Michael Redmond 9P has been providing the English game commentary for the Lee Sedol-AlphaGo match, with American Go E-Journal Managing Editor Chris Garlock. On Friday, the break day before the match resumes on Saturday with Game 3, Redmond sat down with Garlock to recap the first two games and look ahead to the critical third game on Saturday.
Game 1: Underestimating AlphaGo: Lee Sedol, taking black, played a really unusual opening in the first game. I
think he had a plan to try to throw the computer off course by playing an opening it didn’t have in its’ database. But I’ve been talking to the Deep Mind programmers and it’s not actually a database but machine learning that recognizes patterns, so AlphaGo doesn’t need to have an exact match and it doesn’t really matter if some of the moves are slightly different, so AlphaGo can judge the situation, which is quite different from the usual computer programs and closer to the human way of thinking. So Lee’s plan to play something that AlphaGo hadn’t seen before just didn’t work. And in fact, Black 27 was an overplay and Lee got in trouble himself and I think it’s because he was playing an opening that he really wasn’t familiar with and probably underestimating AlphaGo. Even though he mostly recovered from that mistake, White 102 clearly took Lee Sedol by surprise. It’s a move I was expecting, because White had to do something like that, although of course you need the reading to back it up. It’s a move Lee Sedol should have been looking at and I think maybe one of the reasons he might not have been looking at it is that maybe he underestimated AlphaGo and wasn’t treating it like a top human player. In this game, I didn’t find anything I would call original in AlphaGo’s play, it was just plainly strong. Click here for Redmond and Garlock’s Match 1 15 min Summary.
Game 2: AlphaGo’s Exquisite Game: In the second game, AlphaGo had black, and I was looking to see how it would play. In the games with Fan Hui, I wasn’t impressed with AlphaGo’s opening; it was really too orthodox, and too simple for a game with black, because of the big komi. Black needs to play more aggressively or more of a speed-oriented game. So I was interested to see how AlphaGo had changed since last October. So it was rewarding to see that AlphaGo was playing moves that were not conventional and they were speed-oriented and putting pressure on Lee Sedol from the start of the game. And then there was the shoulder-hit at move 37, which was a move that really took me by surprise and I’m pretty sure it took Lee Sedol by surprise too. One of the programmers dug into the files and found that the possibility of playing that move was something like one in ten thousand so it was a really unlikely move but it happened to be on the edge of AlphaGo’s search and in the analysis of the possible variations, AlphaGo decided to knock it up to the move that it would choose. I still don’t really understand the mechanics of it but it’s really interesting. Not only does AlphaGo have good shape or pattern-matching, but it can also think itself out of that and find something completely different, even though it might not have had a good score in the pattern analysis. That game on the whole was an exquisite game by AlphaGo. I actually thought that Lee Sedol had a pretty good chance to win up until the middle game and the game was sort of in the balance for a while there, but then almost before you notice it, AlphaGo had very subtly built up an advantage after Lee Sedol played a few slack moves. In this game, Lee Sedol took territory in the beginning, allowing AlphaGo to sort of dance around the board and take the initiative, and that’s not really typical of Lee Sedol, so once again he wasn’t really playing his own style. Click here for Redmond and Garlock’s Match 2 90-second Summary.
Game 3: Play Your Own Game
If I were Lee Sedol I’d just play an opening I’m familiar with and I wouldn’t worry about whether AlphaGo knows the moves or not and then he should continue into the middle game along a familiar path in which case he’ll be more at home and less likely to make mistakes. When Lee Sedol plays well he’s brilliant, of course. If he plays his own game I think he’s more likely to get into the middle game with a favorable position, which is really necessary. He hasn’t done that in these first two games; he’s been going into the middle game in a fairly difficult position. If Lee Sedol has a favorable position in the middle game, that’ll be something new for AlphaGo, something we haven’t seen yet. Just as a test of the program itself, that’s something I really want to see how it handles that.
The third game in the 5-game Lee Sedol-AlphaGo match will be Saturday, March 12, (Friday night 8p PST, 11p EST). The match will be livestreamed on DeepMind’s YouTube channel with commentary by Redmond and Garlock. And catch Cho Hyeyeon 9P’s commentary with Andrew Jackson starting at 9P PST on the AGA’s YouTube Channel.
photos: (top right) Lee Sedol plays the first move in the first game; (middle left) Lee at the press conference after the second game
