预约 | 3月31日9:00，Randall Berry教授分享基于不可靠观测的观察学习

Observational Learning with Unreliable observations

Many online platforms enable agents to learn from observing other’s actions. Such observational learning can be modeled via a Bayesian game in which agents sequentially choose an action to take. Agents base their decision on whatever private information they may have as well as the observed actions of the other agents. Such models can lead to herding behavior or information cascades in which agents eventually “follow the crowd”. In this talk, we will discuss a line of work that considers the impact of different ways in which an agent’s observations may be corrupted including having some fraction of the agent’s lie about their action or having some fraction of the observations corrupted by noise. Interestingly, these models often show a non-monotonic behavior in the “level" of corruption, so that in some cases poorer observations may lead to better performance. We also discuss the relationship to the classic Blackwell ordering of information structures for single agent decision problems.