Learning to Persuade
In the standard Bayesian persuasion model, an informed sender looks to design a signaling scheme to partially reveal the information to an uninformed receiver, so as to influence the behavior of the receiver. This kind of strategic interaction abounds in the real world. However, the standard model relies crucially on some stringent assumptions that usually do not hold in reality. For example, the sender knows the receiver's utility function and the receiver's behavior is completely rational. In this paper, we aim to relax these assumptions using techniques from the AI domain. We put forward a framework that contains both a receiver model and a sender model. We first train a receiver model through interactions between the sender and the receiver. The model is used to predict the receiver's behavior when the sender's scheme changes. Then we update the sender model to obtain an approximately optimal scheme using the receiver model. Experiments show that our framework has comparable performance to the optimal scheme.
PDF Abstract