The encyclopedia · R&D & Science · Technical decision · 2016–2017
AlphaGo Zero learned to play Go superhumanly from zero human games
DeepMind's AlphaGo Zero beat the champion-defeating AlphaGo 100-0 by learning only from self-play, with no human data or domain knowledge.
Google DeepMind
the move
AlphaGo had already beaten a world champion in Go, but it leaned on a human game database and hand-crafted domain features. That borrowed knowledge put a ceiling on how strong the program could get, and gathering better human data was slow and expensive.
The AlphaGo Zero team removed the human source entirely. The network starts knowing nothing beyond the board rules; it plays against itself, and the network learns to predict which move it would make and who would win. That prediction improves the search, and stronger search makes stronger self-play.
Nothing outside the game is allowed in. The system is its own curriculum and its own grader, compounding strength iteration after iteration without an expert move ever being shown to it.
Starting from that blank slate it beat the original, champion-defeating AlphaGo 100 games to zero — a result that showed the value of a clean, self-generated training loop over borrowed data.
why it works
- Self-play produces an unlimited supply of games matched exactly to the agent's current strength, so the curriculum never gets too easy or too hard.
- Removing human games removes the noise and the ceiling of human play, letting the agent pass the level of the people who wrote the training data.
- Each improvement in the network makes the search better, and better search generates stronger opponents, so gains compound in a positive loop.
- Because the reward is just winning the game, the signal is clean and automatic and needs no labeling.
what transfers
The best training signal is often the one you can generate at will — if you can define the rules, self-play gives you infinite, on-distribution data for free.
what came after
The same self-play idea generalized across games and is now a building block of modern reinforcement learning, opening the door to agents that beat humans in domains with no good expert data. DeepMind went on to apply tabula rasa learning to chess, shogi and other games, and the 'learn from your own rollouts' loop became standard in the field.
references
spotted an error? The archive wants to know.