Rac : Multi-Agent Actor-Critic Framework for Learning Emergent Roles In Team-Competitive Games

Team-competition in multi-agent games is an important setting, where multi-agent actor-critic methods are the state-of-the-art methods. In many such scenarios, the roles for players naturally emerge, thus aiding in collaboration within the teams. However, exiting methods for the team-competition scenario do not attempt to learn such emergent roles. In this paper, we propose RAC, a new technique for learning emergent roles which are dynamic and diverse. RAC uses the actor-critic framework with role encoder and opponent role predictors, that are used in learning an optimal policy. The policies learned by RAC outperform state-of-the-art baselines in direct competitions. Extensive experimentation confirms the quality of learned roles, and utility of the different loss functions used