Multiple roles, multiple teams, dynamic environment: Autonomous netrek agents

We describe the architecture and performance of au-tonomous agents that play the complex, multi-faceted internet game called Netrek. To perform competently within Netrek, an agent must be capable of 1) re-acting in real-time to arcade-like tasks and environ-ment changes within the local surroundings, 2) rea-soning about strategy and long-term tasks at a slower pace and on a global scale, and 3) coordinating both with teammates and against members of the oppos-ing team. Agents capable of operating within the Ne-trek environment, therefore, require a reactive, exi-ble, multi-level framework. We describe in detail our current agent implementation, which is based upon the UMPRS (University of Michigan Procedural Rea-soning System) architecture. As multi-agent coordina-tion plays a signicant role within the Netrek domain, we describe recent research in extending the current communication-based intra-team coordination scheme with a plan recognition scheme that enhances both intra- and inter-team coordination