AI Seminar

Selective Communication to Reduce Uncertainty in Multiagent Coordination

Edmund DurfeeProfessorUniversity of Michigan

To work together effectively, agents often must coordinate their
activities. Coordination can degrade, however, when agents grow uncertain
about how others perceive the current collective situation, and about what
plans and goals others are pursuing. Yet, maintaining continuous awareness
about all other agents' beliefs, plans, and goals is prohibitively costly,
if even possible at all. In this talk, I discuss strategies that agents can
use to decide selectively when to communicate, and about what, so as to
reduce their more important uncertainties at more critical times. I present
work, performed jointly with student and faculty colleagues, that spans a
variety of strategies for different contexts, ranging from where an agent
might be able to estimate ahead of time the expected value of what can be
learned through communication, to where an agent needs to engage in some
communication to discover what, if any, further communication is warranted.

Sponsored by