A New Artificial Intelligence Makes Mistakes—on Purpose
It took about 50 years for computers to eviscerate humans in the venerable game of chess. A standard smartphone can now play the kind of moves that make a grandmaster’s head spin. But one artificial intelligence program is taking a few steps backward, to appreciate how average humans play—blunders and all.
The AI chess program, known as Maia, uses the kind of cutting-edge AI behind the best superhuman chess-playing programs. But instead of learning how to destroy an opponent on the board, Maia focuses on predicting human moves, including the mistakes they make.
Jon Kleinberg, a professor at Cornell University who led the development of Maia, says this is a first step toward developing AI that better understands human fallibility. The hope is that it may therefore be better at interacting with humans, by teaching or assisting them, for example, or even negotiating with them.
One possible use, Kleinberg says, is health care. A system that anticipates errors might be used to train doctors to read medical images or help them catch errors. “One way to do this is to take problems in which human doctors form diagnoses based on medical images, and to look for images on which the system predicts a high level of disagreement among them,” he says.
Kleinberg says he chose to focus on chess because it is one of the first domains where machine intelligence has triumphed over humans. “It is this sort of ideal system for trying out algorithms,” he says. “Sort of a model for AI dominance.”
In addition, he says, chess has been studied intensely, making it something similar to the fruit fly, or drosophila, in biology. “Chess has the distinction of having been called the drosophila of psychology by Herb Simon and the drosophila of AI by John McCarthy,” Kleinberg says, referring to two giants of their respective fields.
Maia was developed using code adapted from Leela Zero, an open source clone of Alpha Zero, a revolutionary AI program created by the Alphabet subsidiary DeepMind.
Alpha Zero broke from conventional AI chess programs by having computers learn, independent of any human instruction, how to play the game expertly. Within the program, a simulated neural network contains virtual neurons that can be tuned to respond to input. For chess, Alpha Zero is fed board positions and moves generated in practice games, and it tunes its neurons’ firing to favor winning moves, an approach known as reinforcement learning. Alpha Zero can use the same approach to learn to play other board games such as checkers or Go with minimal modification.
The Cornell team modified Leela Zero’s code to create a program that learned by favoring accurate predictions of human moves. Other AI chess players, including Deep Blue, the IBM machine that defeated then world champ Garry Kasparov in 1997, may attempt to look ahead in a game by exploring possible moves. But Maia is unusual in how it focuses on finding the most likely move a human will play.
Maia was taught using data from LiChess, a popular online chess server. The result is a chess program capable of playing in a more human way. Several versions of Maia, tuned to different strengths of play, can now be challenged at LiChess.
More sophisticated forms of AI may eventually outstrip human intelligence in all sorts of domains, from mathematics to literature and beyond. But Kleinberg says “there’ll be a long transition period where AI and humans will be working together, and there’s going to be some communication between them.”
Well before that, AI that can predict and mimic human behavior could have immediate applications in chess and other games. “It’s a lovely idea,” says Matthew Sadler, a British chess grandmaster and the author of Game Changer, a book about the chess-playing capabilities of Alpha Zero. “There’s a huge need for club players to have engines that talk their language.”