From what I've read, DeepMind's approach is to just feed in the raw pixel data and the score. No rules, or anything like that.