[Colloq] PhD Seminar Tuesday, February 26, 4pm

[Rachel Bates]

Tuesday, February 26, 2002
149 CN
4:00-5:00
Talk will be followed by tea at 5:00.


Reinforcement Learning: An Overview

Professor Ron Williams
Abstract



Reinforcement learning is a subarea of machine learning that deals with
agents that interact with their environment and try to improve their
performance over time. This is a broad definition that encompasses many
possible learning tasks, but their common characteristic is that that the
agent selects short-term actions but its overall performance is measured by
a long-term reward function.

Included among such tasks are game-playing, process control, and robot
control. The are many dimensions along which such tasks may vary, including
how much prior knowledge of the environment is available and whether the
environment is stochastic or deterministic. Reinforcement learning
algorithms borrow from a variety of techniques from operations research,
artificial intelligence, and optimal control, to list a few. A cornerstone
of many of the approaches is the technique of dynamic programming, which
itself can be shown to be related to tree searching techniques from
artificial intelligence.

This talk will give an overview of the reinforcement learning approach and
how it relates to these other areas. It will also highlight some of the work
that various researchers, including me and some of my students, have focused
on. It will also describe some of the remaining challenges in this area.