Relational Forward Models for Multi-Agent Learning