Computers can now easily beat grandmasters. But as the World Chess Championship approaches, fans still want to see flawed humans battling out. David Edmonds investigates whyby David Edmonds / October 18, 2018 / Leave a comment
What is the difference between “almost” and “nearly”? What rules dictate the selection of one word in a sentence rather than the other? You’re unlikely to know. And yet if you’re a native English speaker, you’ll almost certainly (as opposed to “nearly certainly”) be able to use the two words correctly. You might say: “I am not nearly as good at chess as I am at backgammon.” You wouldn’t say: “I am not almost as good at chess as I am at backgammon.”
Humans are unable to articulate fully many things that we manage to do quite well. The rules of language are one such thing. The game of chess is another. Some people play chess well, but don’t really understand how they do it. What makes Magnus Carlsen the best player in the world? Obviously, the 27-year-old Norwegian grandmaster possesses a superb memory, an ability to calculate ahead and highly-developed pattern recognition skills. But that is true of all the top players. His real edge remains a mystery—even to himself.
On 9th November, Carlsen, the current world chess champion, will face the 26-year-old Italian-American Fabiano Caruana in London. Twelve games over three weeks decide who will become the next world champion. It will be the first time since Bobby Fischer’s 1972 “Match of the Century” against the Russian Boris Spassky that an American-born player has the chance to take the title—something US chess fans are getting very excited about. The Carlsen-Caruana match will be followed avidly by millions of others globally on the internet, or in person for those lucky enough to get a ticket to the Holborn venue. Chess is in rude health.
And that may be surprising to you. Twenty years ago, the continued longevity of the game was in question because of the arrival of an all-conquering machine interloper. Computers threatened to take the element of mystery out of the game, to reduce the flash of inspired brilliance to the whirring of an algorithm—and, as it seemed then, to render human players redundant.
In the annals of chess a watershed game took place in New York in 1997. Garry Kasparov had been world champion for 12 years. But this time he wasn’t playing a human. On the other side of the 64 squares was a computer. Even non-chess fans were enthralled. Kasparov began in a cocky mood. No human had posed him a serious challenge for a decade and he had developed an aura of invincibility. Nor, initially, was he fazed by the IBM-created Deep Blue, which he had beaten a year earlier 4-2. Ever since the invention of modern computers, programmers had tried to see how good at the game they could make them. But reaching a high standard had proved difficult. Chess, with its near bottomless complexity, was regarded as the ultimate reflection of human intelligence—a challenge too far for any silicon wannabe.
You probably know what happened next. Six games later—following Deep Blue’s devastating knight sacrifice in the final game—Kasparov’s crushed face was on newspaper front pages around the world. It hadn’t been a rout: the humbled world champion had notched up one win and three draws, as well as one loss, before the fateful decider. Nonetheless there was a sense that he had let down humanity. The competition, described as the brain’s last stand, heralded a turning point in the relationship between human and machine.
Kasparov was a terrible loser. Following his New York humiliation, he raged against IBM for poor sportsmanship. He had a point. Normally, when elite players prepare for a match they will analyse their opponent’s previous games, identifying their style, probing their weaknesses. But IBM wouldn’t reveal any of Deep Blue’s games—for Kasparov his opponent was a black box. There was also the psychological disorientation of playing against a machine, vividly described in his book,Deep Thinking. He became overwhelmed with dread that the computer had “seen” something that he had not. At the end of Game Two, for instance, Kasparov missed— frankly, didn’t even look for—a fairly simple tactic to force a draw, dazzled by the machine’s play earlier in the game. And so the trajectory of man versus machine was set. Man was improving incrementally. Computers were improving exponentially.
Now, 21 years on, the chess computer has matured. Any one of several chess apps on your mobile phone can outclass top grandmasters. They are almost impossible for humans to win against. To a non-player it may not be obvious why computers would affect chess. But the internet, and the arrival of powerful chess engines, have changed the game in multiple ways, some mundane, others profound.
“Mobile phones would ruin fair competition far more decisively than steroids pumped into a cyclist”
At the more humdrum end, the internet allows chess games to be watched move-by-move. In the last world championship in 2016 between Carlsen and the Russian Sergey Karjakin, Chess.com registered 14m page views. Watching games with the aid of a computer has an unseemly side-effect; not all spectators treat the competitors with the same respect and reverence that elite grandmasters once commanded. Average players, fortified no doubt by a bronze medal in their local club ladder, can now see the verdict of their chess app on the grandmaster’s moves—and then pile in. The comments under live games are full of patzers (weak chess players) masquerading as professionals. “What a blunder by Caruana!”, they gleefully type, having checked their phones.
Computers have changed the way the professionals compete as well. When Fischer squared off against Spassky in Reykjavik, the games were adjourned after five hours, and their “seconds”—chess experts who assist in preparation—would help them analyse before the game reconvened the following day. Now that option is unavailable, because the moment they are out of the hall, a chess engine would instantly reveal the best lines of play. Each Carlsen and Caruana game will be concluded in one day—and might go on for eight hours.
During games, Carlsen and Caruana will be prohibited from using electronic devices. These days, in amateur as well as professional chess, mobile phones are banned or have to be switched off—they would ruin fair competition far more decisively than any cocktail of steroids pumped into a professional cyclist. Three years ago, in one of several high-profile cases, a complaint was made in a Dubai tournament about the then-Georgian champion, Gaioz Nigalidze. He had been making frequent visits to the toilet. The explanation, as his opponent suspected, was not a weak bladder, but a mobile phone, discovered hidden in a cubicle and camouflaged by toilet roll. He was banned from competing for three years.
Computer support in preparation for a game is, however, permissible and it is one reason that players are getting stronger at a younger age. Computers may also explain why very good players now emerge from places where there has been no vibrant chess scene. It is easy to rack up 10,000 hours of practice by competing against players from around the world—or computers—on your laptop. “Computers are powerful learning tools,” says the former British number one Nigel Short (who lost to Kasparov for the world crown in 1993). He points out that computer databases are ever more complete. Fischer had to request a file of all Spassky’s games—a laborious process that required them to be hunted down and bound in a book that Fischer carried around like a Bible in the run-up to Reykjavik. Now an opponent’s games are available to study at the touch of a button.
All the classic games of the past are there to view. Not that they carry much fascination for the younger generation. In the old days every chess player would know their chess history. They might pick up lessons in technique from that early 1960s giant “the magician of Riga,” Mikhail Tal, or would master the art of positional play studying the games of his Armenian successor as world champion, Tigran Petrosian. “People are forgetting history because they regard it as irrelevant for chess instruction,” says grandmaster and chess commentator Danny King. Players believe they can learn everything they need from the machine.
Computers have helped human chess players in myriad ways. But beyond their training application are computers affecting the nature of chess games themselves? That question is more difficult to answer.
Deep Blue was programmed with the aid of strong chess players. Because the number of possible positions in the average chess game exceeds the number of atoms in the Universe, it was thought unfeasible to rely on computational power alone. Instead, the computer had to limit how many moves it assessed. It had to be programmed to examine only a few branches in the tree of possibilities. There were lots of ways of doing this. If a move loses a queen with no compensation, it and all subsequent variations can be safely ignored (though even this is not quite as straightforward as it sounds because the computer has to evaluate what counts as compensation). But, broadly speaking, Deep Blue’s success did not rely on direct human-style understanding—it rested on brute processing.
Naturally a machine has other advantages beyond calculation. Humans suffer nerves, they can be intimidated, they become exhausted, their mood may be affected by a headache or a row with their spouse. Machines are not prey to any such weaknesses. Humans can become gripped by the narrative of a game—that, for example, they have to prise open their opponent’s king side defence before their opponent forces a breakthrough on the queen side. Convinced that this is the essential strategic battle they may miss a shift in the game’s dynamics. Computers, by contrast, do not live out stories in their heads and they have no loyalty to strategies of the past—for each move they react only to the position on the board, irrespective of how that position was reached.
In the end, though, what computers have that humans do not is an incomparable degree of computational power. The latest generation of programs, which use so-called “machine learning,” have taken this to a whole new level. With machine learning the computer spots correlations and patterns from tens, hundreds, thousands, millions, billions of examples, from actual human games or from games in which the computer plays itself. Mining this data allows it to predict what moves will be successful. It need have no “understanding” of the game beyond the basic rules. Instead of the programmer inputting heuristics, the machine identifies what works—and evolves its own heuristics and strategies.
Making the most interesting moves in the application of machine learning to games is DeepMind, a company spearheaded by former British junior chess prodigy Demis Hassabis and now owned by Google’s parent company Alphabet. In 2016 their computer program AlphaGo took on the champion player Lee Sedol at Go, the ancient Chinese game, for a prize of $1m. Sedol has won multiple international titles and is a celebrity in his native South Korea. Go, a game in which you have to claim territory, is far more computationally complex even than chess—for starters it has a much larger board: 19 by 19.
Once the rules of Go had been programmed into AlphaGo, it mastered how to play through countless games against itself. In Game Two against Sedol, AlphaGo made one move—move 37— on the upper right side of the board, the fifth line, that has already become legendary and which no expert would have considered. At the time the commentators were baffled: it looked like a beginner’s blunder. But it was a masterstroke and turned out to be the stone that swung the game. Sedol had to leave the room to compose himself. “I am speechless,” he said after his defeat.
DeepMind then achieved something more astonishing still. In December last year their program AlphaZero played 100 games against the mighty chess engine Stockfish and crushed it by an undefeated score of 64 to 36 (draws count for half a point). Chess players have what is called an Elo rating, which rises or falls depending on how players perform against other rated players. Carlsen and Caruana have an Elo rating of above 2,800—one day a human might break the 2,900 barrier. Stockfish is rated around 3,400. Top US grandmaster Wesley So said of AlphaZero, “I was shocked. This is the new big thing. It totally changes chess. It might be rated, what, 3,700? Close to 4,000? That’s really crazy.” But what was truly remarkable about AlphaZero was that it had reached its astonishing level a mere four hours after it had been taught the rules of the game.
“For all the advances in computer chess, spectators want to watch humans play because of their flaws and frailties”
Machines can now reach these heights without having to “think” like humans. To humans, machine chess can be confusing. When, on Danny King’s YouTube analysis, his software identifies an improvement to a particular game, he will sometimes say, “but that’s a machine move.” Usually that means it is “anti-positional”—that is, it appears to weaken key squares or it undermines the strength or harmonisation of the pieces, and so will be dismissed by humans as a sensible option. For example, the move might strand a piece awkwardly on the edge of the board—something basic chess lessons generally counsel against, because on the side of the board pieces have fewer options. Humans rely on such general maxims to guide them because they can’t process every possible configuration of the unfolding game; but the computer can start from every particular situation and explore exactly how all the geometrical tactics happen to work out. Says King: “a machine doesn’t care if [a move] will pay off in 25 moves.”
The early chess machines were programmed with some understanding of the value of pieces. Chess beginners are taught that a pawn is worth one, a knight and bishop three, a rook five and a queen nine. This is a crude approximation. If the board is closed and crowded, the hopping knight can be more effective than a cramped bishop. At the outset of the chess computer revolution in the mid-1970s, the machines could be inveigled into guzzling pieces—grabbing a pawn, say, even when that led to its pieces being hopelessly uncoordinated.
The latest programs could not be more different. For them a piece is useful only to the extent that it contributes to a final victory. They are the ultimate pragmatists. And so they have what to humans would be a weirdly insouciant attitude to being down on material. As Nigel Short told me, “when I sacrifice a piece, I want to give checkmate in three moves! I want an instant return on my investment.” A chess app doesn’t require such quick gratification. It will give up a rook for a bishop so long as it will eventually reap the reward—a reward too far down the track for humans to envisage. In one of AlphaZero’s games against Stockfish, AlphaZero played a gobsmacking knight sacrifice on move 19 only to regain the material with a win ending 20 moves later.
The impact of chess computers has been to reduce the number of tactical attacking games between humans—games in which a combination of moves aims to force a quick and decisive advantage. Machines have shown how resourceful they can be in defence, how they can wriggle out of apparently desperate positions. Their Houdini-esque dexterity has produced a nervousness, a crisis of confidence, about attack. “Players have far more respect for their opponent’s defensive capabilities,” says King.
This has led to an increase in steady, strategic chess. That’s true of the choice of openings too—sharp openings which produce a pyrotechnical middle game are easily repulsed by engines. Many players now prefer a quieter approach. One opening, the Berlin Defence, which often leads to the queens being exchanged and has a reputation for being dreary and leading to draws, has seen a renaissance in the modern game. It may not be scintillating for spectators, but its defensive subtleties sometimes defy the computer’s single-minded approach.
One of the programmers of Deep Blue, Murray Campbell, a decent player himself, says it’s enjoyable to watch games with the help of a computer. “Having some insight as to what is going on, who is really winning, anticipating mistakes that players may make, marvelling when the players understand better than the computer, all of these make chess more of a spectator sport.” And he also brings a welcome reply to concerns that chess will one day be cracked by computers—meaning that the computer will always either win or draw against any opponent—like draughts was in 2007. “Chess is not solvable with any currently known technology.” In other words, unlike draughts or noughts and crosses, the outcome cannot be simply predicted given perfect play.
For all the extraordinary advances in computer chess, it turns out that spectators want to watch humans play, with their flaws and frailties—indeed precisely because of their flaws and frailties. They relish rivalries, they are intrigued by the ups-and-downs of player form, they love the clash of styles and approaches. No computer thrum can compete with the intensity of human combat.
Despite this, all human players have now come to terms with a painful new worldview. As cars are faster than us, so computers are incontrovertibly superior on the 64 squares. Their calculations are more far-sighted and their rules are more effective. While we might struggle to say precisely when we’d use “almost” rather than “nearly,” for today’s computers, who so effortlessly codify rules from experience, it would be a breeze. The Dutch player Jan Hein Donner summed up our humbled status well. Asked for how he would prepare against a machine like Deep Blue, he replied: “I would bring a hammer.”
Man versus machine
The history of chess-playing computers
1770: The chess-playing automaton known as the Mechanical Turk was invented by Wolfgang von Kempelen. The machine beat Napoleon before being revealed as a hoax—a hidden human player operated the pieces. Yet it did inspire Charles Babbage, an early 19th-century pioneer of artificial intelligence.
1951: Dietrich Prinz, a colleague of Alan Turing, wrote the first chess-playing program, for the Manchester University Ferranti Mark 1 computer. The program took 20 minutes to analyse the board and couldn’t perform basic chess moves like castling, or distinguish between checkmate and stalemate.
1968: Larry Atkin and David Slate (left) began work on the pioneering “Chess” program. In 1976, it won its first tournament against a human opponent. David Levy, a British champion, bet Atkin and Slate that he could beat their computer. He defeated Chess version 4.7 over a six-game match in 1978 and declared it a “victory for humanity.” And for a time it was.
1988: Deep Thought—the name was taken from the work of cult author Douglas Adams—was developed at Carnegie Mellon University and later at IBM by a team led by Taiwanese-born Fenghsiung Hsu. It became the first computer to win against a grandmaster in a regular tournament game when it beat Bent Larsen (right) in 1988, but was defeated by Garry Kasparov.
1997: Deep Thought was renamed Deep Blue by IBM and, assisted by grandmaster Joel Benjamin, the computer grew in sophistication during the 1990s until it faced world champion Garry Kasparov (left) in 1996. The Russian won the first contest 4-2 but a year later, after an upgrade, the computer triumphed 3.5-2.5.
2017: Stockfish is a free, open source chess engine that you can download on to your mobile phone within a few seconds. Difficulty levels range from Passive to Aggressive to Suicidal. In a November 2017 computer-only competition, Stockfish beat Houdini by 10.5-9.5. It was recently superseded by AlphaZero