AlphaZero learned to play chess by playing against itself. After just FOUR HOURS of self-learning, it was able to decisely defeat Stockfish 8.0! (EDIT: this statement is slightly misleading. See downthread.) (100 games match: +28 =72 -0)
What’s really impressive: Stockfish was calculating far more deeply than AlphaZero (at least in terms of nodes per second). AlphaZero is just “smarter.”
There’s a ten-game selection from the match on chess24: this win is one of the most beautiful attacking games I’ve ever seen (Qh1!!!)
[EDIT: This is the final game cited in the arXiv paper, bottom of page 18. The point of AlphaZero’s play is that Black is tied up in knots despite being up a knight and two pawns. This is a more advanced version of the anti-computer type of thinking that humans used to beat computers, back in the day. I wonder whether Black could hold with a move like 34…a5.]
As the paper notes, shogi (more resistant to AI than chess, because it allows for “reincarnations” of captured pieces as in bughouse) has also been “solved” by AlphaZero.
It’s for real. And in a way, it’s not unexpected. No disrespect to our great game, but the accomplishments of the AlphaZero team in go and especially shogi are even more mind-boggling.
One clarification: while AlphaZero was already stronger than Stockfish after four hours, the “fully trained” AlphaZero appears to have been trained for some unspecified longer time frame. (My guess, based on what I’ve read about AlphaGo Zero, is that the chess version of AlphaZero was trained between a few days and a month.)
See Figure 2 on page 7 of the paper: AlphaZero’s rating at tournament speeds is at least 3500! In bullet chess, on the other hand, brute-force calculation is more important than intelligence. At 0.1 seconds per move, Stockfish is still clearly stronger than this incarnation of AlphaZero. At 1 second per move, AlphaZero is slightly stronger than Stockfish.
Not only does AlphaZero play the Berlin (many GMs commented on that immediately), it independently chose the Chebanenko Slav as its primary defense to the Queen’s Gambit!
See page 6 of paper, main line of “D06: Queen’s Gambit.”
The comments section in the article indicated that the two programs did not play under equal conditions of computing power. It would be interesting to see the program tested against known endgame puzzle problems and extreme tactical middlegame positions to see if it solves them in unique ways.
If one gets past all of the puffery and self-congratulatory promotion, the development is impressive. It means that a tremendous amount of computing power, new chips unavailable to the masses, and enormous amount of cash is necessary to destroy the favorite board games (Go, Shogi, chess) of over three billion people in the world. Since the games are “finished”, the proles can go back to being drones, swilling beer and watching futbol, basketball, and cricket for the financial benefit of the entertainment and beverage industries. Chess players can drop their federation memberships and wander aimlessly looking for intellectual diversions on the level of tic tac toe, tiddly winks, and pick up sticks.
You’re talking like a nineteenth-century frame-breaker. Chess remains a game played between people. It has NOT been solved by AlphaZero. (If you read the fine print on games played from various openings, even AlphaZero lost 2% of those games to Stockfish.)
And play through the selected games (granted, selected to show AlphaZero in an extremely favorable light): it routinely sacrifices for long-term positional advantage.
I do NOT pretend to understand any of this! Would appreciate layperson-accessible explanations from folks who do.
P.S. Hikaru Nakamura & others pointed out that Stockfish was not optimally confugured. Stockfish was playing with only a 1 GB hash and no access to its opening book, while AlphaZero had a de facto opening book because of its massive “self-study” program. All this may have magnified AlphaZero’s margin of victory.
I almost dropped your name when I posted the link in my previous post. Even though it was a French that White won, trust me, you have other reasons to enjoy it.
If I am a “frame-breaker” or Luddite, then I am in agreement with a number of economists and futurists who are disturbed by the implications of AI. The Luddites, like their ancestors in previous centuries, were not afraid of machines but were concerned with the consequences for society of how machines were being used to replace labor and the accumulated expertise of workers. 19th century labor had few mechanisms to protect themselves from the growing force of the Industrial Revolution. There were no labor unions. Governments were small and decidedly in favor of those who had capital, wealthy landowners, and the aristocracy who controlled the government. The “breaking” of machines was a long used practice of protest by workers who had nothing like the “safety net” of unemployment insurance, food stamps, social security, or medical care we have today.
Back in the late 1980’s and early 1990’s when the first really strong computer programs, Hitech and Deep Thought appeared with 2450-2550 ratings, the chess Luddites saw the devastating consequences for tournament play if the programs were allowed to enter the lists to compete with humans. Within a short time, the option of signing a sheet so that you did not have to play a program became a ban on the programs being allowed into USCF events. The little designation “NC” was attached to TLAs by organizers to allay the fears of entrants. I recall players offering to “pull the plug” on Hitech in one of its tournament outings. I directed a couple of tournaments that allowed the computers to play. Curiosity and dismissal of the strength of the machines turned into animosity toward the programmers who were charged, rightly or wrongly, with trying to kill chess. Having played and beaten the behemoths, I am well aware of their strengths and weaknesses. The weaknesses were patched over by improved hardware speed, but some persist to this day. DeepMind’s AlphaGo, which is a descendant of the earlier programs in so many ways is now a creature of Google, a multinational corporation with many interests. Whether Google has good or bad motivations may be unclear, but it is worrisome that they have spent so much money win games of Chess and Go, and then acts dismissively of the games as it moves on to new territory to conquer. Any lessons learned are designed to make money, not just engage in theoretical research on neural networks.
As the programmers progress from crushing pastimes like backgammon, checkers, chess, Go, and poker, etc. they have the potential to leave in their wake a diminishing of the popularity of the games and an undermining of people’s fundamental belief in the worth of humanity. In many ways, people express themselves and their creativity in the playing of games. When the games are treated as being “finished” or not worth the time to play, we are diminished as well. AI may destroy the dignity of work along with the dignity of creativity. I am not afraid that AI will kill us; it will need us to serve it. See Hawking, Musk, and Friedman for various scientific, social, and economic views on what AI might bring in the coming decades. I am not so sure it will be good for us to live in a “smart home”, driverless cars, and a dole from the government to pay for basic needs as millions of white collar and blue collar jobs disappear. Policy makers are nowhere near even beginning to understand the possibilities of an AI world. They are too busy trying to give tax breaks to those who will own AI.
The 19th century Luddites couldn’t stop the coming of the Industrial Revolution by trying to break the machines any more than King Knute could stop the tide from coming in by ordering it not to do so. Those trying to stop the coming of AI will have similar success. The question is not whether AI is coming, it’s how to deal with it. I’m not arguing this is a good thing. I’m saying it is what it is.
Santa, please, an algorithm implanted in my brain for Christmas. Skip the other Chessbase products on my list. And I promise to read Kasparov’s Deep Thinking which has been on my book-stand. Promise.
I don’t think there is much of a threat to people playing and keeping popular games such as Go and chess. Humans will accept that they can’t defeat a computer program with or without AI. You correctly note that the problem is centered on how we as humans living and working in a variety of cultures (corporate, democratic, religious, etc) use or allow AI to be used. There is tremendous positive potential in many areas (medical research?), but also tremendous possibilities for mischief and, dare I say it, evil.
Chess remains a game played between people. It has NOT been solved by AlphaZero. (If you read the fine print on games played from various openings, even AlphaZero lost 2% of those games to Stockfish.)
Would appreciate layperson-accessible explanations from folks who do.