Online Learning & Decision-Making

In online decision-making systems, the main objective is to sequentially maximize cumulative reward under uncertainty. In practice, however, feedback is often imperfect, exhibiting heavy-tailed noise, corruption, or even adversarial perturbations. My research studies how to design statistically principled algorithms that achieve robust and near-optimal regret guarantees in such challenging environments. My work focuses on understanding how uncertainty, variability, and contamination in rewards affect learning dynamics, and how variance-aware or robust methods can mitigate these effects without sacrificing efficiency.

• Variance-Aware Decision Making with Linear Function Approximation under Heavy-Tailed Rewards
  X. Li and Q. Sun. Transactions on Machine Learning Research, 2024.
• Corruption-Robust Variance-Aware Algorithms for Generalized Linear Bandits under Heavy-Tailed Rewards
  Q. Yu, E. Baek, X. Li, and Q. Sun. UAI, 2025.