Learning Policy Representations in Multiagent Systems