AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-26
Contextual Decision Making: Predictive or Generative
Contextual decision making can follow predict-then-optimize (PtO), which learns the conditional expected objective, or generate-then-optimize (GtO), which learns conditional uncertainty and optimizes over generated scenarios. Regret bounds show that differences in smoothness can offset GtO's extra dimensional burden, while the optimization tradeoffs for PtO and GtO differ. Synthetic experiments and a real case study show that the better approach depends on data availability and problem structure, with no universal winner.