AlphaGo Zero
A Go-playing AI developed by DeepMind that is trained entirely through self-play reinforcement learning, requiring no human game records, and achieves superhuman performance.
Core Facts
Timeline (last 90 days)
The success of AlphaGo Zero and AlphaStar has made the self-play paradigm highly praised, but successful cases often rely on enormous computing resources and carefully designed training.
Self-play (e.g., AlphaGo Zero) relies on adversarial gameplay between two sides to generate training signals, requiring the task itself to support an adversarial structure.
In each iteration of AlphaGo Zero, the new version plays against the old version, and the old version is replaced if the new version achieves a win rate exceeding 55%.
The success of AlphaGo Zero relies on Go being a closed system with complete rules and verifiable outcomes, which is one of the main obstacles to extending self-play to general-purpose AI.
In each iteration of AlphaGo Zero, the new version plays against the old version and replaces it if the win rate exceeds 55%; after approximately 49 million games of self-play over just 40 days of training, it surpassed top human players.
AlphaGo Zero was released by DeepMind in 2017; it uses Monte Carlo tree search combined with deep neural networks and was trained entirely from scratch without relying on any human game records.
AlphaGo Zero uses no human game records at all and surpassed all previous versions purely through self-play from scratch.
AlphaGo Lee先用人类棋谱做监督预训练再用强化学习自我对弈提升，而AlphaGo Zero完全抛弃人类数据纯粹从自我对弈学习并达到更高水平
All Facts (10)
AlphaGo Zero, released in 2017, relies entirely on self-play reinforcement learning without any human game data, and surpassed all previous versions within 40 days.
80%VerifiedAlphaGo Zero surpassed all human players entirely through self-play.
70%UnverifiedThe success of AlphaGo Zero and AlphaStar has made the self-play paradigm highly praised, but successful cases often rely on enormous computing resources and carefully designed training.
50%UnverifiedSelf-play (e.g., AlphaGo Zero) relies on adversarial gameplay between two sides to generate training signals, requiring the task itself to support an adversarial structure.
50%UnverifiedIn each iteration of AlphaGo Zero, the new version plays against the old version, and the old version is replaced if the new version achieves a win rate exceeding 55%.
50%UnverifiedThe success of AlphaGo Zero relies on Go being a closed system with complete rules and verifiable outcomes, which is one of the main obstacles to extending self-play to general-purpose AI.
50%UnverifiedIn each iteration of AlphaGo Zero, the new version plays against the old version and replaces it if the win rate exceeds 55%; after approximately 49 million games of self-play over just 40 days of training, it surpassed top human players.
50%UnverifiedAlphaGo Zero was released by DeepMind in 2017; it uses Monte Carlo tree search combined with deep neural networks and was trained entirely from scratch without relying on any human game records.
50%UnverifiedAlphaGo Zero uses no human game records at all and surpassed all previous versions purely through self-play from scratch.
50%UnverifiedAlphaGo Lee先用人类棋谱做监督预训练再用强化学习自我对弈提升，而AlphaGo Zero完全抛弃人类数据纯粹从自我对弈学习并达到更高水平
50%