MIT Department of Electrical Engineering & Computer Science
Planning and Learning in Uncertain Environments
Leslie Pack Kaelbling
Brown University
Thursday, May 15, 1997
4:00 PM (3:45 refreshments)
NE43-8th Floor, AI Playroom
EECS Special Seminar
Abstract
Agents are always embedded in environments about which they are
uncertain. This is true of humans, animals, robots, factory
controllers, and web agents. Such agents must act so as to gather
information about their environments as well as to achieve desired
effects in the environment. In this talk, I will discuss partially
observable Markov decision processes (POMDPs) as a model of acting in
uncertain environments. I'll briefly describe the theory, show that
it's intractable to solve these problems exactly, then discuss some
approximation methods. I will present applications of these methods,
including real robot navigation and target identification.
The POMDP model is appropriate when much is known about the world and
only some aspects must be learned. When prior knowledge is weaker,
reinforcement learning techniques seem more appropriate.
Unfortunately, on-line learning in autonomous agents has not generally
been successful. I will conclude with a discussion of some of the
problems involved in autonomous learning and speculation about how
they might be addressed.
Hosts: Prof. R. Brooks and Prof. T. Lozano-Perez
URL of this page:
http://www-eecs.mit.edu/AY96-97/events/53.html
Created: May 8, 1997
|
Modified: Jun 24, 1997
This announcement is from the MIT EECS 1996-97 archive.
|
Current events
To MIT EECS home page
|
Your comments
and inquiries are welcome.
