EE: EAS Trailblazers Department Seminar
Abstract:
Machine learning-driven decision-making agents are drawing increased attention in recent years, yet their deployment is limited by efficiency challenges in complex environments and restricted adaptability to emerging applications. In this talk, I will present my recent research addressing these challenges across both online and offline settings. First, I will show how online decision-making can improve large language models (LLMs)—in particular, enhancing their mathematical reasoning through an online KL-regularized reinforcement learning (RL) approach that is backed by theoretical guarantees and empirical evidence. Next, I will introduce an offline decision-making framework that integrates heterogeneous datasets via a novel offline RL algorithm designed to handle multiple uncertainties. I will demonstrate this algorithm's theoretical and practical efficiency by applying it to wireless network optimization. Finally, I will discuss future directions and interdisciplinary opportunities for advancing intelligent decision-making agents, ultimately paving the way for broader real-world impact.
Biography:
Chengshuai Shi is a postdoctoral fellow at Princeton Language and Intelligence (PLI). Before that, he was a senior machine learning researcher at Bloomberg's Engineering-AI group. He received his Ph.D. in Electrical Engineering from the University of Virginia in 2024, following a B.E. from the University of Science and Technology of China in 2019. His research focuses on developing intelligent decision-making agents, with applications spanning wireless communications, recommender systems, and large language models. Chengshuai has published in premier venues such as NeurIPS, ICML, ICLR, AAAI, IEEE Transactions on Signal Processing, and IEEE Transactions on Wireless Communications. He has received several awards for his research excellence, including the ECE Department Louis T Rader Graduate Research Award in 2023, and the prestigious Bloomberg Data Science Ph.D. Fellowship in 2021.