AlphaZero

AlphaZero is DeepMind's neural-network chess engine that learned superhuman play through self-play reinforcement learning and defeated Stockfish in 2017.

Category
Historical
Also in
Computer Chess
Difficulty
Intermediate
Also known as
Alpha Zero, AZ
First recorded use
2017

AlphaZero is a chess engine developed by DeepMind that reached superhuman strength by teaching itself through self-play rather than studying human games or using hand-crafted evaluation functions.

It combines a neural network with Monte Carlo Tree Search. The network learns two things at once: which moves are promising in any position and how likely each side is to win from that position. During training, AlphaZero plays millions of games against itself, constantly updating the network based on the outcomes. No opening book or endgame tablebases are supplied; everything emerges from raw self-play.

In December 2017 AlphaZero defeated Stockfish in a 100-game match while using far less search time per move. Its style stood out immediately. It often accepted structural weaknesses in exchange for dynamic piece activity and willingly sacrificed material for long-term pressure. Club players who study its games notice how frequently it keeps the position tense instead of simplifying into technically won endgames.

The main practical takeaway is that a well-trained neural network can evaluate positions more accurately than traditional engines that rely on thousands of manually tuned terms. This allows the program to find moves that older engines undervalue or overlook entirely.

Key takeaways

  • AlphaZero learned superhuman chess solely through self-play reinforcement learning.
  • It combined a neural network with Monte Carlo Tree Search to assess moves and winning chances.
  • AlphaZero defeated Stockfish in 2017 while using far less computation per move.
  • Its games show a preference for dynamic activity and tension over material or simplification.
  • Neural-network evaluation proved more powerful than traditional hand-tuned functions.

Frequently asked questions

What is AlphaZero?
AlphaZero is DeepMind's neural-network chess engine that learned superhuman play through self-play reinforcement learning without human games or hand-crafted evaluations.
How did AlphaZero defeat Stockfish?
In a 2017 match AlphaZero beat Stockfish while using far less search time per move, thanks to its neural network and Monte Carlo Tree Search combination.
What is distinctive about AlphaZero's playing style?
It often accepts structural weaknesses for dynamic piece activity and keeps positions tense rather than simplifying into technical endings.
Why does AlphaZero matter for club players?
Its games demonstrate how neural-network evaluation can outperform traditional hand-tuned functions and highlight the value of maintaining tension over early simplification.
Does AlphaZero use opening books or endgame tablebases?
No, it learns all phases of the game solely from millions of self-play games without any external databases.

See also

  • Brilliancy — A brilliancy is a striking, often sacrificial move or sequence whose originality, depth, and elegance transform the position in an unexpected yet forcing way to produce a decisive advantage.
  • Brilliancy Prize — A brilliancy prize is an award given in some chess tournaments for the most spectacular, creative, or original game or combination played.
  • Compensation — Compensation is a positional or dynamic advantage such as initiative, better pawn structure, or attacking chances that offsets a material deficit.
  • Decoy Sacrifice — A decoy sacrifice offers material to lure an enemy piece onto a specific square where it becomes vulnerable to attack or loses its defensive role.
  • Discovered Attack — A discovered attack occurs when one piece moves off a line and uncovers an attack from a second piece standing behind it on the same rank, file or diagonal.
  • Discovery — A discovery is a tactic in which one piece moves off a line, uncovering an attack from a second piece stationed behind it.
  • MuZero — MuZero is DeepMind's reinforcement-learning system that masters chess via self-play by learning an internal model of game dynamics with no pre-coded rules or chess knowledge.

Sources

  1. Google DeepMind: AlphaZero and MuZero

Written by Chesswick.

Reviewed by Maria Russell.

Published
Last updated

About the reviewer

Maria Russell

Maria didn't grow up dreaming of chess. She grew up dreaming of getting through bedtime without a meltdown, and chess just happened to be the thing that worked.

Read full profile