | infeasibility: 0.588 -> 0.038 | Helped |
| Use Odece(0.8) instead of MSE | Constraint parameter prediction · Brass alloy production test | infeasibility: 0.532 -> 0.174 | Helped |
Surprising findings
- CombOptNet and 2sPtO yield very high infeasibility (>85-99%) on several settings, and 2sPtO on Brass alloy has highly unstable results with regret above 850%.
Limitations
- Training requires solving the COP for each instance on every step, which is computationally costly; caching is left to future work.
- Experiments are limited to knapsack and a single covering LP problem, with synthetic data for MDKP.
- Normalized regret is computed only on feasible instances and can be misleading when most predictions are infeasible.
Experimental setup
Datasets: synthetic multi-dimensional knapsack instances Brass alloy production dataset (from 2sPtO work)
Benchmarks: MDKP weight prediction MDKP capacity prediction Brass alloy production
Metrics: infeasibility proportion normalized regret on feasible solutions
Baselines: MSE CombOptNet SFL 2sPtO
Builds on: CombOptNet SFL 2sPtO
Closest papers by meaning
- Solver-Free Decision-Focused Learning for Linear Optimization Problems (NeurIPS 2025 · similarity 0.90)
- DFF: Decision-Focused Fine-Tuning for Smarter Predict-Then-Optimize with Limited Data (AAAI 2025 · similarity 0.88)
- Contextual Optimization Under Model Misspecification: A Tractable and Generalizable Approach (ICML 2025 · similarity 0.85)
- Leaving the Nest: Going beyond Local Loss Functions for Predict-Then-Optimize (AAAI 2024 · similarity 0.85)
- Decision-Focused Learning with Directional Gradients (NeurIPS 2024 · similarity 0.85)
- FSNet: Feasibility-Seeking Neural Network for Constrained Optimization with Guarantees (NeurIPS 2025 · similarity 0.85)
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.