Policy Learning with Competing Agents
研究了当人们为有限资源竞争时，如何通过数据驱动的决策规则来分配资源，并开发了一种基于局部实验的算法，以在竞争环境下最大化社会福利。
In this Issue How to allocate limited resources when people compete for them? Data-driven decision rules are increasingly used to allocate limited resources. However, in domains such as college admissions and job hiring, these policies can be gamed and may perform poorly when people can strategically change their behavior and compete for the resource. Given a fixed selection criterion for scoring individuals, competition causes the threshold for receiving the resource to oscillate and change over time. As a concrete example, the median SAT score of an accepted student at a university can evolve over time because of competition. In “Policy Learning with Competing Agents,” Stanford researchers Roshni Sahoo and Stefan Wager demonstrate that this process can stabilize and reach an equilibrium. They develop an algorithm based on a local experimentation scheme to learn data-driven decision rules that maximize social welfare in the presence of competition.