Comparing Teamwork Modeling in an Empirical Approach

Modeling teams is becoming an increasingly active topic for the research communities of multi-agent systems, cognitive modeling, and decision making. Many research efforts have tried to use software-agents to model teamwork. However, we found little work has been done comparing different methods with common experiments. The Agent-based Modeling and Behavior Representation (AMBR) project has compared different cognitive process modeling in a military simulation environment (Gluck & Pew, 2001), but it has only examined single agent behavior so far. In teamwork modeling, different architectures or models are often evaluated with different domains, which are not comparable. For example, STEAM (a Shell for Teamwork) used a helicopter combat scenario as a simulator (Tambe, 1997), while CAST (Collaborative Agents for Simulating Teamwork) used a Wumpus grid-world as its test-bed (Yen, Yin, Ioerger, Miller, Xu, & Volz, 2001). Consequently, it is difficult to evaluate teamwork performance between two architectures. As a first attempt to address the issue, we conducted an empirical study to compare team performance. First, we built agent teams with two different architectures: CAST and Soar. Then, we tested them within the dTank simulation (Councill, Morgan, & Ritter, 2004). dTank is a tank game simulator for agents in a distributed environment. In the experiments, we compared the agents’ knowledge and behavior.