Title: Contextual Scenario Generation for Two-Stage Stochastic Programming

Abstract:

While two-stage stochastic programs (2SPs) are a staple for decision-making under uncertainty, their real-world application is frequently hindered by the extensive scenario sets required to accurately approximate conditional distributions of uncertain outcomes. To address this, we investigate contextual scenario generation: a process that leverages contextual data to learn the creation of a compact, user-defined collection of surrogate scenarios. When these scenarios are fed into a 2SP, they facilitate high-quality decisions. Current methods for scenario generation typically either overlook contextual data or incur prohibitive computational costs when such information is incorporated.

We introduce contextual scenario generation (CSG), a framework designed to learn a mapping from context to a specific set of surrogate scenarios. Our proposal features two complementary methodologies: (i) a distributional method that establishes a context-to-scenario mapping by minimizing a kernel-based distance to the conditional distribution, and (ii) a task-oriented method that optimizes the mapping to enhance decision quality by differentiating through a learned surrogate of the downstream 2SP objective. Both techniques are versatile, relying solely on the repeated resolution of underlying subproblems and 2SPs constructed from the generated scenarios. We establish finite-sample generalization guarantees and validate the robust empirical performance of our approach across various classes of 2SPs.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC