Contextual Inverse Optimization: Offline and Online Learning

Learning from data are critical across applications. However, in many applications, past data only gives partial information about the future. In “Contextual Inverse Optimization: Offline and Online Learning,” Besbes, Fonseca, and Lobel study a general setting in which historical data are associated with observations of past optimal actions from experts in specific contexts but without the underlying rewards associated with these actions. To what extent can one “reverse engineer” the underlying decision-making process of experts and mimic them? The authors develop results that quantify the performance that is achievable given the data at hand in two types of settings: the offline setting in which data have already been collected and the online setting in which data are collected “on the fly.”