The same rules apply when two computer players play Go. The photo shows a physical board. Image source. Licence.
The earlier AlphaGo studied moves people had played. In 2017, DeepMind described a successor that started without those game records and generated its own experience by playing both sides of Go games. After three days of training, the paper reports, AlphaGo Zero beat AlphaGo Lee 100–0 in their comparison.
The system still started with the rules of Go and a search procedure designed by researchers. It played games, used the results to improve its estimates of promising moves and likely winners, then played again with the updated network. Search helped make the next round of examples better.
The game supplied a clear answer
Self-play works especially cleanly in a game like Go because the program can generate as many legal games as it has time to play and can determine who won each one. The researchers did not have to ask a person to judge whether a move looked good.
Those conditions are part of the result. The system's achievement was learning to play Go at an extraordinary level from experience it generated itself, in a setting where the rules and the outcome were already known.