Chess has long been one of the most effective ways to evaluate a computer program’s ability to “think.” With sixty-four squares on the board, sixteen pieces per player, and a variety of rules for how each piece is allowed to move during the match, chess is a game in which thousands of scenarios are possible, even when considering only a few moves at a time. Consequently, manufacturers like IBM, Google, and some independent programmers have worked to create stronger chess machines over the last fifty years. More recently, advances in technology have propelled machine learning into a level of sophistication far beyond what a human player is able to achieve.
For example, Stockfish, the unofficial world computer chess champion for over five years, has a rating far above that of the best grandmaster players. Stockfish’s programming considers around 10,000,000 moves before making a decision. So how did DeepMind’s AlphaZero, a neural network that considers only around 10,000 moves per decision, convincingly outperform Stockfish after just a few hours of training?
The answer may redefine the way we view chess. AlphaZero’s unconventional style immediately attracted attention from the game’s commentators and experts, who remarked on the machine’s strange, almost human-like thinking and its counterintuitive disregard for what players traditionally see as advantages. For example, chess players start the game with an equal number of pieces, whose values are estimated using a commonly known point system. A pawn is a one, a knight is a three, a bishop is also a three, and so on. The idea is that more pieces on the board translates to an advantage during the match. This concept—and its straightforward evaluation—is convenient for commentators, who can make use of a concrete method to consider which player is “winning.” But AlphaZero blatantly disregards the material on the board. The machine often sacrifices its pieces early in a match, giving space, material, and seeming freedom to its opponent in exchange for a positional freedom that will only benefit it later in the game. It clearly values piece activity, the frequently practiced concept of maneuvering each piece to a place of relevant participation on the board, but perhaps for different reasons than those proposed by traditional theory. It seems that AlphaZero is instead concerned with the natural result of this strategy: manipulation. Some of its games against Stockfish contain truly masterful symphonies of attack, in which the neural network manages to unconventionally pressure not one but both sides of the enemy’s territory. There is no doubt, Chess Grandmaster Matthew Sadler says when commenting on AlphaZero’s style, that its visions and thinking methodology surpass the ordinary, while “its pieces swarm around the opponent’s king with purpose and power.”
AlphaZero’s neural network also trained itself in shogi (Japanese chess) and Go (the world’s oldest board game) until it defeated a world champion program in each case. Just as with chess, the machine started with nothing more than the basic rules of the game, and while it discovered many tactics echoed by traditional human theory, it also invented its own adaptive style. Professionals like Yoshiharu Habu, the only player in history to hold all seven major shogi titles, call it “unique,” “dynamic,” and even “groundbreaking.” This aspect of the machine is most admired by its creators and commentators, for it might mean that DeepMind can apply AlphaZero’s thought to real world problems.
Former world chess champion Garry Kasparov is especially inspired by the machine’s success. “We have always assumed that chess required too much empirical knowledge for a machine to play so well from scratch, with no human knowledge added at all,” he said. “Of course I’ll be fascinated to see what we can learn about chess from AlphaZero, since that is the great promise of machine learning in general—machines figuring out rules that humans cannot detect. But obviously the implications are wonderful far beyond chess and other games. The ability of a machine to replicate and surpass centuries of human knowledge in complex closed systems is a world-changing tool.”
Whatever may be in AlphaZero’s future or in that of the DeepMind division of Google where it was born, its use of reinforcement learning is undoubtedly groundbreaking. What’s more, the machine reminds us that chess, and perhaps other games often reduced to tried-and-true strategies, have sparks of artistic possibility woven into them. AlphaZero’s unique approach to a classical pastime may translate to architecture, research, or mathematical theory. The machine may soon move aircraft carriers instead of pawns, observe physical laws instead of knights, or consider the decisions of world leaders rather than possible moves for a bishop. Yet even AlphaZero is still a piece in a greater game, a masterful stroke of evidence that life, whether analyzed in Google’s research labs or not, is more than just numbers.
Article Sources:
Klein, M. (2017, December 6). Google’s AlphaZero Destroys Stockfish In 100-Game Match. Chess.com. https://www.chess.com/news/view/google-s-alphazero-destroys-stockfish-in-100-game-match.
Silver, D., Hubert, T., Schrittwieser, J., & Hassabis, D. (2018, December 6). AlphaZero: Shedding new light on the grand games of chess, shogi and go. DeepMind Blog. https://deepmind.com/blog/article/alphazero-shedding-new-light-grand-games-chess-shogi-and-go.
TheBestSchools Staff. (2021, April 9). A brief history of computer chess. TheBestSchools.org. https://thebestschools.org/magazine/brief-history-of-computer-chess/.
Image Credits:
(2016). Figures Space Planet Moon Astronaut Chess. Max Pixel. https://www.maxpixel.net/Figures-Space-Planet-Moon-Astronaut-Chess-2314384.