Incentive Design in the Machine Learning Age

Abstract

This dissertation investigates the design of incentives in multi-agent systems where traditional assumptions -- such as fully rational agents and omniscient principals -- do not hold. As real-world systems increasingly involve learning-based decision-makers, either human or algorithmic, this work explores how learning alters the landscape of incentive design. The dissertation is organized into three parts. The first part focuses on incentive design by learning principals, specifically in information and mechanism design settings. For information design, this part introduces novel algorithms that allow a principal to learn an agent's non-Bayesian belief updating process, such as a subjective prior or cognitive bias, via strategic interaction with the agent. For mechanism design, this part examines how a coordinator can learn to compute Bayes correlated equilibria in non-truthful auctions using limited samples of agent types. The second part studies incentive design for learning agents, who are modeled as boundedly rational learners rather than best responders. This part first presents, for a general class of principal-agent problems, a reduction from no-regret learning agents to approximately best-responding agents, enabling a precise analysis of the principal's performance. It then characterizes the convergence properties of multi-agent learning in first-price auction games, identifying when convergence to equilibrium is possible. The third part explores incentive issues in deployed machine learning systems, with a case study on recommender systems. It demonstrates that the strategic behaviors by content creators can exacerbate polarization, even under diversity-promoting algorithms, and proposes alternative algorithmic designs that mitigate these effects. Collectively, this dissertation lays foundational insights for designing systems that are robust to the incentives of learning-based, data-driven participants.