Department of Mathematics - Seminar on Statistics and Data Science - Functional Critics Are Essential for Actor-Critic: From Off-Policy Stability to Efficient Exploration
Supporting the below United Nations Sustainable Development Goals:支持以下聯合國可持續發展目標：支持以下联合国可持续发展目标：
This work revisits several central challenges in model-free reinforcement learning: how to achieve stable, sample-efficient learning and efficient exploration for actor-critic methods under the Deadly Triad setting. The talk will introduce new theoretical results that help bridge longstanding gaps between off-policy actor-critic theory and practice, as well as a foundational connection between functional critics and posterior-sampling-style exploration in reinforcement learning.
Dr. Bai is a founding researcher and head of fundamental research at Isara Labs, a San Francisco-based Al neo lab backed by OpenAl. Isara recently raised US$94 million at a reported US$650million valuation, underscoring its role in a new generation of Al laboratories focused on movingbeyond single-model intelligence toward large-scale, coordinated systems of Al agents. Its missionis to build the scientific and engineering foundations for agentic AI systems that can collaborate,reason, verify, and conduct complex analytical and research workflows at scale - one of the mostpromising directions for scaling AI systems beyond the capabilities of individual large language models.