How does the Stockfish chess machine work

How Google's AI changed chess

Exactly a year ago, Google's (or alphabet) AI company DeepMind stunned with the news that the artificial intelligence AlphaZero had clearly beaten the strongest chess engine in the world - without any external knowledge of the strategy and after just four hours of learning. AlphaZero was originally designed for Go and, as in the chess-like Shogi, had defeated both man and machine. The aim of the project is to develop an AI that masters every game.

A little skepticism was mixed in with the amazement: DeepMind had only published a few games; The hardware performance and the cooling off period are said to have been at the expense of the opponent Stockfish. Now DeepMind publishes the results of their work in "Science" and comes out with a little more details.

AlphaZero vs. Stockfish

AlphaZero combines a Monte Carlo tree search with a neural network, while their opponent Stockfish 8 bases their skill level on alpha beta search and expertise. This approach has been by far the most successful in chess to date. On commercially available hardware, the program achieves an unofficial rating of 3430 - almost 600 points more than world champion Carlsen and only surpassed by the successor Stockfish 9, which has since appeared.

Stockfish ran on a computer with 44 Broadwell cores, AlphaZero on special hardware based on TPUs. The latter are processors developed by Google for tensor calculations. It is difficult to compare the hardware capabilities with one another - also because DeepMind is reluctant to provide information here. In any case, the number of positions calculated is unequal: While Stockfish analyzes 60 million positions per second, AlphaZero only manages a thousandth of them.

Factors: opening book and time to think about it

At the beginning of this year, the two computers dueled each other in a thousand games without an opening book and with a long time to think about it (three hours per game plus increment). Additional matches over several thousand games were based on predetermined opening positions that came from opening theory or computer chess. Eventually, they experimented with reduced cooling-off time for AlphaZero, with the then still experimental version 9 of Stockfish and with the opening book switched on.

In the main match the result is a little less brilliant than the one previously published (57.5% instead of 64%), but 155 wins and 839 draws with only 6 defeats speak for themselves. AlphaZero Stockfish dominated even with a tenth of the cooling time of its opponent, only with further reduction did the ratio reverse. While the results of the two Stockfish versions did not differ much, Stockfish defeated more closely with the opening book than without.

AlphaZero games look human

Strangely enough, DeepMind did not want to publish all the games this time either, but 210 of them can be replayed by interested parties. Some of them are very impressive and demonstrate the AI's superior understanding of the game. While computer games often seem very tough and puzzling, AlphaZero games make a very human impression. The AI ​​does not shy away from long-term speculative sacrifices and is able to develop inconspicuous positions that Stockfish still believes to be in equilibrium.

The LCZero project proves that AlphaZero's success reports are not out of thin air: This open source project is based on the research results of DeepMind and recently achieved third place in a computer tournament.

(bme)

Read comments (84) Go to homepage
display