How Computers Beat Humans At Chess
The best human chess players cannot win against the computer programs anymore. The ironic part? The hard-earned expertise of human experts is the source of their own limitation.
Reading time: 4 minutes
When my son turned four, I offered him a chess board.
A couple of weeks before, during Christmas reunion, my nephew wanted to play chess against everyone. So my son watched this new game being played in front of him. He started to be intrigued. It felt like it was too early for him to learn, but his desire to say “Check!” to an opponent was just too strong. He couldn’t stop talking about chess afterward. Hence the small chessboard as a birthday gift.
On the box, a sentence described the purpose of the game:
Capture enough of your opponent’s pieces to put his king into a checkmate position.
Do you notice anything interesting about this sentence?
In 1997, Gary Kasparov lost a critical game. It was the first time a chess world champion was defeated by a computer program under tournament rules. This computer program was called Deep Blue and its victory was seen as a sign that artificial intelligence was rising to the level of human intelligence, including in some of its most complex intellectual tasks.
Since Deep Blue, many computer programs appeared, each one better than the previous champion. Until the appearance of Stockfish, an open source chess engine that started to dominate the Top Chess Engine Championship.
But in 2017 arrived AlphaZero, a different type of computer program.
Traditional chess programs like DeepBlue or Stockfish rely on thousands of rules handcrafted by human experts. AlphaZero doesn’t. Instead of being taught chess through exposure of the thousands of games played by the best human chess masters, AlphaZero discovered chess from scratch, playing games against itself — at the speed of 2 millions games per hour.
Today, AlphaZero can beat any human being or any computer program at chess.
To understand where its strategic strength comes from, an analysis of AlphaZero’s gameplay has been made. One major insight came up. It seems like AlphaZero is not obsessed about capturing opponent’s pieces.
The common (human) wisdom about chess was that, the more pieces you capture from your opponent, the easier it will be for you to corner his king into defeat. But, because AlphaZero didn’t learn chess from humans, it didn’t pick up this piece of wisdom. It just focused on the most effective ways to win a game, as proven by millions of simulations. And, it turned out, capturing your opponent’s pieces does not rank as high as humans thought.
Let’s come back to our chess box introduction sentence:
Capture enough of your opponent’s pieces to put his king into a checkmate position.
It’s worth noting that this sentence does two things. As expected, it explains the purpose of the game — put the opponent’s king into a position where it can’t move without getting caught (= checkmate). But it also highlights the strategy needed to get there — capture the opponent’s pieces.
The fact that this chess box used this piece of advice as a way to describe the purpose of chess is fascinating. I can’t decide if it is:
A smart shortcut, because it helps beginners have a short-term goal (capture any piece) to aim before going after the long-term one (capture the king)?
An acceptable simplification, because it remains a relevant strategy for human beings who will never possess the calculation power of machines?
A deep flaw, because it frames the game in a way that will ultimately become a limitation?
What do you think?