Abstract:

The rapidly expanding field of Multi-Agent Systems seeks to study the
behavior of groups of agents: autonomous, situated, persistent
entities that interact with each other. Researchers have long since
observed that agents must be able to maintain knowledge about other
agents, in order to execute individual and group plans, ascertain
progress, and detect failures. This ability is called Multi-Agent
Modeling, and is of utmost importance in such diverse application
areas as network security, virtual worlds and training simulations, 
intelligent user-interfaces, distributed applications, collaboration 
support, biology and social-sciences research.

In practice, however, multi-agent modeling poses a challenge. Due to
bandwidth and computational limitations, agents cannot continuously
monitor all other agents, and thus face uncertainty when reasoning
about others. In this talk, I will present some in-depth empirical and
analytical responses to this challenge, based on my experience with
building various multi-agent modeling systems in several complex,
dynamic multi-agent domains.  I will focus on two multi-agent modeling
tasks: Detecting failures in teamwork, and visualizing the state of a
team.  The results I present (i) guarantee failure detection despite
uncertainty, (ii) explore techniques for alleviating uncertainty in
visualizing teams, and (iii) address scalability of modeling large
teams with many agents.