Learning policy representations in multiagent systems