AlphaGo Zero: Starting from scratch
deepmind.com
It is able to do this by using a novel form of reinforcement learning, in which AlphaGo Zero becomes its own teacher. The system starts off with a neural network that knows nothing about the game of Go.