标题： 自我实现的老虎机：算法决策中的内生性溢出和动态选择 显示英文标题

标题： Self-fulfilling Bandits: Endogeneity Spillover and Dynamic Selection in Algorithmic Decision-making

摘要： 在本文中，我们研究算法决策中的内生性问题，其中数据和行动是相互依赖的。 当上下文多臂老虎机模型中存在内生协变量时，由于协变量的内生性会蔓延到行动中，会出现一种新的偏差（自我实现偏差）。 我们提出了一类算法，通过将工具变量纳入领先的在线学习算法来纠正这种偏差。 这些算法还达到了与无内生性情况下的最佳已知下限相匹配的遗憾水平。 为了建立理论性质，我们开发了一种通用技术，以解开数据和行动之间的相互依赖性。 显示英文摘要

摘要： In this paper, we study endogeneity problems in algorithmic decision-making where data and actions are interdependent. When there are endogenous covariates in a contextual multi-armed bandit model, a novel bias (self-fulfilling bias) arises because the endogeneity of the covariates spills over to the actions. We propose a class of algorithms to correct for the bias by incorporating instrumental variables into leading online learning algorithms. These algorithms also attain regret levels that match the best known lower bound for the cases without endogeneity. To establish the theoretical properties, we develop a general technique that untangles the interdependence between data and actions.