ORIE Colloquium: Henry Lam (Columbia)

On the (im)possibility of beating "plug-in" optimization Many modern decision-making pipelines entail optimization problems driven by data and the need to mitigate uncertainty. This has ignited a range of proposed approaches in recent years, including various (distributionally) robust optimization and regularized formulations. We explore the fundamental question of when these approaches can outperform optimization based on simple "plug-in" estimates of unknown quantities. On one hand, we argue "no free lunch is possible" for these approaches in standard settings, in a strong sense that accounts for the entire distribution of regret. On the other hand, when leveraging geometrically effective side information, we argue that outperformance is possible, but via choosing hyperparameters significantly larger than typically suggested in the literature, and via a connection to the control-variate principle in Monte Carlo simulation. Bio: Henry Lam is an associate professor in the Department of Industrial Engineering and Operations Research at Columbia University. His research focuses on data-driven decision-making methodologies, including simulation, optimization under uncertainty and rare-event analysis, especially regarding their robustness and statistical issues when integrating with ML/AI tools. His works have been recognized by venues such as the NSF CAREER Award and NSA Young Investigator Award, industry funding awards from Google, Adobe, JPMorgan and NYC agencies including the Fire Department and the Department of Correction, and other paper awards. He serves as co-Area Editor for the Stochastic Models area in Math of OR, Area Editor for the Stochastic Models and Data Science area in OR Letters, Associate Editor for several journals (such as MS, OR, MSOM), and Chair of the INFORMS Applied Probability Society. Henry holds a Ph.D. degree in statistics from Harvard University and B.S. degree in actuarial science from the University of Hong Kong.