Online learning in Principal-Agent problems
The hidden-action principal-agent framework is a fundamental model to study the interaction between two parties with misaligned incentives. In this setting, a principal obtains a reward according to the outcome of a hidden, costly action taken by an agent. In and by itself, the agent is not incentivized to exert any effort, as they take the full cost of an action without any reward. To incentive the agent to take a favorable action, the principal has first to commit to a contract, which specifies payments according to the outcome of the agent’s hidden action. ; Recently there has been a spike of interest in the computational aspects of this problem, but little attention has been devoted to scenarios with only partial information available. In this project, we address these settings from an online learning point of view. From single to multi-agent settings, we will investigate how to learn an optimal contract over repeated interactions, given little to none initial knowledge on the game’s parameters. At the same time, we will address the tradeoff between the optimality of the contracts employed and their simplicity and robustness. ;
Back to Current Students