A rich class of economic problems — like monopoly regulation, contract design, and Bayesian persuasion — can be understood as games of incomplete information played by a policymaker who commits to a policy and an agent who responds. Typically, optimal policies depend on both the policymaker and the agent’s prior knowledge about the environment. To get around this, researchers often (a) assume that both the policymaker and the agent have substantial prior knowledge, or (b) assume no knowledge and optimize against the worst-case. We propose online learning as a way to combine the superior performance of (a) with the robustness of (b). We study a repeated interaction where both the policymaker and the agent may learn about the environment over time. We develop simple calibrated policies that ensure bounded regret, relative to the best-in-hindsight static policy. Our guarantees are prior-free and hold even in highly non-stationary environments. They require novel behavioral assumptions that capture concepts like "rationality'' or "unpredictability'' without relying on beliefs.

 

Modibo Camara is an Economics Ph.D. candidate at Northwestern University, working at the intersection of economics and computer science. His dissertation on bounded rationality adapts theories of computational and statistical complexity to develop more credible models of behavior. He is a member of the Online Markets Lab at Northwestern and a former intern at Microsoft Research. Prior to that, he earned a B.A. in Economics and Mathematics from the University of Pennsylvania.