Generating Teammates for Training Robust Ad Hoc Teamwork Agents via Best-Response Diversity