In ongoing relationships where both the collusive and non-collusive outcomes can be supported as equilibria, researchers must resolve underlying selection questions for a theoretical understanding of counterfactual policies. One theoretical guide to selection has shown promise in predicting when collusive outcomes will emerge in simple two-player repeated-game experiments. In this paper we both expand upon and experimentally test this model of selection and its underlying mechanism (strategic uncertainty). Expanding the environment to allow for an arbitrary number of participants, we stress test the model. Our results affirm the selection model as a tool for predicting when tacit collusion is likely/unlikely. In a complementary exercise, we then show that the model also predicts the collusive behavior of a standard AI agent (with a learning algorithm typical for pricing tasks), though there is some sensitivity to the form of the learning. As such, our paper serves to further bolster strategic uncertainty as a device for modeling and understanding equilibrium selection, but also the use of AI agents as a predictive tool for designing and stress-testing experiments on human subjects.
If you would like to be added to the distribution list or for further details regarding this seminar, please contact Shayne Trotman at shayne.trotman@nyu.edu