Practical Real-World Reinforcement Learning via Entropy Regularization

ABSTRACT

Reinforcement learning has been among the most popular topics for the past decade due to its superhuman level performance in e.g. game playing or more recently empowering LLMs. However, most of these applications are limited to the virtual world. When we implement RL to real-world applications such as industrial processes, several unique challenges arise. In this talk, I will identify three unique challenges as scalability, safety and robustness. My approach to these problems is known as the entropy regularization which refers to a family of algorithms based on information-theoretic considerations, with well-known examples including PPO, SAC and GRPO.

SPEAKER BIO

Dr. Lingwei Zhu is an incoming assistant professor at the Great Bay University in Dongguan, Guangdong. His primary research interest lies in entropy-regularized reinforcement learning and its applications to autonomous control of large-scale systems as well as sciences. He graduated with the best student honor from Nara Institute of Science and Technology (NAIST) in Japan. He spent two years at the RLAI lab of the University of Alberta and later one year at the University of Tokyo as a postdoctoral researcher.