If multi-agent learning is the answer, what is the question?