A top computer scientist told us the games that artificial intelligence can't win
Google's DeepMind artificial intelligence team is making history.
Its AlphaGo program is up 2-0 on Lee Sedol, one of the top Go players alive. This is the first time a computer has beat a human champion without a handicap.
While there will be three more games between Sedol and AlphaGo, the victory suggests that Go is the latest game that computers are outwitting people in. Checkers fell in 1994, Chess in 1997, and Jeopardy in 2011.
"It there's a social component with players playing together, it's not clear," Littman says.
"If it's a fixed set of rules, with players acting according to preexisting behaviors, then it becomes a sufficiently well-defined problems that you can optimize against and the machine will do better than people," he says. "If it's fluid and changing and contextual, it's much harder to say what it means to do well."
He gave the example of crossword puzzles, which are an interaction between the crossword puzzle maker. The maker can bend the rules of the game, like by letting the player put two letters in a square, and the crossword player.
A team of Littman's students helped make Proverb, a crossword playing AI, that played better than the average player but not as well as a champion. Littman recalls one puzzle that was bedeviling for the program - where all the clues were spoonerized. For instance, one clue was "home is near," meaning "Nome is here," so the answer was Alaska. (For more on Proverb, check this analysis.)
"In puzzles like Sudoku, where the rules are crisp, there's no issue. Computers crush human beings," he says. "But games like crossword puzzles, where the rules are fluid, then it seems like we find something that beats people."
The lines become even more clear when you extend the "games" to include sports. It's not hard to imagine a machine that could beat a human in shotput, Littman says; consider the nearest cannon.
But the more creative the sport is, the harder it is to mechanically replicate.
"With figure skating, you're trying to do something that creates an impact on the judges, surprises them with the artistry," he says. "The idea that we'd have a computer figure skater that beats the best humans is hard to imagine, but building a machine that beats a human in the hundred meter dash, that's obvious."