Playing is believing: The role of beliefs in multi-agent learning