Bootstrapped by supervised learning though: it was trained with the results of 30 million human games before the reinforcement learning kicked in.