Speaker: 刘鑫助理教授，上海科技大学

Inviter: 姚大成

Title: Safe and Efficient Learning and Decision-making in Dynamic Systems

Language: English

Time & Venue: 2026.09.02 14:00-16:00 S615

Abstract: Guaranteeing safety in real-world dynamic systems is challenging because we need to optimize the system utility while minimizing safety violation in the face of system uncertainty. This talk will present my recent work on developing algorithmic and theoretical foundations for safe and efficient learning and control in dynamic systems. This talk will start with a simplified “state-free” system, stochastic contextual bandits with safety constraints and introduce principled safe learning framework that can achieve no-regret performance with zero/minimal safety violation. Then I will present how to generalize these design principles from “state-free” systems into “stateful” Markovian systems and develop safe reinforcement learning with sample complexity guarantees. Applications to LLM routing and fine-tuning will be used throughout to illustrate these ideas.