Swap-regret-minimizing bandits for distributed network optimization

Date

2025

Authors

Huang, Zhiming

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Modern networked systems—ranging from real-time communication platforms to distributed computing infrastructures—operate in increasingly dynamic and strategic environments, where traditional optimization methods often fall short. This dissertation develops a new algorithmic framework for distributed network optimization grounded in game-theoretic bandit learning. We model fundamental problems, such as congestion control and resource allocation, as repeated games involving strategic agents who receive only partial (bandit) feedback. Motivated by practical challenges in computer networks, we design and analyze algorithms that not only minimize regret but also steer collective behavior toward equilibrium. The contributions of this dissertation are threefold. First, we propose a new framework based on swap-regret minimization and online mirror descent, and establish high-probability regret bounds in multi-player bandit settings. These results guarantee convergence to correlated equilibria under decentralized, partial-information feedback. Second, we introduce optimistic learning techniques to accelerate convergence by leveraging predictability in the environment. Third, we apply our algorithms to real-world networking tasks, including TCP congestion control, and demonstrate improved stability, throughput, and fairness through extensive trace-driven emulations. Together, these contributions bridge the theoretical foundations of online learning and game theory with practical considerations in network protocol design, offering robust tools for decentralized decision-making in uncertain and adversarial environments.

Description

Keywords

Multi-Armed Bandits, Online Learning, Game Theory, Correlated Equilibrium, Network Optimization, Congestion Control

Citation