Creating autonomous agents capable of cooperating with previously unfamiliar teammates, known as ‘ad hoc teamwork’, has been identified as an important challenge for multiagent systems. Previous research has assumed that either the task, the role of each agent, or the communication protocol among agents is known before the interaction begins. We consider these three variables simultaneously and show how an ad hoc agent can fit into a new team while handling ambiguous tasks, roles, and communication protocols. We assume a known distribution of possible tasks, roles, and communication protocols. We present experimental results in the pursuit domain showing that our ad hoc agent can join such a team without barely impacting the overall performance compared to a pre-coordinated agent.
In collaboration with Samuel Barrett and Peter Stone @ UT Austin.
Paper with Samuel Barrett, Peter Stone, Manuel Lopes.