         
|
GRASP Seminar Series: Spring 2005
February 4, 11:00 AM, Levine Hall 307.
Sridhar Mahadevan
Autonomous Agents Lab, CS Dpt., UMass, Amherst.
“Hierarchical Activity Modeling Using Dynamic Abstraction”
Abstract: Humans reason about the world at a variety of temporal and spatial scales, and make effective long-term decisions despite significant uncertainty in their state and the effects of their actions. How do we endow artificial systems, such as robots and software agents, with similar capabilities? In this talk, I will provide an overview of our laboratory's work in the broad area of hierarchical activity modeling using techniques from statistical machine learning. Combining ideas from graphical models and reinforcement learning, I will describe a family of hierarchical models for modeling concurrent single and multiagent activities. I will discuss the challenges in developing computationally tractable algorithms for learning and inference, and illustrate how tractability depends on the granularity of temporal and spatial hierarchies. A variety of real-world applications will be used to illustrate the research, including humanoid robotics, intelligent tutoring systems and multiagent vehicle coordination in factories.
Biography: Sridhar Mahadevan is an associate professor of computer science at the University of Massachusetts, Amherst, where he co-directs the Autonomous Learning Laboratory with Professor Andrew Barto. Professor Mahadevan's research interests span several subfields of artificial intelligence and computer science, including machine learning, multi-agent systems, planning, perception, and robotics. Over the past decade has centered around the framework of reinforcement learning and sequential decision-making, where his papers are among the most cited in the field. He was a pioneer in the development of average-reward and hierarchical reinforcement learning, the application of reinforcement learning to robotics, and the use of semi-Markov decision processes to model temporal abstraction in planning and learning. Professor Mahadevan serves as an associate editor for the Journal of AI Research and the Journal for Machine Learning Research. Previously, he served for many years as an Associate Editor for the Machine Learning journal. He has been on numerous program committees for AAAI, ICML, IJCAI, NIPS, ICRA, and IROS conferences, including area chair ( reinforcement learning) at the annual ICML and NIPS conferences. In 1993, he co-edited (with Jonathan Connell) the book Robot Learning published by Kluwer Academic Press, widely considered to be a pioneering book on the application of statistical machine learning to robotics. He is a recipient of the NSF CAREER award, and his research has led to best paper awards at the AAMAS (2001) and ICML (1999) conferences.
