Refreshments: 10:30 AM

Seminar: 11:00 AM

Title: When the Expert Is Not Optimal: Leveraging Human Feedback in Deep Reinforcement Learning

Abstract: In this talk, I will introduce the basic idea behind Reinforcement Learning (RL): how agents learn through trial and error by interacting with their environment. While this paradigm has led to remarkable success, it often requires a massive amount of agent experience, especially for complex domains having high-dimensional states and actions. This limitation motivates the use of human-in-the-loop approaches in Deep Reinforcement Learning (DRL), where human guidance can significantly improve learning efficiency. I will then discuss how learning can be further accelerated by incorporating knowledge from multiple teachers who are not necessarily experts. Rather than assuming perfect guidance, we consider settings where human feedback is imperfect and of varying quality. Towards the end, I will introduce a specific human-in-the-loop RL paradigm known as preference learning, in which RL agents learn from comparative signals instead of explicit reward signals. I will highlight how noisy and inconsistent human rankings pose significant challenges for current preference-based RL approaches, and discuss the different forms of feature-dependent noise that humans can introduce in such frameworks, which can degrade agent performance.

Bio: Dr. Srijita Das is currently an Assistant Professor in the Department of Computer and Information Science at the University of Michigan-Dearborn. She was previously a Postdoctoral Fellow in the Intelligent Robot Learning Lab (IRL Lab) at the University of Alberta, Canada. She earned her Ph.D. in Computer Science from the University of Texas at Dallas, USA. Her research interests include Reinforcement Learning, Human-in-the-loop learning, Human-Machine Collaboration, Active-Learning and Cost-sensitive Learning. She has published several research works in refereed journals and conferences in the areas of machine learning, including IJCAI, AAMAS, AAAI, JAIR & TMLR. She is also a reviewer for top AI and ML Conferences like AAAI, NeurIPS, ICML and AAMAS