Zen computer Go program beats Takemiya Masaki with just 4 stones!

As part of the ’6th E&C Symposium’ in Japan, Japanese pro, Takemiya Masaki 9p played two games against the computer Go program Zen (aka Zen19).

Takemiya Masaki picture

Takemiya Masaki (9 dan).

Much anticipated by both Go players and AI experts, it was an opportunity for Zen to flex its muscles against a world class professional, though many still expected Takemiya to win.

Both games were played on March 17, 2012, on a 19×19 board. Regular readers might recall that Zen played John Tromp in the ‘Man vs Machine match‘ earlier this year.

Zen wins two games

In the first game Zen received a five stone handicap and won by 11 points. After that the handicap was reduced to four stones, but Zen surprised many by winning again, by 20 points this time!

Zen has previously defeated professional players on five and six stone handicaps, but this is the first time it’s won a match against a pro with only four stones. What’s more, Takemiya is not just any pro. He’s a well known former international champion.

Playing the percentages

I have to admit that I found the results quite surprising, especially the win with four stones. Removing the 5th stone at tengen makes a big difference to the type of game that develops in my opinion, and usually black has to fight a lot more.

What we saw from Zen though was something different. It did fight, and it plays quite well of course, but it was also willing to accept many small losses that I don’t think most human players at this level would.

Time and again it backed down, giving Takemiya what he wanted, but also taking a certain amount of compensation.

Losing points, but maintaining the lead.

In that sense, its positional judgement was really impressive and it reminded me of what Martin Müller said when I interviewed him about Computer Go recently:

The programs are generally good in overall balance and counting. They know what it takes to win and will not lose quietly or be overly aggressive when they are ahead.

Takemiya Masaki vs Zen 300x197 picture

Takemiya Masaki giving Zen with a four stone handicap.

About Zen

Zen was written by programmer Yoji Ojima and ran on hardware provided by Kato Hideki, of team DeepZen.

Match details

According to Hideki, the hardware for this match was a mini-cluster of four PCs (a dual 6-core Xeon X5680/4.2 GHz, a 6-core Xeon W3680/4 GHz and two 4-core i7 920/3.5 GHz) connected via a GbE LAN. This is the same hardware used by Zen’s ‘zen19s’ and ‘zen19d’ accounts on the KGS Go Server.

Both of the games were played with 30 minutes main time and 60 seconds byo-yomi. Zen is currently ranked 5 dan when playing under similar time conditions on KGS.

Earlier in the day, another pro, Ohashi Hirofumi 5p played two even games on 9×9 against Zen. The result was one win each.

Will Zen’s march continue?

Winning against a pro with four stones is very impressive, under any circumstances, and shows how far computers have come in Go.

However, it’s clear that Zen was able to win these games by avoiding fighting to a certain extent and relying on its excellent positional judgement. It only played to maintain enough of the handicap advantage to win.

That’s sensible, but it raises the question of whether Zen and other programs will continue to improve steadily as the handicap is reduced and they’re forced to play a more risky style.

Do you expect the current programs to continue improving steadily as hardware gets better, or do you think they’ll plateau at some point?

Let us know what you think by leaving a comment below.

Game records

Zen vs Takemiya – five stones

Download SGF File (Go Game Record)

Zen vs Takemiya – four stones

Download SGF File (Go Game Record)

About David Ormerod

David likes teaching, learning, playing and writing about the game Go. He's taught hundreds of people to play Go, including many children at schools in Australia. In 2010 David was the Australian representative at the 31st World Amateur Go Championships. He's a 5 dan amateur Go player and is the editor of Go Game Guru. You can find David on Google+ and follow Go Game Guru on Facebook, Google+ or Twitter.

Comments

  1. It is a matter of wishing rather than thinking but I hope they keep getting better. Nice to have info on the hardware, I wonder how well it would perform on everyone’s computer?

  2. latinpower says:

    Zen19D is rated as 6d now on KGS

    • David Ormerod says:

      In very fast games Zen is 6d, but in slightly slower games (like this one) Zen is still 5d (see the Zen19S account). It’s worth noting too that this is basically a blitz game by professional standards.

  3. Anonymous says:

    well, the pro got beaten in a speed GO due to time constraint…
    if they play in normal speed, i believe the player will perform better….

    • David Ormerod says:

      You may well be right. If this is any indication of what’s to come, we’ll start seeing more serious games between computers and top human players over the next few years and it would make sense for those to be slower games, like in a serious tournament. We’ll see :) .

  4. Interesting games. In both games black sacrified a not too small group to grab other territory or additional influence: I’m not quite sure normal amateurs would do that. You see that in pro games too: they spend moves that at a certain moment will be used as bait to accomplish something else. I wonder whether the pro’s playing the computer may need a few games to accustom their attitude, not to take the computer lightly (not saying Takemiya did, though). As always I am pro progress, it would great to see the computer play at least at the level of the top pro’s. I wonder how games will look like in those days: the rough-and-tumble style or the quiet style, Lee Sedol or Ishida Yoshio?

    Kind regards,
    Paul

    • David Ormerod says:

      You make a good point Paul. At the moment computers are still very strong in some areas and weak in others. Pros are a lot closer to having a ‘complete game’. If it were a ten game match the human player would be able to learn the computer’s weaknesses and exploit them mercilessly, at least at the moment. It will be very interesting to see what happens though.

      Also, I think players like Lee Sedol would fare well against computers because of their ability to complicate the game and increase the chance that the computer will make a slack move or an overplay. The Koreans have a word for describing an ability that players like Lee Sedol and Cho Hunhyun have. It roughly translates to ‘shake’; meaning that they can shake the game to create opportunities when they find themselves behind and it can be very hard to win a ‘won game’ against this kind of player. I sort of think of it as the equivalent of tilting a pinball machine :) .

  5. in the 4 stone game: move 183???? Someone please explain this!

    • If black tenuki, then white can cut the bamboo connection at E10. Black cannot resist because we would be caught in a shortage of liberties.

    • Anonymous says:

      F9 or E9 can capture outside group

    • Take a look at the board …..

    • The aim of move 183 was to save 3 white stones around d8 from taken by sente. The rescue of 3 stones is, if gote, 6 points in Japanese counting. As this is to prevent sente, which is called gyaku-sente, the maximum value is equivalent to 12 points in gote. The real value of a gyakusente is determined by the total value of the remaining moves, divided by the number of the moves. This calculation is beyond human ability, thus we assume it’s almost double of normal gote, but sometimes it’s smaller than that.

      • no, the aim of 183 is to save the 13 stones. playing at the bamboo joint is atari then when white connects is 2 libs to 2 and blacks turn. So its much much bigger than just saving 3 stones.

  6. If the self aware super-intelligence will exist, you could teach it the rules and it will defeat any pro. As Lasker said, alien minds must have discovered Go, and this alien mind will go beyond of what we CAN imagine.

    I haven’t seen an update of an article of this kind:
    http://www.nytimes.com/2002/08/01/technology/in-an-ancient-game-computing-s-future.html?src=pm

    Anyways, as happened will Deep Blue, the brute force approach of improving algorithms and hardware will eventually create a 9d Go program.

    • David Ormerod says:

      Nice article pore. It is about time for an update.

      As someone who tries hard to promote Go, I also think of these computer Go matches as an opportunity to introduce Go to new people, who might not hear of it otherwise. One way of doing that is through major news stories.

  7. MarcoRosso says:

    It will be interesting. I don’t think that it will be able to be run on personal computers though. As for the positional judgement computer AI’s are pretty good. Though raw fighting seems to be lacking. Until this is improved I doubt we will see a 9pd Go AI.

  8. Tommyray says:

    Playing and winning only with 4 handicap stones against Takemiya Sensei is impressive, but from other perspective it is very high handicap which means Zen is about amateur 3-4 dan at his best. One handicap is not a one stone in strenght difference, it is much more. I’m 4k but to have even game and win with 4d I need just 2 stones – does it mean I am 2d? Certainly not, it just clearly shows how handicap stones change the game.

    • I found that the handicap system works remarkably well and can not agree with your statement. Also, the KGS rating system that Zen’s rank is based on takes stones to be equal to rank difference up to H6. If you can consistently beat 4d players with 2 stones 50% of the time, you are certainly much much stronger than 4k…

      • Agreeing with Nik there, Tommyray’s case is certainly very exceptional. However, Tommyray is still right with the 3-4d statement. KGS 5-6d should not be confused with the corresponding amateur levels, so Zen is still “just” mid dan level. Nontheless I believe (and hope as a programmer and am scared as a player) that Go programs will get to about pro level in my life time.

        Regarding the games shown, I do not consider this kind of exchanges to be as extraordinary as stated. Not saying I’d be able to calculate it through (or that is was intentional) but this happens in my games all the time (I’m KGS 3d) and I rarely consider a game lost only because a large corner died as long as I get something in return – especially in handicap games.

        In fact I remember a book about handicap games that reminded how usual rules do not apply in handicap games. For instance, pushing on the 4th line instead of the 3rd line can be fine.

        • There are several observations to make:
          - it was 1 game with 4H
          - Takemiya may not be a top pro anymore today
          - top pro performance needs top pro conditions
          - amateur 4d is a decent rank but to beat top pros with 4H a stronger rank is needed (5-6d I believe)

          Avoiding any further belittling of Takemiya, whom I admire, we would need 10 games by Zen against the likes of Choi Cheolhan, Park Junghwan, Gu li and Xie He, sponsored with some decent prize money and with varying time limitations. If under such conditions Zen gets 40-60% winning percentage with 4H, I’d be positive about Zen being 5-6d.

          There is no doubt in my mind that this day and better days will come for a computer program. But I’m also sure that in due time humans will boost their capacity with artificial brain equipment (built in chips, you name it) and the world of Go will face similar issues as cycling and athletics are facing today.

        • David Ormerod says:

          Mafutrct, I wasn’t talking about sacrificing stones in particular. I was referring to the feeling of consistently slack play, yet with the ability to turn it into a win.

          My experience has been that playing that way in a low (4 stones or less) handicap game against a pro doesn’t work well. I think that’s because white usually has far superior counting and endgame skills. Which means (usually) the way to maintain a low handicap advantage is to use it to keep the pressure on and gain as much as possible earlier in the game. As a game goes on the handicap advantage inevitably becomes smaller in relation to the number of stones on the board.

          Zen’s approach was different, and it worked in this case. And I wouldn’t be surprised if Takemiya felt like he was on course to catch up, based on his experience with normal human players at four stones. I think computers can confuse human players, even pros, by playing more slackly when they’re given a big handicap. It’s not a normal experience to play someone who gets stronger like that as the game progresses.

          I remember awhile back there was a game where Mogo suddenly killed a big group in a 6 or 7 stone game against a pro. To me it looked like the group should have been fine, but white played in a leisurely manner and underestimated the threat. I thought it was possibly due to underestimating an opponent who had played very slackly up until then.

          If this is going to be a regular occurrence, I think pros will soon learn about the difference between playing a human and a computer on the same handicap. You can be sure they’ll come up with some interesting strategies for dealing with these new opponents too :) .

          • I wonder if what we’re seeing here is that this computer program understands its own weaknesses.

            Most people try to target top level openings from the beginning. However, even if they are given a good opening position – say, through handicap stones – their tactical play is not good enough against a pro to turn a won strategic position into a won game. Once they get into the tactics, the pro creates live groups where with perfect play he should not be able to, and the mid ranked player loses groups that should have been alive.

            I think what we’re seeing here is that Zen, in a sense, “realizes” this, and realizes that it needs a higher density of friendly stones in the opening to allow its lower level of tactical play still to create live groups and win fights.

            Thus, what you see as “slack play” is not really slack play – it’s purposely overconservative play in the opening to create positions where it will be able to hang on to in the end game.

            • David Ormerod says:

              That’s an interesting theory Warren.

              It’s a plausible way of playing for a player that has a high degree of confidence in their ability to judge what ‘enough’ is.

              The average human player wouldn’t be confident enough to pursue that kind of strategy, but I suppose to the algorithm it’s neither here nor there. As long as there’s still a high probability of winning, it can play that way.

    • Handicaps can be deceiving. Everybody has a lower handicap against the person who taught them then they do against the rest of the world. I think that as the ranks spread, we will stay in the range 1 stone for every level to 1 stone for every two levels. Remember that the first stone is only half so the determining factor is how many almost even games are played.

    • Playing and winning only with 4 handicap stones against Takemiya Sensei is impressive, but from other perspective it is very high handicap which means Zen is about amateur 3-4 dan at his best. One handicap is not a one stone in strenght difference, it is much more. I’m 4k but to have even game and win with 4d I need just 2 stones – does it mean I am 2d? Certainly not, it just clearly shows how handicap stones change the game.Playing and winning only with 4 handicap stones against Takemiya Sensei is impressive, but from other perspective it is very high handicap which means Zen is about amateur 3-4 dan at his best. One handicap is not a one stone in strenght difference, it is much more. I’m 4k but to have even game and win with 4d I need just 2 stones – does it mean I am 2d? Certainly not, it just clearly shows how handicap stones change the game. – I disagree on this, I’m around 1d on kgs and I may still lose a fully even game to ie a 3k, but I have also won several games against 3k giving 4-6h. I generally have a very high win percentage when I give 2-6h, but I am pretty sure I would still lose most games against Zen if I received 4h.

  9. A couple thoughts:

    1. It’s possible that Zen’s style lends itself very well to the use of handicap stones. This would give it an advantage over a player of similar strength who, for example, has strengths in territorial games or even fighting games. Being able to judge the position and understand the value of thickness are concepts that might be more difficult for humans at this level.

    2. 6D KGS is definitely stronger than amateur 3-4D, and I’m not even sure what ranking system would align the two this way. (For example, KGS 6D is about AGA 6D and probably even 7D in Japan). A 4 stone game at this level should of course be a challenge for a pro. This is especially true if, as Dieter pointed out, there is no real incentive for the pro to win.

    3. I believe that as Zen continues to get stronger, its style will remain the same. I think Zen can continue to rely on it’s apparent strategy of positional judgement and whole-board sacrifice even in even games. This could be very interesting as it would potentially teach humans to play differently. (This is all of course very hypothetical and relying on the assumption that Zen will eventually progress to and past professional level).

    • I was unclear. I was specifically talking about European amateur level, where 4d roughly corresponds to KGS 5-6d. An EGF 3~4d is considered mid dan level since EGF 7 dan is already about pro level. Thus my comment how Zen is “only mid dan level despite being 6d”. Hope that cleared things up.

      I find the comment about “lends itself very well to the use of handicap stones” interesting. Counting and judging the position is inherently easy for Zen-like programs, they do not require some time (like even a professional does) to properly read and count because they “do this all the time”.

      On the other hand, a human black in handicap games tends to not account for the global position as if it were an even game, even at high level. Or am I wrong here? I believe black likes to “trust” white to know a little better and follows with one eye closed. Not intentional, but simply due to human psychology that cannot be switched off. Thus one could in fact say computers have an edge over humans (even pros) in this regard, making them use handicap stones better.

      Regarding this whole topic, we have to keep in mind also that handicap is (to my knowledge) a terribly under-reasearched topic. The difference in stones can vary greatly depending on what system you use. For instance, according to KGS, I’ll take 6 stones versus a pro, but according to EGR it’s only 5. Neither seems wrong to me, and it works in the other direction (me playing white vs kyu players), too.

      What I’m saying is it could very well be that human handicap is too large typically, to adjust for the psychological disadvantage of black. Meaning that computers should only get (completely random number) 80% of the stones a human of equivalent strength should get in a certain handicap situation. Or in other words, humans get a little more to account for their human weakness.

      Not sure if that is too far-fetched a theory.

      • Concerning your thoughts on handicap, I believe that there are two ways to think about handicap. We’ve always learnt as amateurs to use the handicap stones for attack. However, I think we all have had the experience that it is much easier to use the handicap stones in a territorial manner. As White too, I’ve found it easier to overcome Black handicap supported aggression, due to the difference in fighting abilities, than to catch up with 4H where Black calmly applies a tenuki strategy.

        However, the adagio of “use handicap stones to attack” is not a winning strategy but a learning strategy. Beginners, especially those who learnt with territory as a core concept, as opposed to those bred with the core idea of alive stones, have a tendency to stress territory but are relatively weak in attacking. Handicap stones can help in improving that ability.

        As a winning strategy though, the calm style displayed by Zen may work better than a focus on attack and this confirms my experience with club handicap games.

        Still, it seems to indicate that Zen is relatively stronger at positional judgment than at tactics, which runs counter my intuition on computer programs.

        • David Ormerod says:

          Interesting points about handicap games. Generally the player who wins high handicap games (e.g. 6-9 stones) by ignoring white and trying to consolidate another part of the board doesn’t fare as well once the handicap is reduced. Possibly it’s because they’ve learned less about how to fight, as Dieter says, or maybe because this strategy just becomes less effective as the handicap is reduced.

          Philip’s third point about using the same strategy in an even game is intriguing. My gut feeling is that that won’t work very well – it will have to play differently – against a top player, but I’m just speculating too :) . I’d like to see us learn some new strategies from computers though. That would enrich the game.

          In my opinion, top pros are actually a lot stronger than many people think. And, at least for a human, those last few stones between a pro and a reasonably strong amateur are huge. We’ll see if it’s the same for computers.

  10. terrymac says:

    Will computer go plateau? Yes, but not for long. These games represent something new compared to a year ago; previously, programs did rather badly at handicap go. Go programmers are a highly competitive lot; when their programs plateau, they look for a way to break through to a new level.

  11. I think computer player strength will continue to improve with processing power. However, processing power may plateau.

    Chess programs went from laughable to world championship status during the decades of steadily increasing processor speed. As processor power increased exponentially, the FIDE rating of the strongest chess playing programs increased linearly. Compared to this driver, improvements in the quality of the software were negligible.

    In 2005, processor speed improvement came to pretty much a screeching halt, but processor power in high end computers or clusters continued to increase, approximately exponentially, through increased numbers of processor cores running in parallel.

    Monte carlo methods are well suited to distribution over multiple processors, which is why their application has been so successful here: in this case, they permit the computer player to take advantage of 26 cores in parallel, instead of just one, even if the fundamental programming doesn’t include any new magic.

    Extrapolating from Chess experience, this should meant that computer go should become competitive with top human players when the computer opponents are harnessing somewhere between 1000 and 1,000,000 processor cores in parallel, with a best guess perhaps around 26,000 cores. Above the top of that range – with tens of millions of processor cores in parallel – they should be able to beat human opponents handily.

    Of course, processor speeds eventually stopped increasing. We may eventually see processor parallelism stop increasing at some point as well, perhaps for energy consumption reasons or something like that. That may or may not happen before computers are able to beat the best human players handily.

Speak your mind