Research map / Learning theory and algorithms / Prediction and decision losses
The Perils of Learning Before Optimizing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 end-to-end learning instead of two-stage learning | Demand-weighted facility location · Synthetic benchmark; positive correlation between demand and travel time | test performance | Helped |
| Use end-to-end learning instead of two-stage learning | Demand-weighted facility location · Synthetic benchmark; ρ=0 | test performance | No clear effect |
| Use end-to-end learning instead of two-stage learning | Demand-weighted facility location · Synthetic benchmark; negative correlation | test performance | No clear effect |
| Use end-to-end learning instead of two-stage learning | Demand-weighted facility location · Synthetic benchmark; all tested ρ values | test performance | No clear effect |
| Model each target's expectation separately rather than the expected product | Demand-weighted facility location · Synthetic benchmark; positive correlation | test performance | Hurt |
| Model each target's expectation separately rather than the expected product | Demand-weighted facility location · Synthetic benchmark; negative correlation | test performance | No clear effect |
| Use two-stage prediction with product coefficients instead of end-to-end learning | Optimization with nonlinear target parameterization · Constructed problem with γ(a,b)=a·b | stochastic objective loss: C + \sum_{i=1}^{d} (\frac{N}{2}-\epsilon)^2 1.0 \approx \frac{d N^{2}}{8} | Hurt |
| Use a two-stage method with nonlinear γ rather than an end-to-end method | Optimization with multiple prediction targets · Constructed problem for any nonlinear γ | stochastic objective loss | Hurt |
| Use two-stage prediction when γ is linear | Optimization with multiple prediction targets · Linear target parameterization | stochastic objective loss | No clear effect |
| Use two-stage prediction for product coefficients when targets are conditionally independent | Optimization with multiple prediction targets · γ(y1,y2)=y1·y2 and (Y1 ⟂ Y2)|X | stochastic objective loss | No clear effect |
| Use end-to-end prediction instead of the two-stage proxy | Two-stage minimum cost flow · Stochastic optimization | two-stage to end-to-end loss ratio: \Omega\left(2^{d}\right) | Helped |
| Use end-to-end prediction instead of the two-stage proxy | Two-stage stochastic set cover · Stochastic optimization | two-stage to end-to-end loss ratio: \Omega\left(\sqrt{d} \frac{\log (\log (d))}{\log(d)}\right) | Helped |
Closest papers by meaning
- End-to-End Learning for Stochastic Optimization: A Bayesian Perspective (ICML 2023 · similarity 0.90)
- Leaving the Nest: Going beyond Local Loss Functions for Predict-Then-Optimize (AAAI 2024 · similarity 0.88)
- Multi-Stage Predict+Optimize for (Mixed Integer) Linear Programs (NeurIPS 2024 · similarity 0.86)
- The Bias-Variance Tradeoff in Data-Driven Optimization: A Local Misspecification Perspective (NeurIPS 2025 · similarity 0.86)
- Contextual Optimization Under Model Misspecification: A Tractable and Generalizable Approach (ICML 2025 · similarity 0.85)
- Solver-Free Decision-Focused Learning for Linear Optimization Problems (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.