Friday
Dec072018
AlphaZero vs. Stockfish, the Rematch
Friday, December 7, 2018 at 10:07PM
It was a year ago that AlphaZero took the chess world by storm, beating Stockfish in a 100-game match. AZ won 28 games and drew the remaining 72. This year, there was another match. The time control was slower than the minute-a-move control from last year; this time it was 3 hours for each side for the game, plus a 15" increment after each move. This time they played 1000 games, with 839 draws, six wins for Stockfish (yayy) and 155 wins for AlphaZero.
More about the match, including download links, here.
tagged AlphaZero, Stockfish, computer chess
Reader Comments (3)
The chess programming community has a greater deal of skepticism about this news than the general chess-playing public. Two concerns: 1) The claimed results are against an old version of Stockfish and represent only a 52 ELO advantage. Hardly “crushing.” 2) The commercial motive here is obvious: Google is using the spectacle of a machine learning chess engine to entice corporations with a mountain of Internet-of-Things data to partner with Google to analyze the data to gain insights and competitive advantage. Considering the obvious profit motive, independent testing is required to determine if Google's claimed results are reproducible or just an advertisement.
Winning more than 100 games more than your opponent (even in a 1000 game match) qualifies as "crushing" in my book.
To me the important thing here is, unlike the last match, Alphazero did lose six games. This means it is not playing perfect Chess, yet.