Wikinventia — Atlas of discoveries and inventions · Digital Age

Self-play reinforcement learning generalized to chess, shogi and Go — AlphaZero, DeepMind

2017 AD · Transmission: Global
AIMethodBritish

In December 2017, DeepMind published AlphaZero, a system taught only the basic rules of chess, shogi and Go, which reaches superhuman level in all three games in under 24 hours of training via self-play reinforcement learning, without any database of human games or hand-designed heuristics. In its paper, DeepMind directly pits AlphaZero against Stockfish (chess) and other specialized engines, decisively defeating them. Unlike Deep Blue or Stockfish — heirs to the alpha-beta search and hand-crafted evaluation paradigm Shannon formulated in 1950 — AlphaZero entirely replaces that approach with deep neural networks guiding a Monte Carlo tree search, symbolically closing the cycle begun by Torres Quevedo's Ajedrecista: from a fixed decision tree hand-built by an engineer to a system that discovers its own strategies with no prior human knowledge.

InstitutionDeepMind — London, United Kingdom
Historical regionUnited Kingdom
Primary sourceSilver, D. et al. — "Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm" (arXiv:1712.01815, 2017)
Secondary sourcechessprogramming.org — "AlphaZero"
Original languageEnglish
View this entry in the interactive atlas → View in graph →