The paper extends a deep learning model known as graph neural networks and uses it to learn generalised heuristics. We show that these heuristics generalise to problems with different goals, larger problems, and even problems from different domains than those we trained on. This is the first paper that successfully learns domain-independent heuristics.