STINMALE: Strategic Interactions with Machine-Learning Algorithms: The Role of Simple Beliefs

最近的一个突出技术发展是基于机器学习（ML）算法的自动决策的作用越来越大。例如在线内容提供、产品定价、信用评分和自动驾驶。在对人类代理人之间的战略互动进行建模时，经济学家通常使用博弈论，该理论假设代理人在正确理解其环境中的因果关系和统计关系的情况下追求明确定义的目标。当一些代理是ML算法时，我们需要找到新的、易于分析的方法来建模它们如何与人类或它们之间进行交互。我在这个项目中的目标是开发这样的理论方法，并研究它们在经济环境中的影响，如寡头竞争，信贷市场或在线内容提供，包括对市场监管的潜在影响。我的理论方法的基石是观察到ML算法是“简单寻求”。它们试图从包含可观察特征数据的样本中预测结果，并通过惩罚复杂模型来克服过拟合问题（即复杂估计模型倾向于较差的样本外预测）。简单性也是机器学习算法的“可解释性”的一个方面--这是增强用户与此类算法交互意愿的重要标准。我计划制定战略和市场互动中的均衡行为概念，将简单性寻求作为形成均衡信念的标准。一个概念将侧重于基于样本的预测模型选择，而另一个概念将侧重于可解释性，作为选择模型的标准，与其预测准确性相权衡。我将把这些新的均衡概念应用到经济环境中，如信贷市场的逆向选择，动态信任博弈和寡头垄断定价，拥塞博弈和在线内容提供。

One of the salient recent technological developments has been the growing role of automated decision-making, based on machine learning (ML) algorithms. Examples include online content provision, product pricing, credit scoring and autonomous driving. When modeling strategic interactions between human agents, economists conventionally use game theory, which assumes that agents pursue well-defined objectives with a correct understanding of causal and statistical regularities in their environment. When some agents are ML algorithms, we need to find new, analytically tractable ways to model how they interact with humans or among themselves. My aim in this project is to develop such theoretical methodologies and examine their implications in economic settings such as oligopolistic competition, credit markets or online content provision, including potential implications for market regulation.The cornerstone of my theoretical approach is the observation that ML algorithms are "simplicity seeking". They attempt to predict outcomes from a sample that contains data about observable characteristics, and they overcome the overfitting problem (namely, complex estimated models' tendency toward poor out-of-sample predictions) by penalizing complex models. Simplicity is also an aspect of "explainability" of ML algorithms - an important criterion for enhancing users' willingness to interact with such algorithms.I plan to formulate notions of equilibrium behavior in strategic and market interactions that incorporate simplicity seeking as a criterion in the formation of equilibrium beliefs. One notion will focus on the sample-based selection of predictive models, while another will focus on explainability as a criterion for selecting models that is traded off against their predictive accuracy. I will apply these new equilibrium concepts to economic settings such as credit markets with adverse selection, dynamic trust games and oligopoly pricing, congestion games and online content provision.