BISC Seminar, 6 February 1997, 310 Soda Hall, 4-5:00pm

Michael Lee (
Fri, 7 Feb 1997 12:03:39 +0100

The Basic Ideas Underlying Reinforcement Learning

BISC Seminar

Ron Parr
CS Division
University of California
Berkeley, CA 94720

6 February 1997
310 Soda Hall


Reinforcement learning is a powerful and general method that can be used to produce adaptive, intelligent behavior
on-line. A system using reinforcement learning will maintain estimates of the utilities of various world states. Through
experience, these estimates will be updated and will converge towards the true utility values of these states, providing the
system with the means of selecting an optimal control decision. In this talk I will give an overview of the basic concepts
in reinforcement learning, its relationship to dynamic programming, and how it solves the temporal credit assignment
problem. I will present some examples of successful implementations of reinforcement learning, and discuss some of the
difficulties in applying this approach to larger and more diverse problems. Finally, I will present a brief overview of my
recent work on the use of hierarchical control strategies to improve the speed and versatility of reinforcement learning

Michael A. Lee
Post Doctoral Researcher
Berkeley Initiative in Soft Computing
387 Soda Hall                                      Tel: +1-510-642-9827
Computer Science Division                          Fax: +1-510-642-5775
University of California                    Email:
Berkeley, CA 94720-1776 USA       WWW: