A2067
Title: Causal adaptive learning for decision making at scale
Authors: Maria Dimakopoulou - Uber (United States) [presenting]
Abstract: Sequential decision making and accurate model estimation from adaptively collected data lie at the heart of decision making at scale. Reliable decision making that adapts to new data via contextual bandit or reinforcement learning algorithms requires accurate model estimation. The focus is on how to incorporate state-of-the-art methods from the causal inference literature into model estimation for decision making systems and how to pair them with efficient exploration strategies such as Thompson Sampling. The performance gains unlocked by this approach in the presence of real-world challenges such as selection bias, covariate shift, model misspecification, and bias due to adaptive data collection are discussed.
