Reinforcement Learning for Indirect Mechanism Design

Abstract

Incentive mechanisms such as auctions are tools commonly used for resource allocation. Most of the mechanism design literature concerns itself with direct mechanisms, where agents report their preference type. However, in settings where it is infeasible for agents to communicate their full preference type, we require indirect mechanisms equipped with a simpler message space. We introduce and formalize the Generalized Eating Mechanism (GEM), a large parametric class of indirect mechanisms. We also formulate the mechanism design problem as a Markov Decision Process and use reinforcement learning (RL) algorithms to train good mechanisms within a subclass of GEM. Our RL mechanisms are able to achieve optimal or almost optimal performance in static serial dictatorship, dynamic serial dictatorship, and static price settings.