Google DeepMind Challenge Match - Lee Sedol v AlphaGo - match report

Game 1 - 9th March 2016

You can play through the moves of the game, move by move, using the board to the left (use the right or left arrows to navigate through it). This may not appear in some browsers.

If you click on the board, rather than using the right arrow, it will register only when you play on the point where the next move is actually played!

Lee Sedol is Black, and AlphaGo is White.
Lee Sedol (right) plays a parallel fuseki with 3-4 points, while AlphaGo plays star points. During the live match, Aja Huang (left) of Google DeepMind transferred the moves from AlphaGo to and from the real board. Whereas in the match against Fan Hui AlphaGo played relatively calmly, in this match it started an early fight, playing much more aggressively. AlphaGo also took more time than Lee, taking more than 1 minute over each move.

Almost 100,000 people watched the Youtube channel - it peaked at 99,896 at 7:31 GMT. The downloadable game record is here.

Game 2 - 10th March

A very even game all the way through, with the commentators being uncertain as to whom was winning. However, Lee was more than 10 points behind and resigned.

Over 80,000 viewed online.

Comments from the Press Conference:

Lee Sedol: "Yesterday I was surprised by AlphaGo’s play, but today I’m quite speechless. At no time did I feel that I was leading, and I thought that AlphaGo played a near-perfect game." When asked what were AlphaGo’s weaknesses “Obviously I haven’t managed to find them out…” and when asked about his prediction for the third game said “It’s not going to be all that easy for me, but I’m going to exert my best efforts to win at least one game.. I’ll have to focus quite hard on the beginning of the game.”

Apparently the turning point was move 37 (when AlphaGo played a shoulder-hit on the RH side), a move that AlphaGo thought was good but "would never be played by a human player".

Demis Hassabis: “The team weren’t that confident towards the end of the game as the professional commentators kept changing their minds about who was winning, but AlphaGo’s estimate was increasingly confident that it was winning in the second half of the game."

Michael Redmond (9 dan): "In today’s game AlphaGo made some innovative and dangerous looking moves, but managed to make them work. I thought it played a beautiful game.”

The downloadable game record is here.

Game 3 - 12th March

Lee made a framework, inviting AlphaGo to approach his corner, so that he could attack. Unfortunately, AlphaGo played too accurately and made a special "elephant move" to extend his position, and eventually the attack petered out with much of Lee's framework destroyed. After this Lee was always struggling and eventually had to resign again.

At the post game press conference Demis Hassabis said he was both stunned and speechles after the incredible fight. He said it was amazing that Lee Sedol could keep up with AlphaGo which can evaluate 10000 positions per second.

Lee Sedol apologised for having not satisfied people's expectations. He under-estimated the program in game 1, missed chances in game 2 and thought the pressure too great in game 3.

Michael Redmond said it had been a powerful fighting game, true to Lee Sedol's style, and thought no professional could have beaten him. He thought AlphaGo would inspire a new type of opening style.

The downloadable game record is here.

Game 4 - 13th March

The first 11 moves were the same as game 2. AlphaGo as black made a large centre, which Lee Sedol cleverly found a way to take apart to make the game close.

According to Demis Hassabis, when interviewed on BBC Radio 4 Today, the turning point was move 78 (on the 11-11 point); Alpha Go had given a vanishingly small probability that Lee would play there. Towards the end of the game AlphaGo played some strange moves when it thought it was behind, seemingly to test Lee. However these lost points and, despite being in overtime, Lee played very accurately and eventually at about 08:45 GMT AlphaGo resigned.

At the Press Conference Lee was applauded by the journalists and said with a big smile: "It’s just one game. I’ve never been congratulated so much just because I won one game. This win is invaluable and I would not trade it for anything else in the world”.

Demis Hassabis said that the team would work out why AlphaGo made so many mistakes when they got to London.

The downloadable game record is here.

Game 5 - 15th March

The Korean TV channel KBS suspended normal broadcasting to show the final game.

Another interesting and even more exciting game. AlphaGo made a clear mistake in the bottom right hand corner, but Michael Redmond said that it was probably only worth 5 points or so, so not fatal.

Play continued with AlphaGo creating a large framework, which Lee Sedol needed to invade. He did so, but with some elegant play AlphaGo compressed Lee's group into just two eyes and created a decent sized territory out of the rest of the framework.

Lee then had to invade the bottom left corner, but didn't get much compensation out of this, so was still struggling albeit a little ahead.

The rest of the middle game was a knife-edged struggle, going to and fro, but eventually Lee didn't manage to make use of a cut towards the bottom and it looked like he made a mistake on the left which cost him points again.

The score was still very close, with both players in byo-yomi, but with Lee about 2.5 point behind he resigned as only half-point moves remained.

A sad finish to a close game, which allowed us to see AlphaGo's ability at yose.

Comments from the post-match Press Conference:

Demis Hassabis:" I'm kind of speechless. It was the most mind-blowing game experience so far. It was an incredibly close and tense finish and very stressful for us. The commentators said that AlphaGo made a mistake early on, but it managed to catch up." During the Q&A he said that they had no concrete plans for the development of AlphaGo, either by additional matches or by releasing a version to the public.

Lee Sedol said "I feel sorry because the Challenge Match has come to an end and that I lost." [The sound was partially lost during his comments.] "I don't feel that AlphaGo is necessarily better than me and that humans can do better, but I still enjoying playing Go. I have much studying to do down the road."

Chris Garlock (AGA commentator): "Thank you to the DeepMind team for this opportunity and for five fabulous historic games that we'll be studying for a long time."

The downloadable game record is here.