Without learning from expert human games, it beats an earlier AlphaGo 100–0

Did you know that an AI learned go without first copying human parts? In two thousand seventeen, AlphaGo Zero trains by playing against himself. He knows the rules, then improves the way he chooses moves from his gaming experiences. The principle is to produce parts and use their results to adjust the system. It’s not two robots sitting face-to-face or conscious software. Our animated board explains the alternation of the players, without pretending to reproduce a real training game.

After three days of training in the published experiment, AlphaGo Zero beats a previous version of AlphaGo by one hundred parts to zero. This opponent is another system, not a hundred human champions. The score therefore compares two versions within the specific framework presented by the researchers.

This result shows that learning can benefit from experiments made by the system itself. But the rules of Go provide a clear and testable world. We cannot turn this success into proof that an AI discovers all knowledge without data, rules or verification.