Online learning in the presence of strategic adversary

Abstract

This thesis offers a comprehensive exploration of the online learning problem in which an agent needs to strategise against a strategic adversary (also known as a no-regret adversary). Through examination of three interrelated settings, we have devised novel algorithms that achieve improved performance guarantees and last round convergence to the Nash Equilibrium both in theoretical and empirical contexts. Our findings open the door to further investigation of complex problems in online learning and game theory, where strategic adversaries play a crucial role in a multitude of applications.
In the first of the three main chapters comprising our study, we examine the problem of playing against a strategic adversary under a two-player zero-sum game setting. In this scenario, we introduce a new no-dynamic regret algorithm, namely the Last Round Convergence of Asymmetric Games (LRCA), that achieves last round convergence to the minimax equilibrium. Building on this work, the second main chapter investigates the more general problem of online linear optimization and proposes several new algorithms, including Online Single Oracle (OSO), Accurate Follow the Regularized Leader (AFTRL), and Prod-Best Response algorithm (Prod-BR). These algorithms achieve state-of-the-art performance guarantees against a strategic adversary, such as no-forward regret and no-dynamic regret. Additionally, we show that a special case of AFTRL, the Accurate Multiplicative Weights Update (AMWU), can achieve last round convergence to the Nash equilibrium in self-play settings. In the third and final main chapter, we extend our results to the challenging setting of Online Markov Decision Processes (OMDPs), which have many significant applications in practice. Here, we propose two new algorithms, MDP-Online Oracle Expert (MDP-OOE) and Last Round Convergence-OMDP (LRC-OMDP), that achieve no-policy regret and last round convergence to the Nash equilibrium, respectively, against a strategic adversary.

More information

Identifiers

ORCID for Le Cong Dinh: orcid.org/0000-0002-3306-0603

ORCID for Alain Zemkoho: orcid.org/0000-0003-1265-4178

ORCID for Tri-Dung Nguyen: orcid.org/0000-0002-4158-9099

Catalogue record

Export record