Learning Against a Strategic Agent in Principal-Agent Games

Raj Kiriti Velicheti, Subhonmesh Bose, Tamer Basar

Keywords: Incentive Design, Multiarm Bandits, Reinforcement Learning, Game Theory

TL;DR: The paper tries to understand intricacies in principal learning private type of a strategic agent

Abstract: Principal-Agent interactions, studied within the framework of incentive design problems, deal with the Principal (P) designing strategies such that the Agent's (A's) actions would favor P's cost. It is well known that when A has more information, then P faces a loss in optimality, known as information rent. While a plethora of solutions seek to devise mechanisms to tackle information asymmetry in single-stage games, we consider here the scenario of a principal who learns. Via a prototype incentive design game with continuous types and action sets, we show that P can indeed overcome the information rent through repeated interaction via an explore-then-commit (ETC) incentive policy design, when A responds myopically. We illustrate that the story is more nuanced when the agent responds in a non-myopic fashion.

Track: Short Paper

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 105