[Colloquium] Bagnell talk today 12:15 at TTI
Meridel Trimble
mtrimble at tti-c.org
Thu Feb 19 09:04:58 CST 2004
TOYOTA TECHNOLOGICAL INSTITUTE TALK
Speaker: Drew Bagnell
Carnegie Mellon
Speaker's homepage: www.cs.cmu.edu/~dbagnell
Time: 12:15 p.m.
Date: Thursday, February 19, 2004
Place: TTI-C (1427 E. 60th Street- Press Building 2nd Floor East)
LUNCH PROVIDED
Title: "Learning Policies: Leveraging Supervised Learning for Planning
and Control"
Abstract: Increasingly, learning systems are called upon to not merely predict,
but to take actions that affect the world. The fields of planning and control
are meanwhile asked to make good sequential decisions in complex problems where
simple analytic models are unavailable. This confluence of learning and control
has driven research to identify algorithms that can search for good strategies
while remaining efficient in both sample and computational resources.
Policy search approaches are becoming a popular alternative to classical value
function methods in part because they gracefully handle both large state-spaces
and partial observability. I'll discuss two recent developments that transplant
techniques from supervised learning to learning policies. First, I'll show how
notions from information geometry give simple algorithms with dramatically
better performance than existing reinforcement learning algorithms. Next, I'll
discuss a new approach that discards Bellman equations and searches directly in
a space of policies while retaining the central insight of dynamic programming.
With appropriate domain knowledge, this algorithm is efficient in both
computational and sample complexity and provides strong performance guarantees.
Examples will illustrate how each heoretical advance leads to state-of-the-art
empirical performance on difficult benchmark problems. J.A. (Drew) Bagnell is a
doctoral candidate in Robotics and an NSF graduate fellow at Carnegie Mellon
University. His interests include rich probabilistic models for statistical
learning, robust and stochastic control, and the application of machine
learning
to large, partially observable planning and control problems.
If you have questions, or would like to meet the speaker, please contact
Lori Wilkus at 773-834-2571 or lwilkus at tti-c.org
For more information on future TTI-C talks or events, please go to the
TTI-C Events page: http://www.tti-c.org/events.shtml
Lori Wilkus
Staff Accountant
Toyota Technological Institute at Chicago
Phone: 773-834-2571
More information about the Colloquium
mailing list