Grainger College of Engineering, All Events

Abstract: Bandit learning has been gaining increasing attention and has enabled wide applications in e-commerce and recommender systems. Here, a decision maker sequentially selects an action from multiple actions and receives the reward of the selected action, which is leveraged to improve the action in the next round, with the goal of minimizing the regret compared to always selecting the best action. Recently, the evolution of modern applications has created fresh challenges. In this talk, we present two perspectives on extending classical bandit learning to adapt to these emerging settings. First, we discuss multi-objective bandits, where rewards are multi-dimensional rather than scalar. We present new general formulations and algorithms for Pareto regret minimization, along with theoretical results that reveal a simple yet surprising connection between Pareto regret and classical regret. This theory in turn motivates new algorithms with improved performance guarantees. Second, modern organizations increasingly rely on interacting decision-makers, from distributed services to teams of agents. We discuss multi-agent bandits, where learning proceeds collaboratively under imperfect communication. We introduce a new framework for collaborative learning over random networks and show how network structure affects learning performance. Together, these perspectives represent our attempts to understand how to generalize bandit learning to broader applications.

Bio: Mengfan Xu is an Assistant Professor in Industrial Engineering and an Adjunct Assistant Professor in Computer Sciences at UMass Amherst. She joined UMass in 2024 after earning her Ph.D. from IEMS at Northwestern University. Her research focuses on online learning, stochastic processes, and simulation. Her work has appeared in venues such as NeurIPS, ICML, AISTATS, and ACM SIGMETRICS, including a NeurIPS Spotlight and an honorable mention at the RLC Workshop. She has also been recognized as an Outstanding Reviewer at NeurIPS and received the Meritorious Reviewer Award from the INFORMS Journal on Computing.