Upcoming Event: Center for Autonomy Seminar

Tractable Design of Dynamical Policies for Large Populations of Learning Agents

Jair Guilherme Certorio, Ph.D. Candidate, University of Maryland

Abstract

Designing effective incentives for large populations of agents is challenging, especially when their choices influence an exogenous process that we aim to control. The interplay between agent dynamics and the exogenous system complicates incentive design, particularly when trajectory constraints on the exogenous process must be satisfied.

In this talk, we present a novel approach to designing such incentives within the framework of evolutionary game theory and population games. Our method leverages broad properties of the agents' learning behavior, enabling incentive design independent of the specific learning rule used by the agents. We derive bounds on the system's trajectory related to our incentive's parameters, allowing a policymaker to tune them to meet anytime constraints on the exogenous system's state. Furthermore, we present fundamental results on population games, including an extension of potential games to time-varying settings and an analysis of hybrid learning rules—that is, when agents' learning combines several canonical rules, such as best-response and pairwise comparison. Applications of our work include mitigating epidemics and alleviating traffic congestion, and similar approaches can be applied to various problems in decentralized control.

Biography

Jair Certório is a PhD Candidate at the University of Maryland, College Park, advised by Professors Nuno Martins and Richard La. His research focuses on applying control theory and evolutionary game theory to study large populations of agents interacting with other dynamical systems. Prior to his PhD, he received an undergraduate degree in electrical engineering from Universidade Estadual de Maringá in Brazil and worked as an embedded systems engineer.