Research map / Learning theory and algorithms / Prediction and decision losses
Differentiable Distributionally Robust Optimization Layers
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.
Develops differentiable distributionally robust optimization layers for mixed-integer DRO with parameterized SOC ambiguity sets, using a dual-view energy-based surrogate with importance sampling for gradients, and applies them to contextual decision-focused learning.
Problem
Decision-focused learning for distributionally robust optimization requires differentiating decisions with respect to an ambiguity set, which is unsolved, especially for mixed-integer decisions.
Approach
A dual-view methodology differentiates continuous decisions via implicit differentiation of the cone program and handles discrete decisions via constrained optimization, approximated by a differentiable energy-based surrogate value function with self-normalized importance sampling for the gradient.
Key results
- Decision-focused learning improves over the prediction-focused benchmark by average 21.4%, 18.7%, and 18.1% on the three SOC ambiguity sets in the newsvendor experiment.
- On 40-dimensional continuous portfolio, DFL achieves average profit 0.289 vs 0.082 for prediction-focused and 0.268 for the prior end-to-end DRO method.
- On 60-dimensional mixed-integer portfolio, DFL achieves 0.2145 average profit vs 0.0479 for prediction-focused (347% improvement).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train ambiguity-set parameters by directly minimizing decision loss rather than using prediction-focused learning | Contextual DRO for multi-item newsvendor · Three SOC ambiguity sets | percentage optimality gap: 21.4%, 18.7%, and 18.1% average improvements | Helped |
| Use decision-focused learning instead of prediction-focused learning | Portfolio management, continuous, 40 assets · Portfolio management, continuous, 40 assets | average percentage profit: 0.082 -> 0.289 | Helped |
| Use the proposed decision-focused learning method instead of the prior end-to-end DRO method | Portfolio management, continuous, 40 assets · Portfolio management, continuous, 40 assets | average percentage profit: 0.268 -> 0.289 | Helped |
| Use decision-focused learning instead of prediction-focused learning | Mixed-integer portfolio management · 20 assets | average percentage profit: 0.0583 -> 0.1561 | Helped |
| Use decision-focused learning instead of prediction-focused learning | Mixed-integer portfolio management · 40 assets | average percentage profit: 0.0634 -> 0.1763 | Helped |
| Use decision-focused learning instead of prediction-focused learning | Mixed-integer portfolio management · 60 assets | average percentage profit: 0.0479 -> 0.2145 | Helped |
| Use decision-focused learning instead of prediction-focused learning | Mixed-integer portfolio management · 60 assets, 15 SOC constraints | average percentage profit: 0.0357 -> 0.0762 | Helped |
| Use decision-focused learning instead of prediction-focused learning | Mixed-integer portfolio management · 60 assets, 30 SOC constraints | average percentage profit: 0.0491 -> 0.1349 | Helped |
| Use decision-focused learning instead of prediction-focused learning | Mixed-integer portfolio management · 60 assets, 60 SOC constraints | average percentage profit: 0.0479 -> 0.2145 | Helped |
Surprising findings
- The prediction-focused benchmark uses full knowledge of the conditional distribution, yet decision-focused learning still outperforms it.
Limitations
- Heavy computational burden for large-scale problems since cvxpylayers and Gurobi run on CPU.
- Uniqueness of discrete solutions is assumed only almost surely.
- Surrogate convergence holds only asymptotically as the energy temperature lambda goes to zero.
Experimental setup
Datasets: synthetic multi-item newsvendor data synthetic portfolio return data
Benchmarks: multi-item newsvendor (toy) 40-dimensional portfolio management mixed-integer portfolio management (20, 40, 60 dims)
Metrics: percentage optimality gap average percentage profit improvement over prediction-focused benchmark
Baselines: prediction-focused learning end-to-end robust portfolio method (Ma et al. style DRO learning)
Builds on: cvxpylayers SPO+ differentiating through a cone program
Closest papers by meaning
- Distributionally Robust Optimization via Generative Ambiguity Modeling (ICLR 2026 · similarity 0.89)
- Wasserstein Distributionally Robust Optimization through the Lens of Structural Causal Models and Individual Fairness (NeurIPS 2024 · similarity 0.88)
- Large-Scale Non-convex Stochastic Constrained Distributionally Robust Optimization (AAAI 2024 · similarity 0.87)
- Distributionally Robust Optimization with Bias and Variance Reduction (ICLR 2024 · similarity 0.86)
- Drago: Primal-Dual Coupled Variance Reduction for Faster Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.85)
- Geometry-Calibrated DRO: Combating Over-Pessimism with Free Energy Implications (ICML 2024 · 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.