Learning Sequential Tasks from Human Feedback
Virtual agents and physical robots need to be able to learn so that they can perform novel or unanticipated tasks. Reinforcement learning is a powerful framework that allows agents to learn to maximize an environmental reward, but there are many cases where no environmental reward signal is present. However, we know that people are able to train via evaluative feedback: consider all of the impressive tasks that dogs can learn to accomplish! This talk will discuss learning algorithms that rely on non-technical user feedback to train agents to perform sequential decision tasks in a variety of settings.
Matt is currently Research Director of the Edmonton Borealis AI research center and holds adjunct appointments at Washington State University and the University of Alberta. He received his doctorate from the University of Texas at Austin in 2008, supervised by Peter Stone. Matt then completed a two year postdoctoral research position at the University of Southern California with Milind Tambe. Since then, he has had positions at Lafayette College and Washington State University, where he held the Allred Distinguished Professorship in Artificial Intelligence. Current research interests include intelligent agents, human-agent interaction, multi-agent systems, reinforcement learning, and robotics.