【学术讲座】面对战略参与者的实用绩效性策略学习

 Practical Performative Policy Learning with Strategic Agents 

课程信息 | INFOMATION

讲者：黎波 (清华大学经济管理学院)

时间：2024年4月24日（周三） 10时

地点:   同济大厦A座502教室

语言：中文报告（英文演示文稿）

内容摘要 | ABSTRACT

There has been a recent growing interest in training predictive models in strategic environments, including strategic classification (Hardt et al 2016} and performative prediction (Perdomo et al 2020).  Methodologies in strategic classification predominantly hinge on the specification of an agent's utility function and decision-making process, while algorithms aimed at achieving performative optimality typically rely on parametric assumptions regarding high dimensional data distribution map or employ canonical bandit experiments to learn this map in a quite often impractically slow pace. This paper approaches the performative policy learning problem as a complex causal question and derives tractable low-dimensional structure in the distribution shift pattern in many typical strategic settings. Multi-stage learning algorithms are proposed to learn the distribution map and the policy with good performance on policy value. We conduct extensive experiments on synthetic and real data to demonstrate the practical efficacy of our methods.

嘉宾介绍 | GUEST BIO

黎波，清华大学经济管理学院管理科学与工程系副教授，2002年本科毕业于北京大学数学科学学院数学系，2006年博士毕业于加州大学伯克利分校统计系。他近年的研究兴趣包括平台实验设计、因果学习与推断、数据驱动的决策制定、可信机器学习、机器学习与经济学的交叉等。已在经济、管理、统计、人工智能等多个领域的期刊及会议发表论文60余篇，包括Biometrika, JRSSB, Management Science, TKDE，管理科学学报等期刊以及NeurIPS, ICML, ICLR, AAAI, KDD等会议。