Mastering the game of Go without human knowledge

ABSTRACT

A long-standing goal of artificial intelligence is an algorithm that learns, tabula rasa, superhuman proficiency in challenging domains. We introduce AlphaGo Zero, an AI that achieves superhuman performance in the game of Go without any human data, guidance, or domain knowledge beyond the game rules. AlphaGo Zero uses a single neural network to evaluate positions and select moves, trained exclusively through a novel reinforcement learning algorithm from games of self-play. Starting from random play, AlphaGo Zero surpassed the performance of the version of AlphaGo that defeated the human world champion, Lee Sedol, in just 40 days of training.