Dissecting the Impact of Model Misspecification in Data-driven OptimizationOnline
Data-driven optimization aims to translate machine learning predictions into decisions, but the right way to do so depends on whether the model is correctly specified. This talk dissects when the standard two-stage estimate-then-optimize approach outperforms integrated estimation-optimization, and when it does not, drawing the line in terms of how misspecified the predictor is. The speakers also connect to related work on the bias-variance tradeoff under local misspecification, with implications for learning-based decision systems in pricing, retail, and revenue management.
Speakers: Prof. Adam N. Elmachtoub, Haixiang Lan (Columbia University)