Decision Focused Scenario Generation for Contextual Two-Stage Stochastic Linear Programming

We present a framework for scenario generation in contextual two-stage stochastic linear programs. A neural network generates scenario sets from context inputs. Rather than learning conditional distributions explicitly, the approach computes first-stage decisions through a log-barrier regularized formulation with efficiently computable derivatives via implicit differentiation. Training minimizes actual downstream costs on observed data without relying on value-function surrogates or differentiation through standard LP solvers. The method learns multi-scenario representations while avoiding high-dimensional density estimation requirements. We detail the mathematical formulation and demonstrate competitive performance with existing approaches, even when trained on small amounts of training data.

OpenReview

Recommended citation:
Download Paper