AlphaZero is a computer program developed by the artificial intelligence research company DeepMind to master the board games chess, shogi, and Go. Unlike earlier game-playing systems, it learned entirely through self-play, starting with no human knowledge beyond the game rules. The program uses a neural network combined with Monte Carlo tree search, an approach first demonstrated in its predecessor AlphaGo Zero. The company released a preprint paper introducing AlphaZero on December 5, 2017, followed by a peer-reviewed publication in the journal Science on December 7, 2018.
AlphaZero's training method required no opening books, endgame tables, or human-crafted heuristics. Instead, it generated its own experience by playing against itself millions of times, guided by a reinforcement learning algorithm. This approach allowed it to achieve superhuman performance in all three games within hours, despite searching far fewer positions per second than traditional engines. DeepMind described the resulting style of play as unusual, with some moves appearing counterintuitive to human observers.
Training Method
The training process used 5,000 first-generation tensor processing units (TPUs) to generate self-play games and 64 second-generation TPUs to update the neural network weights, all running in parallel. The in-training system was periodically matched against a benchmark program in short exhibition games to gauge progress. Against the chess engine Stockfish, AlphaZero exceeded benchmark performance after about four hours of training; against the shogi engine Elmo, it took roughly two hours; against the three-day version of AlphaGo Zero, about eight hours.
The neural network was updated continuously during training. Unlike its predecessor system, AlphaZero did not use game symmetries to augment its training data, and it incorporated the possibility of draws, which are absent in Go but possible in chess and shogi. The algorithm generalizes the earlier AlphaGo Zero approach to handle these additional draws.
Match Results
In chess, after nine hours of training, AlphaZero played a 100 game match against Stockfish 8 (the 2016 TCEC world champion), with each side receiving one minute per move. Stockfish used 64 threads and a hash size of 1 GB, while AlphaZero ran on a single machine with four TPUs. The result was 28 wins, 0 losses, and 72 draws. In a series of twelve 100-game matches starting from popular human openings, AlphaZero won 290, drew 886, and lost 24.
In shogi, after two hours of training, AlphaZero defeated the Elmo engine (World Computer Shogi Championship 27 summer 2017 version with YaneuraOu 4.73) in 100 games, winning 90, losing 8, and drawing 2. In Go, after 34 hours of self-learning, AlphaZero played against the three-day version of AlphaGo Zero, winning 60 games and losing 40.
The number of positions evaluated was dramatically different. AlphaZero searched roughly 80 positions per second in chess and 40,000 in shogi, compared to 70 million for Stockfish and 35 million for Elmo. It compensated with the selective focus of its deep neural network on promising variations.
Evaluation of Performance
DeepMind researchers stated that the algorithms demonstrated that a generic Reinforcement learning approach could reach superhuman levels in multiple board games without domain knowledge beyond the rules. They noted that state-of-the-art chess engines typically searched millions of positions using handcrafted expertise, while AlphaZero searching a thousand times fewer positions within hours. Demis Hassabis of DeepMind, himself a chess amateur, described AlphaZero's style as alien, referencing counterintuitive sacrifices like offering a queen and bishop for positional advantages.
However, some experts expressed caution. Grandmaster Hikaru Nakamura and Komodo developer Larry Kaufman suggested that the outcomes could have been more contested if Stockfish had used an opening database, since the engine was optimized for that scenario. Stockfish developer Dario Romstad noted that it is not optimized for rigidly fixed time, and the version used was already outdated by the match. Similarly, some shogi observers thought Elmo's hash size was too low, and the resignation or entry settings may have been inappropriate.
Legacy
DeepMind never released the full AlphaZero code. The algorithmic description in the Science paper, however, enabled the public to reproduce similar systems. In 2019, DeepMind published a more powerful generalization called MuZero, which achieved strong performance on Atari games and board games without being taught the rules or a representation of the game, showing that the core idea is not limited to human-defined game rules.
Observers compared AlphaZero's debut to historical milestones in Chess computer development. In 1997, Deep Blue became the first computer to beat world champion Garry Kasparov on a match. AlphaZero represented a different frontier: Instead of using specialized, search-heavy heuristics focused on millions of positions, it learned that its own code and out-competing by Machine learning, where neural networks trained via self-play surpassed the performance of human-engineered systems.