AlphaGo Zero teaches itself Go from scratch
On this day in 2017, DeepMind introduced AlphaGo Zero. Earlier versions of AlphaGo learned partly by studying thousands of human games. AlphaGo Zero was given only the rules and started from random play, improving by playing millions of games against itself with reinforcement learning, so in effect it became its own teacher. After just three days of self play, it beat the version of AlphaGo that had defeated world champion Lee Sedol by 100 games to 0. Along the way it discovered unconventional strategies of its own. The work, published in Nature, showed that an AI could reach superhuman skill without human examples, and its creators hoped similar methods could help with problems like protein folding and new materials.