Description
Date and Time: September 12, 2025, 10:30 – 11:30 (JST)
Venue: Online
Title: Heavy-tailed Linear Bandits: Huber Regression with One-Pass Update
Speaker: Jing Wang (Nanjing University)
Abstract: Stochastic Linear Bandits (SLB) provide a fundamental framework for sequential decision making, with broad applications such as recommendation systems and close ties to the theoretical foundations of reinforcement learning. In this talk, I will present our recent progress on the Heavy-tailed Linear Bandit (HvtLB), which addresses the scenario with heavy-tailed noise, published at ICML 2025. While earlier work has established near-optimal and moment-aware regret bounds for HvtLB, existing algorithms require storing all past data and scanning it at each round, which is impractical in online scenarios. To overcome this challenge, we propose a one-pass Huber regression algorithm based on online mirror descent. Our method processes only the current data at each step, reducing the computational cost to a constant level while still achieving near-optimal, variance-aware regret guarantees. Our approach further enjoys the potential to be applied to broader decision-making scenarios involving heavy-tailed noise, such as online linear MDP and online adaptive control.
Bio: Jing Wang is a Ph.D. student in the LAMDA Group from Nanjing University, supervised by Prof. Zhi-Hua Zhou and Assistant Prof. Peng Zhao. His research focuses on online learning, bandit algorithms, and resource-aware machine learning. His work has been published at top conferences, including ICML and AISTATS. He also serves as a reviewer for conferences such as ICML, NeurIPS, ICLR, and AISTATS, as well as for the ML journal.
