More than 20 years ago, Deep Blue beat the world chess champion, but that system didn’t make much of a real creative contribution. It’s different now. It’s a standard casual game, played in the bedroom, from opening to checkmate in about 9 minutes. On one side is world chess champion Magnus Carlsen, a worthy child prodigy who became a grandmaster at the age of 13. On the other side is the Play Magnus App, an iPhone App that mimics Carlsen’s chess habits. Carlsen lowered the age of the machine to 18 and then fought the program, but Carlsen was still challenged. For the first few minutes, Carlsen was flummoked by an unexpected attack, then continued to fight for a level with the App, but eventually surrendered. We seem to see apps condescending to say, “You need to hone your chess skills, let’s try again!” Carlson could only smile back.
There is nothing special about the event. In fact, Carlsen has posted several videos of himself playing against virtual chess players of different ages. The video makes it clear: win or lose, the computer is Carlsen’s least favorite opponent. The problem cannot be avoided. Carlsen may indeed be the best chess player in human history, but how did such a player get beaten over and over again by computers, and how did humans get here?
The story of 1997
For those of you who don’t know much about chess, the story of how computers conquer chess begins with Deep Blue, which defeated world Go champion Gary Kasparov in 1997. Since then, the machine has shown its advantage, easily munching on the beautiful patterns and strategies thrown at it by humans. But modern analysts come to a different conclusion: the machine is fragile, Kasparov made a lot of mistakes, and there were obvious missteps on both sides.
The first game was won by Deep Blue, but in the second game, there was a change in the last move. Deep Blue had a chance to take a pawn, but it retreated, and Deep Blue took another step. It blocked the possibility of Gary Kasparov’s counterattack. The computer’s behavior exceeded Gary Kasparov’s expectations, and he was so disturbed that he missed the opportunity to draw. After the game, Gary Kasparov accused Deep Blue of cheating, arguing that a supermaster helped the computer make unexpected moves.
The controversial move may have been an accident. Years later, Murray Campbell, the scientist who helped IBM design Deep Blue, explained that the move was the result of a bug that the team had quietly fixed before the third game began. Unfortunately, the damage has been done. Later in the game, Gary Kasparov wasn’t so confident. Unable to understand Deep Blue’s move, Kasparov wasted a lot of time trying to trick the computer with an unusual human move, only to make an early mistake in the sixth game that would decide the match.
In short, deep Blue won, but it was no feat for the computer industry. It won because of human error. What this tells us is that human beings have weaknesses, such as hesitation, fear, guessing, and fatigue, that make them vulnerable to attack. Deep Blue doesn’t perform well, but it’s tireless and consistent. When Kasparov’s intuition went awry, the computer won easily.
Human despair chart
Chess may be an elegant game, but Deep Blue’s strategy is aimed at ugly brute force. Deep Blue didn’t use neural networks or machine learning strategies. In contrast, Deep Blue speculates potential movements with great raw force, reaching speeds of 200 million steps per second.
The Deep Blue system evaluates each step against a variety of different parameters, and then assigns a value to each parameter. The researchers analyzed nearly a million chess games played by grandmasters and assigned weights to the parameters, which the grandmasters then optimized. Deep Blue’s way of playing chess is the equivalent of putting together countless master games, because the system has enough raw computational power to predict the future and avoid big mistakes.
Today, there are more than a dozen computer chess engines around the world, all running on standard hardware, and they rely heavily on 200 years of chess history. In competition, a chess engine can search vast databases to find the opening gambit before the game begins. When it comes to the plate, the system can ensure that it is in a good position. Before the end of the game, the system can use a variety of strategies, and it constantly searches the database to make every move nearly perfect.
As for the rules by which chess engines evaluate weights, they were developed with the help of a large group of chess masters. Contributors suggest changes to the algorithm, and then a beta version is created, with the old version pitted against the new until the researchers decide which version is better.
Chess uses the Elo rating system, which means that weights are given based on the likelihood of beating an opponent. But it is difficult to compare the performance of computers with that of humans, because few people can compete with them, and few are interested in doing so.
Machines can easily play 1,000 games in a row, so comparisons between computers and humans can only be estimated. Still, if you look at the data of today’s top humans and top chess engines, you can see a picture of human despair.
Statistically, the computer rules, but it’s not perfect. They can’t predict the end of a chess game because there are more possible outcomes than there are atoms in the universe. The engine doesn’t have to be perfect to beat the human world champion. The computer just has to be consistent, tireless and not make obvious mistakes.
AlphaZero has a good shot
Chess actually values accumulation, which may be overlooked by the layman. There are many chess champions who say that the new generation will eventually beat the old, not because they are younger or more energetic, but because they have more knowledge. When it comes to making chess moves, computers are inferior to humans, but something has recently changed.
In 2017, Google-funded company DeepMind demonstrated AlphaZero, the first generation of its deep learning system. AlphaZero starts with no built-in knowledge of chess, no list of opening moves or millions of master games, only the rules and nothing more.
But AlphaZero learns, and it learns fast. He plays chess with himself, and in a few hours he can reach master level. By the end of the day, AlphaZero was skilled enough to beat the limited Stockfish chess engine. Last year, Stockfish beat the full Stockfish.
As AlphaZero learns, humans can observe its progress, watching it evolve from beginner to master, and then continue to evolve.
AlphaZero and Stockfish use basically the same hardware, but AlphaZero analyzes one-thousandth of the steps per second that Stockfish does. AlphaZero’s advantage is not in speed, but in learning. Analyzing the moves, Kasparov lamented that AlphaZero, like himself, had a dynamic style. Matthew Sadler said: “AlphaZero has found the secret notebook of the best player of all time.” It’s like a chess-playing alien descending on humans.
AlphaZero has one big difference from previous computer chess programs: Instead of simulating humans, AlphaZero is equipped with a neural network and can understand the game itself. AlphaZero didn’t just beat humans, it may have helped them understand chess, the first computer system ever to do so. AlphaZero seems more significant than Deep Blue’s defeat of Kasparov