EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
2898. Learning Decision-Focused Surrogate for Decision-Dependent Problems with Samples
Invited abstract in session MA-28: Advancements of OR-analytics in statistics, machine learning and data science 1, stream Advancements of OR-analytics in statistics, machine learning and data science.
Monday, 8:30-10:00Room: 065 (building: 208)
Authors (first author is the speaker)
| 1. | Zhuojun Xie
|
| Laboratoire Génie Industriel, CentraleSupélec, Université Paris-Saclay | |
| 2. | Yiping Fang
|
| CentraleSupélec | |
| 3. | Adam ABDIN
|
| Industrial Engineering Research Department, CentraleSupélec, University of Paris-Saclay |
Abstract
Decision-Focused Surrogate (DFS) framework has been recently proposed to tailor parametrically tractable proxies for intricate optimization problems with nonlinear and/or nonconvex nature. In this work, we extend the frontier of DFS to a data-driven contextual stochastic decision-dependent setting where the endogenous uncertainty is affected by decision-dependent features. Moreover, the conditional distribution remains unknown, yet historical observations are available. Specifically, we substitute the unknown conditional distribution with a parametric linear model. To ensure the decision quality for specific tasks, we employ techniques from Decision-Focused Learning (DFL), including optimization differentiation and surrogate loss construction, to acquire gradient of decision quality concerning model's parameters. We examine the performance of DFS in a battery energy management problem within a renewable power system considering decision-dependent degradation. Our results demonstrate that compared to a surrogate model trained using standard regression loss, DFS can achieve better decision quality, especially when the model is mis-specified. Despite the higher computational expense of learning DFS, our empirical analysis reveals that the trained surrogate can generate high-quality solutions for instances under similar conditions.
Keywords
- Decision Theory
- Optimization Modeling
- OR in Energy
Status: accepted
Back to the list of papers