AlphaGo defeats Lee Sedol in first game of historic man vs machine match

Go players around the world were shocked on Wednesday, when the computer program AlphaGo defeated top professional Go player Lee Sedol 9p in the first game of the Google DeepMind Challenge Match.

AlphaGo-Lee-Sedol-Demis-Hassabis-Eric-Schmidt

From left: DeepMind CEO Demis Hassabis, professional 9 dan Lee Sedol, and Executive Chairman of Alphabet Inc (formerly Google) Eric Schmidt.

DeepMind challenged Lee Sedol to a five game match against AlphaGo in January 2016, after announcing that their artificial intelligence (AI) had defeated Fan Hui 2p 5–0 in a similar contest.

The match began on Wednesday, March 9, 2016, at the Four Seasons Hotel, in Seoul, Korea.

Lee is playing for one million dollars and, perhaps more importantly, the pride of countless humans around the world who don’t yet wish to see computers triumph in the ancient board game Go.

DeepMind, on the other hand, seek to test the abilities of their machine and make another step along the road towards a general purpose learning algorithm.

The match begins

Lee Sedol sat at the Go board with Google DeepMind programmer and 5 dan amateur Go player, Aja Huang, as a storm of cameras clicked away and the world’s media looked on.

AlphaGo-Lee-Sedol-Aja-Huang

Aja Huang (standing in for AlphaGo, left), and Lee Sedol 9 dan.

Aja stood in for AlphaGo, providing the physical presence necessary to place stones on the board and negotiate with Lee at the start of the game.

Lee Sedol drew black in the nigiri, meaning he would play first.

An unorthodox opening

Lee (playing Black) placed his first two moves on the 3-4 points in adjacent corners, while AlphaGo chose to play on the star points as White.

This set the initial tone of the game as a contest between Black’s territorial formation and White’s center influence.

AlphaGo-Lee-Sedol-first-move

Lee Sedol plays the first move of the match.

Everything was fairly standard until Lee played Black 7, which was quite unusual. It appeared that Lee had prepared a strange move to test AlphaGo’s strength in the opening.

Perhaps he also wanted to play an opening that AlphaGo was unlikely to have seen before, while crunching millions of positions as part of its training process.

Regardless, it wasn’t a good idea, because it didn’t lead to a good result for Black.

White’s play in the top right corner (from 11 to 14) appeared to be heavy at first, but Black 15 was also questionable.

AlphaGo allowed Black to take territory on the right side, while developing power in the center, and then took the initiative in the fighting with White 18.

The opening up to 18 was slightly better for White.

A difficult battle ensues

AlphaGo-Lee-Sedol-commentary-Michael-Redmond-Chris-Garlock

Michael Redmond 9 dan (right) and Chris Garlock provided live commentary of the game on Youtube.

Black 23 was an overplay, and White 24 to 26 was the correct sequence to punish Black.

Black 27 was also too much, and Black 39 was a bad exchange. The game proceeded smoothly for White up to 42.

Black resisted with 43, but the overall result up to White 50 still favored White.

Black’s group lived at the top, with 55, but White 58 was a sharp attachment.

Black 61 put up a powerful resistance, but White’s haengma from 62 to 76 was seamless.

Lee Sedol turns the tables

White 80 was slack, and Black started to catch up from 81 onwards. White should have defended the bottom left corner with a knight’s enclosure instead.

White’s play from 84 to 88 was also questionable, and the game was reversed up to Black 93.

White 102 was a sharp invasion, but White 106 was a bad exchange.

The trade up to Black 115 was satisfactory for Black.

AlphaGo-Lee-Sedol-Lee-Soyong-Kim-Seongryong

Lee Soyong and Kim Seongryong 9 dan comment the game in Korean.

Black stumbles

Black 119 was questionable. It would have been better to play Balck B5, White B4, Black D3, aiming to squeeze White’s corner.

Black 123 was another questionable move. He should have attached at R4.

The game became even again, up to 128, and Black 129 became the losing move. Black should have blocked off the corner at White 132 instead.

The game was reversed again by 136, and White consolidated his lead with White 150 and 154.

AlphaGo didn’t give Lee any chances afterwards.

AlphaGo-Lee-Sedol-media

The assembled media watches game one of the Google DeepMind Challenge Match.

Questions about Lee’s form

It looks like Lee wasn’t in his best form, and he might have been struggling with the pressure.

He made a couple of overplays in the opening, and AlphaGo’s responses were accurate and efficient.

Lee reversed the game due to White’s mistakes, but he wasn’t able to maintain his lead until the end.

As expected, AlphaGo’s endgame was excellent, and Lee didn’t have any chances after he lost the advantage in the bottom right corner.

Reactions to AlphaGo’s first win

DeepMind

DeepMind CEO Demis Hassabis was understandably excited:

AlphaGo-Lee-Sedol-Team-AlphaGo

Team AlphaGo sitting at the back of the room.

Other top pros

Lee Changho 9p said,  “I’m so shocked by AlphaGo’s play!”

Meanwhile Cho Hanseung 9p remarked, “AlphaGo is much stronger than before, when it played against Fan Hui 2p!”

“When Google said the odds were fifty-fifty, it seems they weren’t joking. I still can’t believe its performance even though I just saw it with my own eyes.”

Lee Sedol’s comments

In a post-game interview, Lee Sedol was visibly startled by AlphaGo’s strength.

“I was so surprised. Actually, I never imagined that I would lose. It’s so shocking.”

“Regarding the game, I got off to a bad start and AlphaGo played well right until the end.”

“Even when I was behind, I still didn’t imagine that I’d lose. I didn’t think that it would be able to play such an excellent game.”

“I heard that the DeepMind CEO Demis Hassabis said that he respects me as a Go player, but I have great respect for both of them [referring to Demis Hassabis and Eric Schmidt] for making this amazing program.”

“I also respect all the programmers who helped to make AlphaGo.”

The match continues

AlphaGo is off to an amazing start, but there are still four games to go.

Game two of the match is scheduled to take place tomorrow, Thursday March 10, and there will be a one day rest break on Friday before play resumes over the weekend.

Check our match schedule for details and visit the DeepMind AlphaGo vs Lee Sedol page for regular updates.

Game record

Lee Sedol vs AlphaGo – Game 1

 

Download SGF File (Go Game Record)

 

Younggil plans to post a full commentary of the game soon.

More photos

 

David Ormerod, with Younggil An 8p

Related Articles

About David Ormerod

David is a Go enthusiast who’s played the game for more than a decade. He likes learning, teaching, playing and writing about the game Go. He's taught thousands of people to play Go, both online and in person at schools, public Go demonstrations and Go clubs. David is a 5 dan amateur Go player who competed in the World Amateur Go Championships prior to starting Go Game Guru. He's also the editor of Go Game Guru.

You can follow Go Game Guru on Facebook, Twitter, Google+ and Youtube.

Comments

  1. Anonymous says:

    Hello, while awaiting the full commentary (I hope) could you tell us what white should have played on the move 80 instead the “slack” move?

    • Respond to Black’s approach on the lower left?

    • Anonymous says:

      ist written defending the lower left Corner with a Knights move

    • I wonder about it (move 80), too. Doesn’t it close the area in the upper left? Doesn’t it eliminate a bad aji? Doesn’t it mean A LOT of points?
      I know that not answering in the lower left corner cost white a lot of points in that area as black was able to make a large territory there. But what’s the math here? How many points is it playing on G13 or C6 for white and black respectively?

      • Younggil An says:

        White 80 wasn’t urgent because Black didn’t have any special move around that area at the time. White could answer the bottom left, and she will be able to return at 80 any time later.

        Generally, double approach is urgent and valuable than any other normal moves, and Black reversed the game due to White 80 in the game.

  2. Ortenix says:

    So by how many points did white lead in the end?

  3. GigaGerard says:

    Ortenix my question too.
    I am searching the web and the fact that nobody dare calculate the end difference in this game seems to suggest they all have an awesome respect for the human champion that clouds their judgement.
    Or perhaps a Go player can play agressively to change the odds up to a certain point where all is lost, so the final counting would not reflect a strength difference then… What is happening here?

    • someoneelse says:

      I (EGD 3dan) counted something between w+2,5 and w+5,5; probably w+3,5 or w+4,5

      • Younggil An says:

        Yes, right. Thanks both of you for your question and answer.

        White might be winning by 6.5 points in the end.

    • The game ended with a resignation, so it does not have a final settled score, and we cannot really calculate what it would have been — there are some points at stake still, depending on the play. It appears that Black is slightly ahead on the board, but not ahead by 8 points which he would need to win, and has no way to increase his lead to 8 points without some serious mistakes by the opponent.

  4. Could Black 163 have been a much deeper invasion of white’s top left territory, or does black have to worry about the cut at E7?

    • Anonymous says:

      Black 163 is not an invasion in white’s area. It is closing off the black area. Black could probably blay something like F11 to reduce white more but then white could also reduce the black territory. Even if the cut at E7 doesn’t directly work white could probably reduce black’s territory more because of the aji there.

      • That’s what I’m saying, Seedol played to close off his own area, but to me it appears at first that black’s reduction against the top left would be bigger than any white’s reduction of black’s territory, so could it have been an reduction _instead_? But yes, I think I underestimated the overall aji.

        • Younggil An says:

          Thanks Bacchus. You’re right that Black didn’t try to reduce White’s territory because of the weakness at E7.

          If White goes out at 163, it’ll be difficult for Black to stop.

  5. Crackerjack says:

    AlphaGo has grabbed the four corners, and that usually a winning game…

  6. If you count the points on the table it looks like black has won..but he resin

  7. What was the unexpected move that “no human would play” that threw Lee? He mentioned it in the press inference but I couldn’t figure out what he was referring to.

    • *press conference

    • bobiscool says:

      One of them is probably the first cut at the beginning, the second one is possibly (note possibly is less probable than probably) the move on the right side that was separating black’s stone and 4th line group.

      Although Zhou Ruiyang 9p predicted that move, all the other pros laughed at it before it was actually played

    • Anonymous says:

      White 102? Lee Sedol froze for quite a while after that move, based on the replay.

      • David Britt says:

        Apparently Ke Jie was playing that variation in the commentary before alpha go played it. Still could be what Lee was referring to, but I think it was the early peep and cut.

    • Younggil An says:

      Yes, the move Lee mentioned was White 102.

      Lee didn’t expect that AlphaGo could invade that deeply, and Lee was mentally shocked by that.

      Lee made a few mistakes on the right side and the bottom right corner afterwards, and it seemed as if Lee didn’t maintain his calm and concentration power after White 102.

  8. I wonder if Garry Kasparov has reached out to Lee.

  9. As i said to previous comment and even Top Pros don’t realize Alphago will play for the win and not to “humiliate” his opponent.
    If there is move A with 40+ points victory and 95,1% probability of winning and move B with o.5 point victory and 95,2% probability of winning Alphago will select move B even if from a human perspective the A is the best move and B very Bad.
    I think Alphago style reminds Lee Changho on Steroids letting his opponents go to an endgame were they thought there were pretty closed when actually Changho knew he was winning
    (even by half point) long before endgame take place.

    • Kari Antero says:

      That’s actually comforting though in my mind. I’d hate if the bot would try to win as much as it can, instead of as sure as it can. This can teach us a lot about the game.

    • Younggil An says:

      Thanks Billy for your comment. Yes, that’s very impressive and I understand the point.

      The commentators would have also learned about the system from yesterday’s game I think.

  10. Ellery Bann says:

    Google found an old blood-stained goban in a Japanese flea market and just now encapsulated Fujiwara-no-Sai’s essence inside a “neural net” computer… just saying!

  11. Davide Ferreira says:

    SAI = Simply Artificial Inteligence.

    And he has learned modern joseki 😉

  12. Hi.
    I ve seen alphaGo was running on 1202cpu and 176 gpu.
    Could someone give me a better understable comparison please?
    As i m not really knowleageable in informatic stuff xd.
    Thanks

    • CPU is Central Processing Unit and means processor. But it’s generally somewhat vague if it means just processor core or a chip (today’s processor chips contain up to tens of processor cores). Your laptop or mobile phone has one processor chip with a few cores; high end desktop would have few tens cores in 1 or 2 chips. Servers can have more chips. Core can execute 1 or 2 (depending on design; there were few designs with more, but they’re infrequent) chains of machine instructions (called threads) at any given moment (it can alternate between tens of thousands but only 1 or 2 are executed at once). Instructions are basic operations like add a number to another, compare values, skip to another instruction depending on a previous comparison result, etc. One core can execute ~10 billion such instructions per second. CPUs execute general programs, they are good at dealing with complex dependencies between instructions, programs can be big, they can alternated between tens of thousands of threads easily.

      GPU is Graphics Processing Unit or sometimes also called General Processing Unit. It’s primary use is to render graphics (mostly in games). But rendering realistic graphics requires quite a lot of non-trivial computation and non-fixed function GPUs *(i.e. almost everything produced int the last 10+ years) can do stuff non-related to graphics at all. First major non graphical use was providing physics simulation to 3D Game environments, but it can do more, like run artificial neural networks which evaluate Go positions or choose moves to investigate. GPUs are less flexible than CPUs (it’s program code memory is limited, if it’s operations tightly depend on each other results it won’t be fast, it’s much harder time alternating between many different programs quickly), but can do large amount of computations in parallel (few trillion operations per second) if there are akin to parallelization at instruction level. Neural Network computations are.

      Typically CPU is an executive and GPU a “physical worker”. GPUs will crunch a lot of numbers fast but would struggle at complex managing of total application flow. CPUs are more flexible but the number of math operations they can do at once is limited.

  13. “Go players around the world were shocked on Wednesday, …”
    Not necessarily. I wasn’t, he he! 😉 As I wrote in comments to the 5th game with Fan Hui: ‘We all still remember how Kasparov was flabbergasted after losing to Deep Blue, only shaking his head after not being able to match its skills and its calculation. I WOULDN’T be surprised if we saw something similar after the AlphaGo – Lee Sedol match!’ Well, it’s just the first game, but still …. it says a lot about AlphaGo’s strength.

    “It looks like Lee wasn’t in his best form, and he might have been struggling with the pressure.”
    Well, it’s exactly the same what people thought after the Kasparov – Deep Blue match! People then said perhaps he wasn’t in the best form, or he didn’t choose the best tactic against Deep Blue. Perhaps it’s difficult to choose the right tactic, if you face a stronger opponent… Perhaps it looks like you’re making unusually high number of mistakes, if you face an opponent that is able to punish them no problem.

    “He made a couple of overplays in the opening, and AlphaGo’s responses were accurate and efficient.”
    Didn’t Lee Sedol play the overplay moves ON PURPOSE?! Such an experienced player wouldn’t make “a couple of overplays” just for nothing, would he? It looks more like testing AlphaGo’s skills to me. But AlphaGo wasn’t mistaken…

    Let’s see if Lee chooses a more solid approach in the next games. But once again, I wouldn’t be surprised if he said after the match the same as Fan Hui said: “It was like playing against a wall.”

    Whatever the overall result of the match will be, this is historic! Wow!

  14. Anonymous says:

    Ok. Ready, set,…….. Go.

  15. Lee Sedol spoke about a winning move by AlphaGo that a humen player would never make, and that he was impressed by it. Which move was this?

    Kind regards,
    Paul

  16. I think Lee Sedol’s bad move was his last — resigning.
    I claim he should keep on playing to beyond where humans would normally agree to stop. Because he shouldn’t think of himself 100% as a “go player,” he should think of himself as 50% a “software tester.” And if you want to be a good software tester for some highly experimental, hairy, ultra-parallel system, which has been very little tested by anybody outside google, then you don’t just assume it must be fine. That is not the way to reveal bugs. At any moment alphago could go nuts and just commit suicide. I’m not saying its likely, but the history of human-vs-computer and computer-vs-computer tourneys has shown that computers have often found ways to die in very very embarrassingly stupid manner. Especially highly experimental leading edge computers. As a software tester, your attitude has to be “prove it. Prove you can go all the way to the end without screwing up.” Not “wow, seems great so far, I’ll assume it must be good.”
    I mean, if Lee Sedol was going to lose anyhow, why not keep going just to see? In Kasparov vs Deep Blue, K actually resigned in a position where he had a forced drawing line, because he ASSUMED
    that since the computer had been outplaying him all game it “must” not have missed that line. Actually, it had not seen it and perpetual check draws were a known tough issue for a lot of computers.
    (In a later game K blundered in a book position into a book trap loss, because he’d decided to intentionally play an opening he was unfamiliar with…)

    • Anonymous says:

      Could not agree more, I do not know the habits of Go professionals, but to resign in such a close game against a completely unknown system, is quite unprofessional. I had the impression that also the stream commentators were very surprised that Sedol resigned…

    • Endgame should be the strength of the computer, but you’re right, there could be bugs. On the other hand, it’s not without cost for Lee to play out a presumably lost game, especially in games 1 and 3, when the time between games is shortest. He needs to pick his spot to bug-check Alpha Go–probably game 5 if he somehow gets to 2-2.

      We haven’t seen Alpha Go lose, so I wonder if it’s programmed to resign, and under what circumstances. A strong machine that won’t resign would be a nightmare in go–much worse than chess.

  17. For me, the white sequence from 60 to 78 was the most revealing. That kind of smooth development of position from the cut was something that looked very “human” and sophisticated. I’m not shocked to see a computer out-calculate Lee, but this is play with a sense of position and balance, which is really something.

  18. were i in charge of publicity for Google, i would ensure that the human champ lost the first game so as to ramp up public excitement for the ensuing 4 games. hard to imagine that Lee would throw the match, even if both An and Kim saw several mistakes in his play. but the thing that strikes me most is the first picture in your article, which shows an unusually high table compared to the height of the players’ chairs. Lee is not tall, so when he is seated comfortably in the deep armchair, his eyeline is just about at board level! even when he leans forward to play, you can see his view of the board is at a narrow angle, whereas in Japanese pro tournaments the board is well below the players’ eyelines so they get the necessary birds-eye view. was this a setup?

    • Younggil An says:

      Oh, that’s interesting. I didn’t pay attention about the height of the table, but it doesn’t seem too high. If Lee felt uncomfortable, he must have asked to set up changed.

  19. Anonymous says:

    white 68 creates a heavy empty triangle; would not G12 have been more fluid and supple? black’s centre group was thick and out at G10 and working well with M6 group so surely white should focus on escaping quickly and ensuring there was no aji left in G14?

    • Younggil An says:

      Yes, right, but White wanted to take sente, so she played at 68, and then pushed through from 70 to 76.

      As we can see, White 68 was to take sente with earning one more liberty.

  20. Anonymous says:

    Kim criticised black 26; it seems to go against all the principles of sabaki. Was Lee hoping to cut at P15? Redmond showed that black cannot cut there. Does An have a better move for black than M15? For example, since the cut at P15 does not work, black could counter-peep at O15 and then live on the upper side.

    • Younggil An says:

      That’s an interesting question.

      I have to study the game again later, but it seems like Black 25 was necessary. Even if peeping at O15 is sente, Black should still come back to 25.

      I guess Black could have better options for 7 and 15.

  21. Anthony Cheng says:

    Here’s Ke Jie commentary: https://www.youtube.com/watch?v=IAAY532UQZA

  22. Since the computer already beat the human for the prize, I would like to suggest that the fourth match use the following format: The allotment of time for each player’s move should be 45 minutes for the DeepMind, and 3 hours 15 minutes for Lee. This is to equalize the great difference in speed of human vs. machine thinking. The fifth match allotment should be 45 minutes -/+ delta for DeepMind and 3 hours 15 minutes +/- delta for Lee.

  23. Dam who is dat female in red with nice legs i wanna lick it up n down