All organisms, from bees to humans, and intelligent agents – such as
game playing programs and robots – are faced with the fundamental
problem of deciding a course of action that promises long-term
rewards. This problem of sequential decision making is made more
difficult due to sparsity of rewards, the curse of dimensionality, the
exploration-exploitation problem, and the inherent stochasticity of
the real world. This talk will review research in artificial
intelligence on a multi-scale probabilistic theory of sequential
decision-making that grapples with these challenges. Recent work on
proto-value functions, a new unified framework for learning
representation and behavior will be described, which exploits
mathematical properties of smooth functions on a manifold. Examples
from a variety of applications will be used to illustrate the issues,
including factory automation, humanoid and mobile robotics, and single
and multiagent plan recognition.
BRIEF BIO: Sridhar Mahadevan is an associate professor in computer
science at the University of Massachusetts, Amherst, where he
co-directs the Autonomous Learning Laboratory (with Andrew Barto). His
research work has spanned a variety of subfields of artificial
intelligence and machine learning. He is an associate editor of the
Journal of Machine Learning Research, and the Journal of Artificial
Intelligence Research. He was an invited speaker at the National
Conference on Artificial Intelligence in July 2005, celebrating the
25th anniversary of the American Association for Artificial
Intelligence.