SEOUL, SOUTH KOREA — Go grandmaster Lee Sedol regained a sizeable slice of human pride on Sunday night when he won the latest game in his historic match with an artificially intelligent machine built by Google researchers. But on Tuesday, in the final game of this best-of-five series, he hopes to regain far more.
In one sense, the match is already lost. Google’s system, known as AlphaGo, won the match’s first three games, taking home a million-dollar prize and becoming the first machine to beat a top human at the ancient game of Go, a pastime that’s exponentially more complex than chess. Lee Sedol openly apologized to the Korean public and the wider Go community (and, perhaps, humans in general) for losing the match, tapping an undeniable melancholy among those gathered to watch the match inside Seoul’s Four Seasons hotel. But he completely reversed the mood in Game Four.
Game Five is, in a way, the last frontier.
When AlphaGo resigned nearly five hours into the game, the Korean press cheered. They cheered even louder when Lee Sedol walked into the post-game press conference. “Because I lost three matches and then was able to get one single win, this win is so valuable that I wouldn’t exchange it for anything in the world,” he said through an interpreter, fueling still more cheers. “That’s because of the cheers and the encouragement that you all have shown me.”
But another significant moment arrived at the very end of the press conference, when the Korean unexpectedly turned towards Demis Hassabis and David Silver, two researchers from DeepMind, the London-based Google AI lab that built AlphaGo, with an unexpected question.
Black Versus White
In Game Four, Lee Sedol had won playing the white stones while AlphaGo played black. In other words, AlphaGo made the first move, and he played second. Until changes to the rules of Go in the early 20th century, playing second was a disadvantage. But now, if you play second, you receive a sizable head start in points, and that disadvantage goes away. “It’s even,” says Andrew Jackson of the US Go Association, who has been broadcasting an online commentary during the match.
This week, however, playing second suited Lee Sedol. His best effort prior to Game Four was in Game Two, when he also played white. And during the press conference following Game Four, the Korean explicitly said that AlphaGo was weaker when it played first and he played second. “It struggled more when it was holding black,” he said.
For Game Five, under the official rules of the match, the two opponents were set to randomly choose who would play first and who would play second. But then came that moment at the end of the press conference following his victory in Game Four. Lee Sedol turned towards Hassabis and Silver and asked if he could play black in Game Five. To wit, he was asking for the bigger challenge. He was asking for the hurdle he still hasn’t cleared. “I really do hope I can win with black,” he said, “because winning with black is much more valuable.” Hassabis and Silver conferred—ever so briefly—and then granted his wish.
Some Added Spice
There was more applause from the international press. Granted, this applause was led by me. But, well, it was another wonderful moment from the Korean grandmaster. And it lends some added spice to the fifth and final game. This is no dead rubber. It’s not just about Lee Sedol clawing back to within one game. It’s about Lee Sedol showing that he can beat AlphaGo no matter which stone he holds. If he can do that, the machine’s match victory isn’t quite so complete.
Yes, if Lee Sedol wins, there will be talk of a rematch. But that will by no means favor the Korean. The trick with AlphaGo is that it’s powered by machine learning—technologies that allow machines to learn tasks on their own. Google’s creation beat European Go champion Fan Hui in a closed-door match this past October. After Hassabis, Silver, and their team continued to retrain the system over the past five months, its skill level rose significantly. Before a rematch, that level would rise yet again.
Game Five is, in a way, the last frontier. And it will by no means be easy for Lee Sedol. Yes, he now has the advantage of having watched AlphaGo play two games with the white stones. So he has more experience to draw from. And yes, the pressure to win the whole match is now off, as it was in Game Four. But clearly, AlphaGo is stronger when playing white.
Just before Game Three, I asked David Silver if AlphaGo played differently when it played one color as opposed to the other. “I think it’s hard to say,” he told me. “I would have to defer judgment to a pro player on that.” Though he has helped build a machine that plays Go at professional level, he is still an amateur and feels he can’t really judge the play of the machine.
Well, the best good pro player to defer to is Lee Sedol, who clearly thinks that AlphaGo struggles when playing black. And the Korean has chosen the opposite scenario for Game Five. That indeed deserves a cheer. But the cheers will be far louder if he can grab a win from this position of weakness. Whatever happens tonight, the bigger contest is by no means over.