AI Seminar
    -------------------------------
    Tuesday, October 28th, 2003
    4:00 pm - 5:30 pm
    175 ATL (Large Conference Room)

    "Cooperative Real-Time Control Plans for Resource-Limited Agents"

    Ed Durfee
    Department of Electrical Engineering and Computer Science
    University of Michigan




----------------------------------
A resource-limited agent operating in real-time environments needs to make
trade-offs. Because it generally cannot monitor and respond to every
time-critical event quickly enough, it needs to decide which reactions to
ensure will occur in real time, and which it can best afford not to
guarantee.  Such trade-off decisions can be dramatically improved if the
agent can communicate with other agents with whom it shares its environment,
to model their reactions better, and potentially to influence them to change
their reactions.  At the extreme, the agents can in fact formulate a joint
policy that they can execute decentrally.  In this talk, I will describe the
protocols and algorithms that we have developed through which agents that
initially have formed stand-alone real-time control plans can negotiate and
converge on more effective, complementary control plans.  I will also
outline our ongoing work in extending techniques for constrained Markov
Decision Processes to decentralized problems requiring hard guarantees on
resource utilizations. Applications for these techniques, such as unmanned
aircraft control and planetary robotics, will be used for illustration.

This work has been done jointly with Haksun Li, Dmitri Dolgov, and Kang
Shin.