AlphaZero, Google’s Newest Contribution to Chess AI

AlphaZero, Google’s Newest Contribution to Chess AI

The powerful new chess bot taught itself chess in four hours and plays like nothing we’ve ever seen.

AlphaZero, Google’s Newest Contribution to Chess AI

Google’s DeepMind AI Lab has invented a fully autonomous chess computer with an estimated Elo rating of 3600. Observers say it plays the game not like a human, and not like a machine, but more like an alien. What new things can AlphaZero teach us about this ancient game that it mastered in only 4 hours?

The History of Artificial Intelligence in Chess

People have been fascinated by the idea of artificial intelligence in chess since before they had the language to describe such an idea, and well before the technology existed to be able to accomplish it. The spectacle that the Mechanical Turk created when it was presented to audiences all across Europe as a machine that could play chess on its own, often beating its human opponents, is a testament to this fascination.

Of course, we now know that the famous Mechanical Turk was actually operated by a human grandmaster hidden inside its complex-looking mechanism. The feat of engineering and showmanship that it took to keep up the hoax of the Turk for so many years is on par with the level of ingenuity needed to create real chess playing machines, we think, but the idea was well ahead of its time.

The Mechanical Turk

The Mechanical Turk

It wasn’t until the early 20th century that a chess computer was actually invented. It was Alan Turing who first experimented with the idea of a computer being able to play chess, out of pure interest in whether a computer could be made to “think like a human” so to speak. Claude Shannon expanded on Turing’s work, though he considered it to be of no practical purpose, merely theoretical interest.

The first functional chess-playing computer program was written in 1951 by Turing’s colleague, Dietrich Prinz. It ran on the Ferranti Mark I computer which was housed at Manchester University, and while it couldn’t complete a full chess game due to computational and memory limitations, it did successfully solve the “mate-in-two” problem it was given, where the position stood as two moves away from checkmate and the computer had to determine the best move to make.
Dietrich Prinz

Dietrich Prinz

AI vs Human Players

Clearly, chess artificial intelligence at this moment in history was no match for even a modestly skilled human player, but the technology continued to improve over time, and the strength ratings gradually rose until they could actually challenge and even defeat a human grandmaster.

That Grandmaster was Garry Kasparov, and the year was 1997. He had handily beat the chess computer called Deep Thought in prior matches in the previous years, but it had been acquired by IBM, worked on by their top minds, and renamed Deep Blue. It was ready for a rematch and, in a game that shocked the world, it won. That was the moment that the tide essentially turned on chess artificial intelligence. If Deep Blue could calculate 200 million moves per minute, a number that now seems comically low in comparison to what modern chess computers can do, and beat the strongest chess player in the world, what hope did humanity have left?

Garry Kasparov vs Deep Blue, A game that shocked the world

Garry Kasparov vs Deep Blue, A game that shocked the world

Nowadays, almost everyone walks around with a device in their pockets that’s capable of handily beating anyone they ever meet in a quick game of chess. Even the most mundane chess software is stronger than any human player could hope to become by several orders of magnitude. The field of chess computing has gotten so advanced that we can hardly even understand any improvements that would be made at this point, so it would seem that the work there is pretty much done.

Enter AlphaZero, an entirely new kind of chess AI.

What Makes AlphaZero So Special

The driving force behind all of the chess computer programs to date has been, essentially, human ingenuity. It was human scientists and mathematicians who developed the algorithms that allowed the computers to perform the calculations necessary to choose the right chess move. In essence, humanity taught the computers how to play chess.

What makes AlphaZero so unique is that it has managed to teach itself the game of chess with absolutely no human input on strategy, completely mastering it in a matter of hours.

The team of researchers from Google’s DeepMind AI Lab programmed AlphaZero with the rules of chess only, no strategy, and left it to do its thing – its thing being in this case surpassing all human knowledge of the game of chess garnered over thousands of years in the same amount of time it might take to take in a double feature at your local movie theater.

Perhaps the most interesting thing about this new AI is that, in contrast with previous versions that have learned strategy partly by playing against strong human players, AlphaZero refines its strategy exclusively by playing against itself. Researchers initially thought that this fully autonomous design might be a handicap for the chess AI, but it actually has turned out to be a boon. Apparently, versions that learned from watching humans picked up bad habits and mistakes from their games. AlphaZero’s play is free from these blunders and, as far as we can tell, perfect.

AlphaZero vs Stockfish

Since humans are no longer any match for chess AI, the DeepMind team tested AlphaZero’s smarts by pitting against the current strongest chess AI out there, Stockfish. In the first test, AlphaZero won 25 games as White, 3 games as Black, and there were 72 games in which neither AI recorded a win or loss. It’s only the first test of 100 games, and we’re sure that there will be plenty more to come, but for now, it certainly seems that AlphaZero is stronger than the reigning champ Stockfish.
Stockfish vs AlphaZero

AlphaZero vs Deep Blue

Deep Blue may have been the first chess AI to show that it’s possible to beat a human grandmaster, and it’s certainly one of the most well-known names in chess AI, but when it comes to strength against modern chess machines, there is just no comparison. AlphaZero has not actually faced off against Deep Blue, but honestly, there would be no point. Both Stockfish and AlphaZero are much, much stronger than Deep Blue. For comparison, Deep Blue had an ELO score of about 2700, and AlphaZero’s Elo score is estimated to be something more like 3600.

A League of Their Own?

Since chess AI cannot be sufficiently challenged by human players, only by each other, you may be wondering if we’ll ever start seeing intra-AI matches as a spectacle. Perhaps they could even start their own chess league!

The answer to that question is unclear. It’s certainly possible, and if there is interest in the possibility, it may yet come to fruition. The machines that are currently being built are primarily used for study and research purposes, but there’s really no reason that they couldn’t play recreational games in their “downtime.”

Robots Playing Chess

A league of their own? It’s certainly possible

It may not turn out to be what audiences expect, however. Modern chess AI “think” in a very different way than humans do, and that shows in the way that they play chess. Researchers familiar with the games that AlphaZero has played describe it as being completely alien. AlphaZero’s playstyle certainly doesn’t match a human’s, but it also doesn’t really look like a machine’s. One observer described it as being like watching, “an alien civilization inventing its own mathematics.” Another added,

“I always wondered how it would be if a superior species landed on Earth and showed us how they played chess…Now I know.”

Given these differences in the way that humans and chess AI play the game, there would be a significant difference in the ambiance at chess matches, and some people may even find the experience downright eerie!

Chess Champions React to Chess AI

Given the extreme strength of these new chess computers, you may expect human grandmasters to be hostile toward them. After all, they could be putting them out of a job! Actually, opinion is split. Garry Kasparov, the former world champion who was famously beaten by Deep Blue, has probably thought about this particular issue more than most people, and he says that human chess players should not be trying to beat chess AI, but rather merge with them to form a new kind of augmented intelligence.

Human and Machine Hands

Human chess players should not be trying to beat chess AI, but rather merge with them to form a new kind of augmented intelligence

After seeing what AlphaZero was capable of, Kasparov was wowed, saying, “AlphaZero has found things we didn’t know about the game.” He sees the role of this new class of artificial intelligence as not replacing humans, but rather offering them a promotion. As machines become better at performing tasks that humans have traditionally done, the humans themselves will be free to pursue their unique talents and engage in more creative tasks.

The Future of Chess

While it may be true that by the year 2080 we will all be watching giant robots facing off against each other with a giant chess set complete with giant chess (hover) board, in the meantime we’ll content ourselves with watching Magnus Carlsen beat all competitors at the World Chess Championship. After all, his games, and the games of all other human grandmasters, have a little something that chess AI still can’t replicate. Call it heart, soul, or even imperfect play, but it’s what makes the game so fun to watch.

2 comments

  1. Being an AI developer and a EE who’s designing a customized CPU, I’d like to say that the Alpha-Zero vs Stockfish match is 100% useless data. Alpha-Zero ran on a high end customized cpu designed for this type of neural networking. Stockfish ran a general cpu. Also Deepmind stripped Stockfish of it’s book moves, gave it just 1GB of cache, and the settings were not well set. As for the Alpha-Zero only doing 80,000 positions / second. That’s incorrect. Alpha-Zero tree search calls the NN 80,000/s, which in turn analyzes the entire board. In short, Deepmind has no idea how many equivalent positions per second Alpha-Zero analyzes.

Leave a Reply

Your email address will not be published. Required fields are marked *

Share via