Adaptive Team Coaching Using Opponent Model Selection

ABSTRACT In multiagent domains with adversarial and ooperative team agents, team agents should be adaptive to the urrent environment and opponent. We introdu e an online method to provide the agents with team plans that a \ oa h" agent generates in response to the spe i opponents. The oa h agent an observe the agents behaviors but it has only periodi ommuni ation with the rest of the team. The oa h uses a Simple Temporal Network to represent team plans as oordinated movements among the multiple agents and it sear hes for an opponent-dependent plan for its teammates. This plan is then ommuni ated to the agents, who exe ute the plan in a distributed fashion, using information from the plan to maintain onsisten y among the team members. In order for these plans to be e e tive and adaptive, models of opponent movement are used in the planning. The oa h is then able to qui kly sele t between di erent models online by using a Bayesian style update on a probability distribution over the models. Planning then uses the model whi h is found to be the most likely. The system is fully implemented in a simulated roboti so er environment. In several re ent games with ompletely unknown adversarial teams, the approa h demonstrated a visible adaptation to the di erent teams.