PRISE and NeurPRISE: Problem-Driven Scenario Reduction for Robust Optimization
Two-Stage Robust Optimization (2RO) with discrete uncertainty is computationally prohibitive. Scenario reduction selects a small representative subset to enable tractable computation, but existing methods are problem-agnostic. The paper introduces PRISE, a problem-driven sequential lookahead heuristic that evaluates each scenario's marginal impact. PRISE yields high-quality subsets but is computationally expensive. To address this, NeurPRISE uses a GNN-Transformer neural surrogate to encode per-scenario structure and cross-scenario interactions, reducing computational cost.
Key facts
- Two-Stage Robust Optimization (2RO) with discrete uncertainty is challenging.
- Scenario reduction selects a small representative subset of scenarios.
- Existing methods are problem-agnostic, ignoring feasible region and recourse structure.
- PRISE is a problem-driven sequential lookahead heuristic.
- PRISE evaluates the marginal impact of each scenario.
- PRISE is computationally expensive at scale.
- NeurPRISE is a neural surrogate model with GNN-Transformer backbone.
- NeurPRISE encodes per-scenario structure via graph convolution and captures cross-scenario interactions.
Entities
—