DeepMind and Kramnik tapped AlphaZero’s ability to learn a game from scratch to explore new variants more quickly than the decades or centuries of human play that would reveal their beauty and flaws. “You don’t want to invest many months or years of your life trying to play something, only to realize that, ‘Oh, this just isn’t a beautiful game,’” says Tomašev.
AlphaZero is a more flexible and powerful successor to AlphaGo, which laid down a marker in AI history when it defeated a champion at Go in 2016. It starts learning a game equipped with only the rules, a way to keep score, and a preprogrammed urge to experiment and win. “When it starts playing it’s so bad I want to hide under my table,” says Ulrich Paquet, another DeepMind researcher on the project. “But seeing it evolve from a void of nothingness is exciting and almost pure.”
In chess, AlphaZero initially doesn’t know it can take an opponent’s pieces. Over hours of high-speed play against successively more powerful incarnations of itself, it becomes more skilled, and to some eyes more natural, than prior chess engines. In the process it rediscovers ideas seen in centuries of human chess and adds flair of its own. English grandmaster Matthew Salder described poring over AlphaZero’s games as like “discovering the secret notebooks of some great player from the past.”
The nine alternative visions of chess that AlphaZero tested included no-castling chess, which Kramnik and others had already been thinking about, and had its first dedicated tournament in January. It eliminates a move called castling that allows a player to tuck their king behind a protective screen of other pieces—powerful fortification that can also be stifling. Five of the variants altered the movement of pawns, including torpedo chess, in which pawns can move up to two squares at a time throughout the game, instead of only on their first move.
One way of reading AlphaZero’s results is in cold numbers. Draws were less common under no-castling chess than under conventional rules. And learning different rules shifted the value AlphaZero placed on different pieces: Under conventional rules it valued a queen at 9.5 pawns; under torpedo rules the queen was only worth 7.1 pawns.
DeepMind’s researchers were ultimately more interested in the analysis of the other great chess brain on the project, Kramnik. “This is not about numbers, but whether it is qualitatively, aesthetically pleasing for humans to sit down and play,” says Tomašev. A technical paper released Wednesday includes more than 70 pages of commentary by Kramnik on AlphaZero’s explorations.
Kramnik saw flashes of beauty in how AlphaZero adapted to the new rules. No-castling chess provoked rich new patterns for keeping the king safe, he says. A more extreme change, self-capture chess, in which a player can take their own…
Read More News: AI Ruined Chess. Now, It’s Making the Game Beautiful Again