AlphaZero is a Reinforcement learning system developed by DeepMind that learned to play chess, shogi, and Go at superhuman levels using a single general-purpose algorithm and no human-generated game data. DeepMind announced AlphaZero in a paper published in December 2017, describing a system that, starting from only the rules of each game, learned entirely through self-play, using a deep neural network combined with Monte Carlo tree search to evaluate positions and select moves, then improving that network through the outcomes of games played against itself.
Relationship to AlphaGo
AlphaZero generalized the self-play approach first demonstrated by AlphaGo Zero, the version of AlphaGo that had learned Go purely through self-play without human game data, to additional games. Where AlphaGo and AlphaGo Zero were specialized for Go alone, AlphaZero used the same underlying architecture and training procedure, with no game-specific tuning beyond the rules themselves, to reach superhuman performance in chess and shogi as well, demonstrating that the self-play reinforcement learning approach was not a Go-specific trick but a genuinely general method for mastering games of perfect information.
Chess result
AlphaZero's chess result drew particular attention within both the AI and chess communities because chess had a much longer history of computer engine development than Go, dominated for two decades by traditional engines using deep, hand-tuned brute-force search, most famously Stockfish. After just four hours of self-play training, using no opening books, endgame tablebases, or other chess-specific human knowledge beyond the basic rules, AlphaZero defeated Stockfish in a 100-game match. Commentators, including grandmasters who reviewed the games, noted that AlphaZero's chess play displayed a distinctive style, favoring long-term piece mobility and king safety over the material-counting heuristics embedded in traditional engines, and frequently made sacrifices that produced durable positional advantages rather than immediate material gain, a style some described as more intuitive or aesthetically striking than that of prior engines.
Significance
AlphaZero's demonstration that a single self-play reinforcement learning algorithm could master multiple distinct games of perfect information without domain-specific engineering was seen as an important step toward more general-purpose learning systems, though AlphaZero remained limited to fully observable, deterministic, two-player games and did not extend to the kind of open-ended, real-world tasks associated with Artificial general intelligence. Its architecture and self-play training methodology influenced subsequent DeepMind research, including MuZero, a follow-up system that learned to play games, including ones with unknown rules, by learning an internal World model of the environment rather than being given the rules directly. AlphaZero's chess games were subsequently studied extensively by professional chess players and engine developers, and some of its stylistic tendencies influenced the design and evaluation of later chess engines, including open-source neural-network-based engines that adopted similar self-play training methods.
Reception
AlphaZero is widely cited in AI research as a landmark demonstration of the power of self-play reinforcement learning combined with deep neural networks and search, extending the significance of AlphaGo's earlier results, including its 2016 win over Lee Sedol, and reinforcing DeepMind's broader research narrative that general learning algorithms, given sufficient compute, could match or exceed decades of accumulated human and hand-engineered domain knowledge in well-defined domains.