Lee Sedol took home the victory late last night/early this morning. He started the game very copying his opening moves from Game 2, expecting AlphaGo to follow suit, which it did. Lee didn't follow his old moves for very long, however. From reading some commentary on GoGameGuru, it sounds like he went for a tough all-or-nothing strategy instead of letting the field get divided up into lots of little battles that AlphaGo could weigh and evaluate well.
Around 78 moves in, he made a move that was really highly regarded by others, and 10 moves later, AlphaGo started making a bunch of moves that onlookers thought were very bad. Is it because there's an inherent weakness in algorithms that utilize Monte Carlo Tree Search? Or is it because AlphaGo couldn't properly see the strength in Lee's move at 78? If that move only works if a few other exact plays get made, then perhaps AlphaGo never tried those paths and didn't see what was coming.
I am really excited to see that Lee, representing Team Humans, managed to recognize and exploit a weakness in AlphaGo. Although he might not have another winning strategy to employ in the next game (as he'll be playing as Black instead of White) I still hope he can manage to push out a win.
A Domino-Covering Problem
1 month ago