Can AlphaGo defeat Lee Sedol?

[The following is a guest article written by Ben Kloester.]

Nature-Go-game

The front cover of Nature, in late January, 2016.

The Go world was shocked and intrigued in January, when news broke of DeepMind AlphaGo’s victory over top European pro Fan Hui 2p.

A computer playing Go at professional level was seen by many as being at least a decade away, a view held even by experts like Rémi Coulom, author of the previous state-of-the-art AI Crazy Stone.

News of Fan’s defeat was met with awe, and in some cases even fear.

Since the publication of DeepMind’s paper [PDF] in Nature, and the release of the game records, professionals around the globe have had time to analyse AlphaGo’s play in more detail, and have drawn less sensational conclusions.

A consensus has emerged that although this is a great advance in computer Go ability, DeepMind would not be celebrating victory if it had been a top professional sitting across the Go board back in October.

Expert commentaries of Fan Hui vs AlphaGo, including Younggil’s SGF commentary and video commentary, Myungwan Kim’s epic video analysis, and even commentary from match referee Toby Manning [PDF] of the British Go Association (based partly on Fan’s comments), have all identified mistakes made by AlphaGo that a stronger player would have capitalized on.

Professionals have also pointed out some areas of perceived weakness for AlphaGo, and speculated on its potential limitations.

And let’s not forget that Fan actually won two of the ten matches played.

All this suggests the AI’s strength, as seen in October, is well below the top-100 ranked professionals.

Predictions Marching to different tunes

Lee-Sedol-at-AlphaGo-press-conference

Lee Sedol 9 dan: “I am confident that I can win.”

Attention now turns towards next week’s showdown against Lee Sedol, which begins on March 9.

Many expect the Korean professional to demonstrate the ongoing superiority of man over machine.

But are they right? Or does AlphaGo have more of a chance than they think?

DeepMind certainly seems to be more confident than the consensus predictions – at a Korean press conference Demis Hassabis suggested it would be “too close to call” – fifty-fifty odds.

David Silver’s comment, that he would be “very disappointed” if they lost, suggests even stronger confidence.

DeepMind counts among its numbers some strong amateur Go players, as well as pioneers in Go AI, so we’d expect them to be reasonable judges.

What do they know that the pros don’t?

Forget dog years, consider AI years!

Demis Hassabis appears on screen at the AlphaGo press conference.

Demis Hassabis appears on screen, at the AlphaGo press conference.

Well one thing they, and no-one else, knows is what they’ve been doing since October. And 6 months is a looong time when you are talking about machine learning!

As Hassabis mentioned at the press conference, AlphaGo has already done the equivalent of over a thousand years of human playing!

The ability to rapidly play games and learn from them means that AlphaGo, like our furry friends, passes the years faster than we do.

But instead of the 7-to-1 of dog years, we must consider ‘AI years’, which might allow thousands of years of learning per human year!

Though a striking image, simply having AlphaGo play itself for 6 months is an oversimplification of what DeepMind are actually likely to do to strengthen the program.

There are several potential improvements that might give AlphaGo extra strength, from most to least likely to be carried out:

  • Improving the accuracy of the value function
  • Tweaking model parameters
  • Adding more hand-crafted features to the rollout policy
  • Retraining the supervised learning network on professional data.

But before we delve into this in any more detail, it’s important to understand the source of AlphaGo’s strength that distinguishes it from other computer AIs.

Questioning your values

One of the things that separates professionals from strong amateurs is their ability to look at even a complex board position and tell who is ahead.

This question of ‘value’ of a board position has been a non-trivial problem in computer Go since inception, and DeepMind’s solution to it is the main thing separating its program from other Go AIs.

Deterministic, zero-sum games (like Go) actually have an objective value function across all board positions, but Go has too many combinations to ever calculate this precise value.

AlphaGo uses a neural network model to approximate the value function, and this model was created in three steps, building two other models along the way:

  1. A ‘policy network’ (i.e. a model giving a probability distribution over possible moves) built using ‘supervised learning’ (SL – where we get the model to make a prediction, then we give it the answer and it adjusts the model to ‘learn’ from the answer) to predict a human’s move, given a board position.

    AlphaGo’s supervised learning policy network successfully predicted human moves 57% of the time, when trained on 160,000 6–9dan KGS games, with a total of 30 million board positions.

  2. Another policy network, built by ‘reinforcement learning’ (RL) – taking the supervised learning network and getting it to play subsequent versions of itself and learn from the game outcomes, to predict the move most likely to result in a victory.

    The reinforced learning policy played 1.28 million games against different versions of itself, resulting in a very strong policy network for selecting moves.

  3. Finally, the ‘value network’, which was built by supervised learning & regression over board positions and values generated from the SL and RL networks, and predicts the expected value (i.e. probability of a victory) of a board position.

    To do this, AlphaGo generated 30 million games, playing the first n-1 moves with the SL network, then selecting a random legal move, and then using the RL network to select all moves until the game ends and a value (i.e. win/lose) is known.

    The value network was then trained on just one board position from each game – the one subsequent to the first RL network move – to minimize the error in predicted value.

This complex process resulted in a value function that is closer to the ‘real’ value function for Go than anyone has ever achieved before.

In fact using the value network alone, AlphaGo beat all other computer AIs!

DeepMinding the gap

As a result of this, the value function is one area where DeepMind might be able to easily gain extra strength to close the gap with Lee Sedol.

By doing more reinforcement learning to build a better move policy, and then generating a much larger corpus of games and board positions and using them to retrain the value function, they could further improve its accuracy.

All that is really required to do this is time and computing power.

Split the difference

Another area where I’d expect DeepMind to at least experiment and perhaps obtain modest improvement, is in some of the structural aspects of their model.

AlphaGo uses a hybrid approach that combines the more traditional Monte Carlo Tree Search technique of semi-random playouts with the above-described value function to assess moves.

At the moment those two techniques are given equal weight, but the optimum balance may differ from that.

There are also several other modelling constants where trial-and-error tweaking is inexpensive and could improve results.

Better than (better than?) random

Like other Monte Carlo based AIs, AlphaGo builds up a set of complete games from the current board state by playing lots of fast random(ish) games all the way to the end to see who wins.

But it is only randomish, because they use a ‘rollout policy’ to select which moves it’s more likely to explore.

AlphaGo’s rollout policy is built in a similar way to the SL policy, but is designed to be much (~1000x) faster.

Like the SL policy it uses some handcrafted features, such as whether a move is an atari, ladders, and a ‘pattern book’ of 3×3 patterns, to decide the move probabilities.

Though not as likely as the previous two options, DeepMind might try adding other such features, or tweaking the existing rollout policy to improve the Monte Carlo search.

Learn from the best

One other possibility suggested by several observers is that DeepMind could use kifu from professional games to get an edge.

This would involve going back to the start and retraining their supervised learning network, but instead of the KGS games, using pro games.

There are at least 80,000 pro games out there, around half the volume of KGS games used to train AlphaGo’s SL network.

However this approach seems unlikely, not only because it requires re-doing a lot of work, but also because if there were no obstacles, DeepMind would have used it to begin with.

Whether due to copyright uncertainties or some other reason, they will probably stick with the KGS dataset they’re already using, or add data from strong players on other Go servers (like Tygem).

What forest? All I see is trees

All of that may not be enough to beat Lee if professionals’ observations about AlphaGo’s weaknesses reflect inherent limitations in its approach or structure, rather than simply not yet having learnt from enough board positions.

The critiques all touch on similar themes, which centre around a lack of whole-board awareness, or high-level play.

Among the weaknesses suggested are a lack of understanding of sente, no useful conception of aji, and a lack of ‘creativity’, or following common patterns (albeit patterns often used by strong amateurs or even pros) where the specific context calls for deviation.

Myungwan Kim’s comment about “5 dan mistakes” seems particularly prescient. It may make sense that AlphaGo struggles with whole-board interconnectedness given the structure of its underlying models.

Convolutional neural networks (CNNs) are typically local by nature, and don’t build a good understanding of the whole board.

According to Rémi Coulom, AlphaGo’s architecture uses 1 layer of 5×5 convolution, and 11 layers of 3×3 convolution, meaning “it can propagate information at a distance of 13 points, but not more”.

“So if a large dragon has one eye each, on opposite sides of the board, the neural network is completely blind to figure out whether it is dead or alive.”

“Maybe it will suppose that any such large dragon must be alive, which works well in practice most of the time.”

Though AlphaGo is remarkable when compared to all previous Go-playing CNNs, this may still be a limitation for certain positions with important non-local interactions.

A known unknown

Where does this leave us, in terms of predicting who will win next week? Sadly, there’s just no reliable way to know.

While it is clear that October’s AlphaGo would be most unlikely to win, it is also pretty clear that despite speculation on some limitations there is no obvious upper bound on how much AlphaGo will improve by then.

Demis Hassabis and the DeepMind team have expressed quiet confidence.

Unfortunately, when we contacted Demis for this article, he was unable at this stage to answer further questions “until well after the match”.

Perhaps we’re best off heeding the words of someone who knows more than most about AlphaGo’s preparations, and sticking to their prediction.

After all, who can think of a more exciting match than one that’s too close to call?

Lee Sedol (left) and Demis Hassabis `high five' on screen at the AlphaGo press conference.

Lee Sedol (left) and Demis Hassabis `high five’ on screen at the AlphaGo press conference.

Match details

Full details for the match and ongoing reports can be found on the DeepMind AlphaGo vs Lee Sedol page.

Who do you think will win?

Who do you think will win the upcoming match between Lee Sedol and AlphaGo?

Can Lee Sedol see off this challenge to the superiority of humans in Go?

Or do you think DeepMind is about to pull another rabbit out of a hat?

Share your prediction by leaving a comment below.

 

Related Articles

About Guest Writer

The article you've just read was submitted by a guest writer. Go Game Guru is always happy to accept guest articles from readers. If there's something you'd like to write about the game Go, why not let us know? You can leave us a short message below or send us a quick email. Please ensure that any articles you send us don't require a lot of extra work before they can be published, otherwise your article might be delayed.

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

Comments

  1. I think Lee Sedol will win, but it won’t be easy.

  2. As long as Lee wins tomorrow, or if the loss doesn’t upset him, he should be fine.

  3. I think go players underestimate the progress that AlphaGo will have made since October. From move 100 onwards it will basically make no mistakes anymore and Lee Sedol will. Lee must win the games in the opening. It’s possible, but my money is on AlphaGo, 4:1.

  4. Sedol needs to win to keep the magic ✌😜

  5. Lee will win I have zero doubt of that anyone thinking otherwise is over estimating the skills of the bot.

  6. Daniel Rich says:

    AlphaGo will win.

    One thing it was mentioned that Fan won 2 out of 10 games and that is fairly misleading.

    5 games were played with 1 hour main time and 30s byo-yomi
    Alpha-go won all of these.
    5 games were played with shorter time controls(forget what)
    and fan won 2 of these.

    So AlphaGo struggles with shorter time limits which these games are going to be 2 hours main time so AlphaGo may see a significant boost just due to more rollouts. The two games Fan won actually make me think AlphaGo is more likely to win now since it was shown to be fairly tight on time and now it will have even more.

    • Ben Kloester says:

      Hi Daniel, thanks for the comment! I didn’t intend to be misleading – earlier versions of the article actually mentioned the byo-yomi but I took it out for the sake of brevity.

      One thing you may not be aware of is that AlphaGo keeps searching the tree during opponents moves as well, so more main time is actually doubly bad for its opponents.

      I think even with more time, the AlphaGo from October is nowhere near strong enough to beat Lee Sedol. So the real question is less about the game conditions such as time/byo-yomi, and all about how much stronger it has gotten in 6 months.

    • Claudio says:

      “AlphaGo is more likely to win now since it was shown to be fairly tight on time and now it will have even more.”

      Well, sir, whilst Alphago us stronger this time, given a longer time control and improved AI, don’t forget that the human player is also significantly stronger. Your analysis would be reasonable if both human players were of equal strength. But you are simply not considering one simple fact: Lee Sedol > Fan Hui.

      Computer might win, though. But reread your comment and you’ll notice that you are stating it’s “more” likely that it will win only because of the time control alone. You seem to be overlooking too many other factors that will affect this new match :/

  7. I’m assuming this program can’t be run on a modern computer, but how strong would your PC have to be to be able to run an AlphaGo software program? How long before people can purchase the program and play against it?

    • Ben Kloester says:

      Travis, DeepMind were kind enough to (somewhat) cover this in the paper. The standard (non-distributed) version of AlphaGo ran on a single machine, with 48 CPUs and 8 GPUs. So not exactly average, but not inconceivable for a home build.

      The Distributed version that played Fan ran on 1202 CPUs and 176 GPUs. By the way, the CPUs run the MCTS and rollouts, and the GPUs are used to evaluate the neural networks(s).

      It is not clear whether DeepMind plan to commercialize AlphaGo – most people think not; at least their public comments seem to indicate it is not on their radar. However, the paper contained enough detail to replicate what they built very closely, and there is already a Free Software effort to replicate the AI begun on Github.
      https://github.com/Rochester-NRT/AlphaGo

  8. Shai Simonson says:

    I very much hope that Lee trounces the program. Because that will mean a few more years of interesting progress. I am too young to have this problem solved 🙂

  9. Tony Collman says:

    Thanks for that explication of the architecture of AlphaGo. I was wondering if, having settled on an adversary, Deepmind would have “fed” all Lee’s games to the machine, but it seems from what you say that that would mean going back to square one in some sense.

    Travis:

    The Fan Hui match used a huge network of computers controlled by Google. I read somewhere (from someone at Deepmind as far as I recall) that it is extremely unlikely that there will ever be a commercial release of Alphago.

    • Ben Kloester says:

      It is not just that it would mean going back to square one. It is that there is not really any way to train the network specifically to Lee’s games, because they would represent such a small portion of the overall training data. Deep neural networks respond best to large training datasets.

  10. I hope AlphaGo plays better against Lee than Iyama did in their latest game (Mar 4 on gokifu.com). I had the impression that Iyama had a very bad opening and never recovered. He had to play many moves on the inside to save a bullied group, even in the midgame, all while Lee built up a very thick outside. I could be wrong—hope to see some commentary on the game.

  11. David Holland says:

    Perhaps the question should be what will win?

  12. Lee Sedol would have won easily last year. Uless the program improves certain aspects from the ground up, I find it hard to see how it will overcome the obvious flaws in its game at a high level – it will just be more accurate at that basic professional level. I am also still completely unsure of its ability at ko, as the earlier games were unusually devoid of ko fights.

    So, for me, unless it can stake the game on a very complex branched fight where it has the advantage of pure computing power, or it has made some step changes in its model, I think Lee still has a big edge.

    Lee by 4-1 or 5-0 for me.

  13. The article makes some shrewd points. With all due respect for the awesome abilities of professional Go players, they really know very little about what Lee will be up against in a few days. In contrast, the Google DeepMind team knew exactly what they were taking on when they challenged Lee. In their Nature paper they estimate the October version of AlphaGo at Elo 3100 and Lee’s strength at Elo 3500. This coincides pretty much with the balanced assessment of Myungwan Kim who judges that October AlphaGo would have decent chances of winning against Lee taking black with some reverse komi. So the Google DeepMind people knew very well that to have a fair chance of toppling Lee, they have to come with a version of AlphaGo that is at least one or two stones stronger than the October version. And given that they’re still “quietly confident”. I can’t think of any reason why they would be bluffing, knowing very well that they’re not up to the task, that would be ultimately counterproductive from their point of view. They could be overconfident — this happens to very smart people — or simply wrong: but they don’t come across as very brash types, and when they judged that they had a version of AlphaGo that should be somewhat stronger than Fan Hui, they were not wrong, they were spot on. Lee Sedol, on the other hand, is the one who is coming across as overconfident, at least in his February conference statements. I hope he has had time to reflect, or that some wise people like Myungwan Kim are giving him some good advice and telling him to come fully prepared for a really strong opponent, because otherwise he is in for quite a shock.

    • Ben Kloester says:

      Thanks! I’m flattered that you found parts on my analysis to be shrewd.

      If anything since arriving at the 50/50 split I have been sorely tempted to up the odds in the AI’s favour! This is mainly due to thinking about the small amount of evidence we have about the personalities and motivations of the DeepMind team.

      Most notably, that they sought the match with Fan when they did (eg not sooner), and then the AI won 5-0. They also had the choice of when to challenge Lee, so I’ve been wondering if they waited until they were similarly confident?

      By the way, the thing that originally got me interested in this question was a topic on the Good Judgement Project, which is a forecasting platform. If you’re a fan of making predictions based on shrewd objective assessments of probability, come check it out.

    • Anonymous says:

      I’m not sure if this is true at all, but I somehow remember something about him expressing more confidence in plain terms than other pros when he was younger, e.g. wanting to be first board in the Nongshim cup once hoping to win many many games), and finding it interesting that he then had to face Ke Jie in the MLily Final who gave him a “5% chance” which turned to be a lot more than 5% chance when winning the series by half a point. Well, I’d give Alphago 5% chance, so it would have 50% chance!

      Despite the clear fact that they must have created a bot the was a win ratio against it’s older versions showing it is about two stones stronger than it’s October version, which suggests it can obviously play at the top pro level I still think they’d need an actual match of it playing against a top level pro just to iron out any tiny flaws which could completely undermine or hide it’s full ability. Then they can have another match a couple of months later, in which AlphaGo could have a serious edge over the competition.

      [1] It is possible that calculating the stone-wise difference in strength from elo’s suggested from win ratios against the same level of players may over (or under) rate it’s level. See Iyama Yuuta on Goratings.org

  14. Elias Westerberg says:

    I’m a little surprised the article didn’t consider adding more computing power, as a way to improve AlphaGo. The distributed AlphaGo scales fairly well.

    Has it really been confirmed that it will use the same set up as in the match against Fan Hui?

    This article has some interesting comments on this: http://www.milesbrundage.com/blog-posts/alphago-and-ai-progress

    • Ben Kloester says:

      Hi Elias, I considered this when researching the article, but it didn’t make the list of likely improvements because DeepMind have already pushed this about as far as they can.

      There’s a very useful table in the Nature paper showing ELO level vs hardware used, and it looks something like this:
      Search threads CPUs GPUs Elo
      12 428 64 2937
      24 764 112 3079
      40 1202 176 3140
      64 1920 280 3168

      See the last step, where a 60% increase in hardware yielded less than a 1% increase in strength?

      • Ben Kloester says:

        I just had a look at the linked article – it is great! I wish I’d found it before I wrote mine.

        But I think the writer makes a twos error in their extrapoliation. The first is including data points from both the asynchrous and the distributed version on the same plot.
        The second is using linear extrapolation to project strength returns to hardware.
        If you remove the leftmost point, which is the asynchronous version, the data falls into what looks pretty neatly to be a logarithmic shape. Certainly it is sublinear.

      • Hanspeter Schmid says:

        This is beacuse of the nature of artificial neural networks. At some level of hardware complexity they will not perform better because they have reached the level of what they can do with their present training state.

        However, if AlphaGo has now really trained for five more months by playing against itself, it may very well reach a neural-network training state where the level at which more hardware is useless would be higher up. And then, we can be sure, Google does have more hardware to throw in 🙂

      • “40 1202 176 3140
        64 1920 280 3168

        See the last step, where a 60% increase in hardware yielded less than a 1% increase in strength?”

        That 1% increase in strength was based on games playing at 2 secs per move.
        If more time per move were allowed, the increase could be even smaller!

  15. We should not forget that AlphaGo is allowed to “cheat,” in a way that is invisible outside the computer(s): It “tries out” moves, and for each move selects one of the moves that the “trial runs” credited with the highest “best outcome” rank. Of course, so does the human, but in his/her mind alone, whereas AlphaGo is physically executing the trial move sequences, making notes, and going through evaluation algorithms. This is equivalent of having a staff of skilled go coaches giving advise, and is NOT allowed in professional go games.

    If Lee Se Dol was allowed to confer with a group of high-ranked players for each move, I would bet BIG for Lee Se Dol to win, all five games, hands down. But since only AlphaGo will be allowed to “cheat” in this way, I only wager a smaller sum … That Lee Se Dol son saeng nim STILL will win all five games!

  16. So much buzz about this game, I think tonight’s game for the Nongshim cup final is much more interesting.

  17. Roland CA says:

    The Nongshim cup final between Lee and Ke is about to start in 5 minutes. I heard youtube will stream the game records an hour after the start. Anyone knows the website for live video coverage showing the players in person?

    Lee has been very impressive lately, winning big vs. Lian and Iyama.
    For faster games of 4 hours total time or less, I think Ke has a much better chance, 4:0 so far.

    Overall, it’s been 7:2. Will it be 8:2 or 7:3 after tonite? does not seem to make too much difference. If Google Ai can mimic Ke and play like him, it can win 🙂

    • Lee got killed, but wasn’t he just bad out of the opening when he refused the main line of the taisha?
      Black had an ideal formation on the left. When Lee came in desperately, way too deep, his fate was sealed. As I was mentioning a few weeks ago, many pros seem to avoid the taisha. This may be a weakness Ke Jie had prepared for.

  18. Andy Z. says:

    Wouldn’t it help DeepMind if they would hire a top pro to mentor AlphaGo and point out where it goes astray? But maybe they think they don’t need this help!

    • Hanspeter Schmid says:

      I wondered abou the same thing. From the point of view of training an artificial neural network, it is less than trivial to figure out a way in which a top pro could make AlphaGo’s neural networks learn something.

      I don’t think Google are above needing this help, I just think they cannot possibly have found a way to make AlphaGo undestand what a human pro wants to teach it. If they have, I would find this much, much more exciting than ALL the rest of the project.

    • Well, according to
      http://www.milesbrundage.com/blog-posts/alphago-and-ai-progress
      they have hired Fan Hui to help. Not a top pro perhaps, but still pretty strong: also the ability to analyse, communicate/teach is just as important as pure go playing strength in this instance, and Fan is a known teacher. This tends to indicate that feedback is important to DeepMind, meaning they have probably been tweaking their program and checking its progress. They can also test its strength against Fan (can it give him a two-stone handicap ?) and see where they stand much better than just by thrashing other go-playing programs.

  19. “DeepMind counts among its numbers some strong amateur Go players, as well as pioneers in Go AI, so we’d expect them to be reasonable judges.

    What do they know that the pros don’t?”

    AlphaGo is their baby. Can we actually expect them to be reasonable judges? When a company like Apple announces that it has released some new product, do we just accept that at face value or do we know that Apple will always try to promote it’s products as fantastic regardless of the flaws?

    I see comments like “AlphaGo has already done the equivalent of over a thousand years of human playing!” and I don’t think they have an appreciation of what that actually means. All one has to consider is the fundamental changes to the game that happened from the 1900s to the 2000s. We’ve seen new approaches to the fuseki, the introduction and adjustment of komi, etc… So has AlphaGo experienced the same cultural shift as the Go world did when komi was introduced? Does it understand why certain fusekis have been abandoned or reintroduced? Does it understand why certain josekis have fallen out of favor? Can it develop novel ways of playing? Does it understand the implications of changing rules from Japanese to Chinese? Or even Ing?

    All of these things are something that human players can do with out a thousand “human years” of playing.

    Could you imagine if some of the great players had over a thousand human years of playing? What would players like Go Seigen been able to do? We admire Shusaku’s never losing a castle game because we have context of what the castle games were. We appreciate how strong Lee Changho is because we understand how strong the opposition he faced and we see how many titles he still won despite that opposition. The same is true for Lee Sedol. The same should be true for AlphaGo.

    Except that AlphaGo’s only claim to fame so far is that it beat Fan Hui. It won easily to be sure, but it has advantages that Fan Hui does not. I have nothing but respect for Fan Hui’s strength as a Go player. He is far above my skill, but he is still a human being. If he is tired, it will affect his ability. If he was irritated, hungry, dehydrated, so on and so forth.

    Not only is AlphaGo a machine (that never suffers from things like hunger) but it also it had the luxury of anonymity. Imagine that Fan Hui had won easily. Would the AlphaGo developers have trumpeted on the rooftops that they had lost completely? I suspect not. They would probably go back to developing quietly and the world would never know. They could have done this many times until they achieved success. Now the world is aware of them and the world is watching. So if they do lose…it’s a setback to be sure. However, they still have the advantage of time. They can just go back to developing quietly and wait till next time. Then they’ll try again. Maybe they’ll lose again, and they’ll go back to developing and try again later. They can do this over and over again, as long as they have the backing of Google and the interest. Make no mistake though, that when they win…they will be out there trumpeting about their success. Just like they are doing with Fan Hui now. As they put it “Mastering the Game of Go.” AlphaGo hasn’t “mastered” anything.

    “Look at how smart our program is!” Well, sure it is quite “smart”. However, none of it’s “smartness” is due to it’s own initiative. It relies on having studied millions of positions played by strong human players. It relies on a team of very smart programmers with access to a very large budget. Without those…what exactly is AlphaGo?

    I’ve said it before and I’ll say it again. I’d be very interested to see CrazyStone and Zen playing on the same level of hardware that AlphaGo played on with Fan Hui. CrazyStone won against a professional player with 4 stones on much less impressive hardware. I don’t see Remi Coulom submitting a paper to Nature.

    In short, I suspect that Lee Sedol will win. Even if he does though, will that matter? The AlphaGo team will just go back to developing and will try again. They’ll keep trying and eventually they’ll win. Then they will beat their chest and proclaim “total victory” over the game of Go. The world will herald a “new era” for the game and then most of the world will forget the game exists. AlphaGo will mothballed and what will the game of Go have actually gained? Not much, I think. It strikes me as sort of sad, that a game that is thousands of years old, with a great deal of history, that is constantly changing and developing, is reduced to being “mastered” because a human player lost.

    Can we actually expect them to be reasonable judges? I see comments like “AlphaGo has already done the equivalent of over a thousand years of human playing!” and I don’t think they have an appreciation of what that actually means. All one has to consider is the fundamental changes to the game that happened from the 1900s to the 2000s. We’ve seen new approaches to the fuseki, the introduction and adjustment of komi, etc… So has AlphaGo experienced the same cultural shift as the Go world has when komi was introduced? Does it understand why certain fusekis have been abandoned or reintroduced? Can it develop novel ways of playing? Does it understand the implications of changing rules from Japanese to Chinese? Or even Ing?

    All these things are something that human players can do with out a thousand “human years” of playing.

    Could you imagine if some of the great players had over a thousand human years of playing? What would players like Go Seigen been able to do? We admire Shusaku’s never losing a castle game because we have context of what the castle games were. We appreciate how strong Lee Changho is because we understand how strong the opposition he faced and we see how many titles he still won despite that opposition. The same is true for Lee Sedol. The same is true for AlphaGo.

    Except that AlphaGo’s only claim to fame so far is that it beat Fan Hui. It won easily to be sure, but it has advantages that Fan Hui does not. I have nothing but respect for Fan Hui’s strength as a Go player. He is far above my skill, but he is still a human being. If he is tired, it will affect his ability. If he was irritated, hungry, dehydrated, so on and so forth.

    Not only is AlphaGo a machine (that never suffers from those) but it also has the luxury of time and it had the luxury of anonymity. Imagine that Fan Hui had won easily. Would the AlphaGo developers trumpeted on the rooftops that they had lost completely? I suspect not. They would probably go back to developing quietly and the world would never know. They could have done this many times until they achieved success. Now the world is aware of them and the world is watching. So if they do lose…it’s a setback to be sure. However, they still have the element of time. They can just go back to developing quietly and wait till next time. Then they’ll try again. Maybe they’ll lose again, and they’ll go back to developing and try again later. They can do this over and over again, as long as they have the backing of Google. Make no mistake though, that when they win…they will be out there trumpeting about their success. Just like they are doing with Fan Hui now.

    • Hanspeter Schmid says:

      @Wasuji: I read your contribution with great interest, and it gives me a good new perspective on the whole AlphaGo thing.

      There is one thing, though, where I think you are quite wrong.

      You say “CrazyStone won against a professional player with 4 stones on much less impressive hardware. I don’t see Remi Coulom submitting a paper to Nature.” etc.

      Speaking not as a go amateur now, but as a professional scientist, I think you misunderstand what a Nature publication is and how difficult it is to get one accepted. A bit of boasting and a misrepresentation of reality won’t do, the reviewers would reject it. Google’s work is not impressive because AlphaGo beat Fan Hui, but because AlphaGo got very, very little Go knowledge programmed into it (basc rules, and reading ladders, and that’s it) and learnt everything else by looking at a game database and by playing against itself. If I haven’t missed something out there, this differs AlphaGo from EVERY other strong go program.

      The Nature publication is completely merited, not because of some ELO number, but because they successfully tackled an old problem in a totally different way. I know something about artificial neural networks and have worked with a few, and I have read the paper, and I am totally impressed by what that team have accomplished.

      Even if AlphaGo gets mothballed, the advance in the theory and application of artificial neural networks is massive, and this is why Nature printed the, not because of the game (and rest assured that I love the game, it’s just that Nature’s priorities are not right with Go, but with science).

      • I have no doubt that submitting a paper to Nature and having it accepted is difficult. I simply ask that if AlphaGo had lost the match…would they have submitted the paper still? I don’t know. I’m not denying that they have contributed novel ideas to Go playing programs.

        However, I do think that Anders K. is right when he states in his blog post. “Massive amounts of hardware are helpful too.” from https://smartgo.com/blog/google-alphago.html

        I’ll ask again…would this have been as impressive if it lost? As novel as the solution is…it’s only because it won that we are talking about it. To further elaborate on my point, how well would this program scale down to a similar level of hardware that Crazy Stone played on when it beat Ishida Yoshio? Why use the “distributed version” if the single machine version is (by the own estimation) just as strong as Fan Hui? If it doesn’t have access to GPUs, what can we expect of it? Note that the program that beat Fan Hui was running on 1920 CPUs and 280 GPUs. Crazy Stone played on 64 cores (I’m not sure what the makeup of that hardware was, though if you look at AlphaGo’s paper, no GPUs were used in their internal tournament). So I certainly believe that it is a novel solution, but it’s a novel solution that relies on tremendous amounts of processing power.

        Which raises another question for those who think this is going to be a teaching tool…who is going to be hosting this program for professionals (let alone amateurs) to play against and study with? Is there going to be a cluster computer like some universities run and dedicating all it’s resources to this program, so the professional can play a game against it? How is this going to scale down to a home PC ( or even a cellphone )? Can it?

        I don’t mean to sound so negative and I don’t mean to imply that the AlphaGo team hasn’t accomplished anything. I am just skeptical. I remember AI history and how we were assured as far back as the 1960s (even earlier), that the “future” was right around the bend. The developers assure us that this technology will lead to improvements in other areas. This may be true, but again, who is hosting the cluster for these applications? I suspect, like I said prior, that it will win (eventually) and it will be mothballed. Even if the technology is used to improve say, self driving cars…you still have to get humans to accept self driving cars to make that technology viable and useful. That is historically a very tough thing to do and may take decades to accomplish.

        • Hanspeter Schmid says:

          @Wasuji: Thanks for clarifying this that. I think that if they had lost, they would still have submitted, because they have all comparisons to other Go programs in their paper, evaluated in a tournament, and AlphaGo is massively stronger. And as I said, I agreed with most of what you said except with what I perceived as an attack on the team’s scientific integrity. I apologize if I overreacted.

          Regarding your question “I’d be very interested to see CrazyStone and Zen playing on the same level of hardware that AlphaGo played on with Fan Hui:” They tried 🙂

          It is not exactly the same level of hardware, because one AlphaGo unit has 48 CPUs and 8 GPUs, and no other program can make use of this, but they let AlphaGo with 48 CPUs/8GPUs play against CrazyStone Version 2015 with 32 CPUs in a tournament and got 2890 ELO for AlphaGo vs 1929 ELO for CrazyStone.

          (A GPU is a Graphics Processor normally used for generating what a computer wants to show on its screen.)

          On the one hand, AlphaGo clearly takes its power from the additional GPUs, as its ELO drops to 2181 (still above CrazyStone) if only 48 CPUs/1GPU is running. It cannot run without any GPUs because those evaluate the policy and value networks. If all GPUs are switched off, AlpahGo falls below 1600 ELO and is weaker than CrazyStone or Zen.

          On the other hand, CrazyStone cannot actually make use of a GPU, so there’s where it gets difficult (near impossible) to compare the two programs on a similar level of hardware.

          For this reason, I also think that AlphaGo will scale down very, very badly to cell phones, because those don’t have a GPU that can be used. But maybe, in the future, when deep learning and neural networks become more generally used tools, cell phones will have GPU-like co-processors, and THEN they will be in the business again.

    • Wow, dude! I’m really impressed with the length of your post, but you should know I didn’t bother to read it.

  20. Sorry about the duplication :-). Only human

  21. I guess it might be 4-1 in the favour of Lee Sedol. It would be nice to see AlphaGo win one game, showing its particular strength. But I would love to see Lee win the match, because I like his style of play, shown in the recent Nongshim match: is Ke Jie from a different planet? After the match, I hope AlphaGo and its successors will become 20p in real strength, and that they will show us the direction to the truth of go, the right strategies, the value of sente, and getting rid of the connotations of go to zen and the like mumbo-jumbo.

    Kind regards,
    Paul

  22. Warren Dew says:

    I will make a prediction: not only will Lee Sedol win, but AlphaGo will never be improved enough to beat people in the top 6.

    Human players improve by “training” on games at their current level. As they get better, they train on higher level games.

    AlphaGo cannot improve substantially merely by analyzing more games played by moderately good players. It would have to analyze a similar number of games by top players, and not enough games exist.

    Some other computer program, likely not yet written, will beat the top human on the originally predicted schedule.

  23. I think anyone who says Alphago will easily win Lee Sedol does not really understand Baduk (or Go). While there are nearly infinite number of different cases in Baduk games, Alphago has been trained with only limited number of prior games. Although Alphago can learn from those and may possibly create new tricks, I don’t believe it is immune to small or large mistakes when it meets with the situation it never encountered before. Top professionals not only can easily take advantage of any small mistakes, but also they can create new tricks while playing games. For this reason, I believe Lee Sedol wins at least this time as Lee Sedol said in an recent interview. Only one external factor that may play in the game is that playing with machine is awkward format for Lee Sedol, which may cause human errors.

  24. I see that people seem pissed that top players could be beaten by computers, but I think you should see this from a different perspective. Top players lack strong partners to train and improve their game. It happened in the past with Chess, and now all top players use computer chess engines as “sparring partners” to improve their skill level. And you can see that for the last decade or so, their level has really increased significantly, and I believe the computer is a big part of it.
    Computers will not “solve” Go any time soon, they will probably beat top humans sometime soon and it could even be good news for them.

  25. Graahm Laight says:

    AlphaGo. I think it’s likely that the experts commenting on the previous game have noticed lots of weaknesses, but haven’t noticed some strengths because they don’t match the normal human understanding of the game.

  26. I could retire from the winnings if I could bet against the people who predict a 5-0 Lee win.

  27. From what I understand, Alpha go still can’t read deeply ahead in novel situations. Lee should seek to play unique moves and fights. Playing an orthodox game will mean playing to Alpha Go’s strength, as the program is based on learning most-appeared patterns from human pro games.

  28. Brennus says:

    Lee will win. They’re doing this all half-assed, though: KGS 5-dan games for training and no whole-board heuristic. What they should do is convert an entire google server site to a Beowulf cluster then hire Cho Chikun to play the thing 24/7 🙂

  29. Hanspeter Schmid says:

    I predict that Lee Sedol will win 5:0 or 4:1, and if it is 4:1, he will
    lose the first game and win the others.

    The reason is that Lee is a self-confident and imaginative player. He
    knows from Fan Hui’s five games where to look for AlphaGo’s weaknesses,
    and he will come well prepared. Of course AlphaGo will have learnt a
    lot in the past five months, but it will not have acquired true
    creativity. So I assume that Lee Sedol will probe AlphaGo in the first
    game, risking that he loses because of those probes, and then adapt and
    win games 2 to 5. Lee Sedol will be able to adapt to AlphaGo, and he
    will not lose faith in himself and become nervous if he loses the first
    game. AlphaGo, however, is not equipped to adapt to Lee Sedol after
    just one game.

  30. Ben Kloester says:

    Alphago has been trained on 30 million positions. In the 6 months since October, there is no practical barrier to Deep mind increasing that to 300 million. No human has ever seen that many positions. It is only because the network is not as good at learning from a position as we are (because of abstracted thought, imagination and generalisation) that humans would even stand a chance against such experience.

  31. GoPlayer says:

    Lee Sedol is very famous for his creative playing style. Unlike other top players, he often does not follow traditional or known strategies. He playing style is also aggressive. I remember he beat an 9 dan opponent using a cascade format, which is very unusual because cascade format is known to be a mistake only beginners make and even amateurs try to avoid.

    I know Alphago has been trained with prior games, but I cannot imagine Alphago can play consistently by putting stones in the most optimum place with a strategy in mind.

    Other thing western players might be missing is that the level of European champion was only 2 dan while Lee Sedol is the top end of 9 dan. If those two play together, the chance Fan Hui beat Lee Sedol is close to nil.

    Putting everything I know together, I can hardly imagine Alphago beats Lee Sedol this time. It might be improved in the future, but it seems too early for Aphago beat the human champion.

  32. In any case, i don’t find the conditions of this challenge fair. The DeepMind team keeps every piece of information secret. Lee Sedol does not have access to any of the games of Alphago, other than the ones played in October. He also does not have a clue about the estimated Elo strength of Alphago. I think it was _very_ risky to accept such conditions. It may take more than 5 games to realize what the computer’s weaknesses are, if any. Lee Sedol should have insisted to have access to at least 10 recent games of Alphago (possibly played against itself). How does he prepare against something he has no information about?

  33. Baldomero says:

    As a mere kyu player, I’d bet on AlphaGo, even though I’m lifetimes away from appreciating and understanding the depth and greatness of Lee Sedol’s GO and know zilch about AI programming algorithms.
    Here’s why.
    AlphaGo went from 30k to 2/3p in… months? a year? So, several stones stronger every month. It takes the pliable mind, discipline and dedication of an aspiring Asian Go child player from childhood to early or late teens to reach 1p, and that’s an elite few. AlphaGo was probably already 3p in October, accounting for Fen Hui’s actual 1p or so actual rank, being rusty, what with playing mostly weaker European circuit rivals. Still, he remains a young, active Chinese pro, no slouch. And he got trounced. The difference between a 3p and a 9 p is what? about a stone? The “vast differences” among pros are not vast at all. Just read pro games commentaries. They are just very important, because the tiniest error costs you a game and maybe a tournament, from which you extract your livelihood. Important differences. Not big ones. A slight positional advantage here, a bit of bad aji there, slightly sub optimal yose later. So, improving even half a stone a month would make it unbeatable by now. Unless a plateau is assumed, due to the missing magical human touch of creativity or whatnot. A creativity that top chess software either somehow have, or do not require. A creativity that is probably just the still poorly understood human problem solving algorithms. So, barring such magical human quality, AlphaGo has likely kept on improving, trying all kinds of tactics against itself over millions of games. That’s a lot of games. Enough to cover centuries of evolution of the game.
    The programmers, as mentioned before, assessed very accurately the strength of the program against a weakish pro last October. Plus they have access to servers such as Tygem where they can pit their monster against unaware known pro playing there. Online games are more casual, sure, but pros are competitive and don’t like to lose anywhere. I’d bet they’ve been losing big since December or so, unless this newcomer Pro that insists on slow games. A cybernetic Sai. And there’s a repressed excitement is the statements of the AlphaGo team, like they want to boast but do not want to ruin the coming worldwide shock. I read the between the lines and they are definitely biting their tong.
    As to Lee Sedol’s creativity and unexpected moves: The software has assimilated every pro game ever recorded, icluding all new joseki and more eccentric opening moves ever tried. To deviate much further would simply handicap Sedol, as the moves would be bad. Does an insei beat a no nonsense, conservative Meijin playing tengen or 10-10? Probably not. if AlphaGo is 10 or 11p by now, tricky moves will be punished with their correct response and will merely precipitate the loss.
    My prediction: 50%/50% (50% 10p rank – 50% >10p) 5-0 no contest

  34. I think there has been some over-estimation over just how much AlphaGo could have improved over these past few months. Sure, it’s playing itself thousands of times a day or whatever, but is it capable of correctly evaluating these games? If it cannot say where those games are being won and lost and then learn from it then all those games don’t amount to much. High level human players, however, are able to evaluate games brilliantly.

    Also, AlphaGo has only ever played itself and Fan Hui, whereas Sedol has pitted himself against the best of the best and came out on top on so many occasions. Sedol regularly plays high level games and I’m sure as part of his regular training has access to strong coaching, teaching, and input from other cream of the crop players. His environment is fine tuned to ensure he is formidable on the board.

    Has AlphaGo learned from amateur games? It should just learn from high level pro games. Learning from amateur games means learning amateur play and likely amateur mistakes. What use would Lee Sedol have from studying an amateur game? Apart from pointing out clear errors and sub-optimal lines of play, not a great deal, I would imagine.

    My money is on Lee Sedol.

    • Hanspeter Schmid says:

      @Damon: “Has AlphaGo learned from amateur games? It should just learn from high level pro games. Learning from amateur games means learning amateur play and likely amateur mistakes. ”

      No, this is not so. It means learning ABOUT amateur mistakes. A neural network is essentially a classifier for good vs bad, strong vs weak, etc. If you don’t give it weak and bad moves to look at, it will learn much less rapidly, and will not know all the weak moves to exclude from the Monte-Carlo tree search.

      This is something we also see, e.g., in devices that recognize bird calls with neural networks. If one is trained with flawless studio-quality sound samples, and a second one with noisy samples, the second one will outperform the first one by far.

      Same for humans, I believe. An amateur may get a bit better by only ever replaying high-level pro games, but reading a book like “How to not play go” can help much faster 🙂

      What I do find really interesting is where AlphaGo got WITHOUT looking at any pro games a all.

  35. Note that AlphaGo didn’t explicitly learn from a Fuseki library before the games against Fan Hui, which explains why it sometimes played strange.
    I think it’s natural Google will add the Fuseki libraries before the game against Lee Sedol.
    Yet another cheap improvement…

    • Baldomero says:

      From what I understood, the first step in developing the software was a predictive program that analyzed positions based on a database of pro games and predicted the pro’s move with 57% accuracy. I that is part of the program, then it does take game knowledge into account. It also explains why it plays in a much more natural way than conventional Monte Carlo algorithms.

  36. It’s true. That explains why it played openings “like most pros”, but it also means that there was no guarantee that it would play the fuseki like it should (according to the current consensus).
    So I guess Lee Sedol may try to take advantage of this during his preparation : either by going out early of traditional fuseki or by leading the game into a rarely played fuseki that AlphaGo may not have learnt well.
    But if Google has added fuseki explicitly inside the AI (by forcing fuseki instead of letting the engine search), then the strategy will not be effective at providing an early advantage to Lee Sedol.
    It should be easy to see during the first moves : if AlphaGo takes time to think,mit’s not using Fuseki and Lee Sedol could take advantage of it. If it plays instantaneously, Lee Sedol will have to wait until the end of the fuseki library moves to attack, I guess.

  37. it seems Lee Sedol decided to get out of Fuseki early, as expected.
    I was still surprised that AlphaGo took a minute or two to play his first move although it was obviously in Fuseki libraries.
    I was also surprised to see the way it managed its time keeping only 5 minutes for the end-game, but it seems it was the right decision to improve its position before the end, as it probably takes very little time for AlphaGo to calculate end-game moves…

  38. Hanspeter Schmid says:

    “I was still surprised that AlphaGo took a minute or two to play his first move although it was obviously in Fuseki libraries.”

    This seems to indicate that Google have NOT given AlphaGo a fuseki library.

    The October version playing Fan Hui also did not have a fuseki library, but there were speculations whether Google would add one for the Lee Sedol games. The answer seems to be “no”.

  39. Yes, that was exactly my point. But I still don’t understand why they would make such an apparently cheap decision. I don’t believe that they think AlphaGo’s neural network is already stronger than fuseki libraries…

    • Anonymous says:

      For a computer scientist, implementing an opening book is hard coding the answers into the code. In programming/CS, hard coding is considered inelegant, but often extremely useful.

      These guys are research computer scientists, so winning is great, but winning with an elegant algorithm is better.

  40. Daniel Rich says:

    In addition to the elegant algorithm argument though these guys aren’t interested in Go for Go. They are interested in using generic techniques that they can then apply to hard interesting problems(including those that may make money).

    An opening book is throwaway work if they want to apply their techniques to something else.

  41. Well, watching the second game, it seems that DeepMind feels more confident on their neural network’s ability than in baked fuseki.
    If they are right, AlphaGo could soon become a source of new/creative fusekis in the future…

  42. Baoliang Zhang says:

    Can Lee Sedol win the third game?

    I think Lee still have a chance to win the third game but the change is about 49:51

    My reasons are as below

    the positive factor

    1 Although Lee had tried very hard in the game and have done his best but he has not played as good as his top level due to stress and he may think too much about to find the weakness of AlphaGo If he play with a calm and normal attitude and treat the AlphaGo as normal player he will has a better change to win

    the negative factors

    1, Lee had not played move which make AlphaGo do not understand , Al through Lee has played some creative move but Alpha Go have played some unusual moves which Lee can not do much about it so AlphaGo played some move which Lee do not fully understand and also the other professional player do not fully understand

    2 AlphaGo looks like know when to fight hard and when to compromise and always keep the game under control

    3 .AlphaGO may not play the best move but every move it played may be the move to have a big probability to win so when it is leading the game it will not take much risk to fight hard but when the game is close it will start try harder

    4. AlphaGO may have a better ability to foresee the game so unless lee kill its big group or clearly leading over 10 point we can not amuse lee is ahead because we may see less than AlphaGo does

    5 AlphsGO may has better ability to calculation (reading and judge) on the position than Lee do

    6, If Lee has not shown his full ability due to the stress in the two games he played then AlphsGO also has not shown its full ability since AlphaGo will be stronger when it play with stronger player

    Baoliang Zhang

  43. Baoliang Zhang says:

    How can Lee Sedol win the forth or fifth game?

    It is very disappoint that Lee Sedol had not done well in his third game against AlphaGO.

    Although Lee Sedol had tried very hard in the game and had done his best but he has not played very well due to he is under huge pressure and very stress. He was behind in the opening due to one move which may be a lose move. he had tried very hard and used his all ability trying to catch up the game but he did not have any chance in the whole game.

    Then will Lee Sedol still be able to win the forth or fifth game?

    I think it is still possible for Lee Sedol to beat AlphaGo in the forth or fifth game if he relax and play with a calm and normal attitude and treat the AlphaGo as normal player. The chance for for Lee Sedol to win the forth or fifth game is more than 50% (51:49).

    My judgement is based on the following fact

    As Lee Sedol said from the new conference, Althrough AlphaGo is very strong but AlphaGo is not perfact. That mean if Lee Sedol play near perfectly (it is very difficut but possible ) he still can win.

    Apart from the third game on which Lee Sedol do not have any chance to win,
    In both the first and second game Lee Sedol had some chance to lead the game. If he play better and take the chance he could already win a game

    The third game does not represent the real level of Lee Sedol.. Lee Sedol had never lost a game in such a way on which he was behind in the very early stage game and struggled hopelessly with every effort and used his all ability but did not work at all. It is partly because that AlphaGO is strong but the main reason is Lee Sedol had not played on his best level due to the pressure and stress. Also as some professional said Lee Sedol may had think too much during the game.

    I believe if Lee Sedol play his best level he still has chance to win.

    The question if it is possible and how can Lee Sedol play his best level?

    I believe that it is possible for Lee Sedol to play his level. Because Lee Sedol has noting to lose after the third game and he should be free from pressure and stress in the fourth and fifth game

    It will help Lee Sedol to play with his best level if he do the following things:

    It is very important for Lee Sedol be fully relaxed before and during the game and do not think too much.

    Every move is count.It is very important for Lee Sedol to concentrate on every move he palyed is the best move and avoid any mistake from the begining to the end of game. Since AlphaGo do not always play best move the possibility for Lee Sedol to win the game if every move he palyed is the best move and avoid any mistake from the begining to the end of game. I believe Lee Sedol is capable on this if he is totally relaxed and concentrated.

    Assume that AlphaGO may have a better ability to foresee the game and AlphsGO may has better ability to calculation (reading and judge) on the position, it is important for Lee Sedol to avoid wristwrestling again AlphaGO.

    It may be beneficial to review the principle of go ten tactics and try to apply them during the game:

    1, Don’t be greedy;

    2, move cautiously into contested areas.

    3, Offend while defending;

    4, sacrifice to take the lead.

    5, Release the small and seize the big;

    6, throw away the ones at risk.

    7, Don’t be careless and hasty;

    8, coordinate your moves.

    9, Be moderate in coping with a strong opponent;

    10 aim for parity when in trouble.

    In summary, it is still possible for Lee Sedol to beat AlphaGo in the forth or fifth game if he relax and play with a calm and normal attitude and treat the AlphaGo as normal player. The chance for for Lee Sedol to win the forth or fifth game is more than 50% (51:49).

    Mark Baoliang Zhang

  44. Baoliang Zhang says:

    Can Lee Sedol win the fifth/last game ?

    I think the chance for Lee Sedol to win last/fifth game is very high more than 80% (81:19).

    Lee has find the weakness of AlphaGo and will play very sharp in the Fifth/final Game

    I think some of the weakness of AlphaGo may be as below

    AlphaGo will confused and play badly after it misread the position due to the unusual move of Lee Sedol

    As I mentioned before [1] the top professional go players have the some ability which AlphGo can not learn on this stage.That is imagination , creativeity and telent as a humen. As David Silver mentioned ‘HUmans inevitably have alot of tricks up their sleeve that we are not able to train against.’

    what trick will drive AlphaGO Mad?

    I think it will difficult for AlphaGO to read the the game correcly if some strange shape can combine the kyo with the other tesuji such as connect and die which works well in the game four

    AlphaGO do not sharp on the shape and potentil (aji) combination use of ko with the following tricky such as ,two-stone edge squeeeze,capture and recapture
    connect and die , sekito shibori (the stone-tower squeeze) snapback will cause AlphaGo suffered from blind spot and may become again

    Lee Sedol is a very great and sharp player he will be well prapared for the last game

    and win the game

    The former world champion XiHeLuo believe that it is possible for him to gave AlphaGO 4 handy by play with a unusual way to amplify/enlarge the its weakness.

    With the reason I have mentioned before[1-2] I think the chance for for Lee Sedol to win the last/fifth game is more than 80% (81:19).

    [1]How can Lee Sedol win the forth or fifth game?

    It is very disappoint that Lee Sedol had not done well in his third game against AlphaGO.
    Although Lee Sedol had tried very hard in the game and had done his best but he has not played very well due to he is under huge pressure and very stress. He was behind in the opening due to one move which may be a lose move. he had tried very hard and used his all ability trying to catch up the game but he did not have any chance in the whole game.
    Then will Lee Sedol still be able to win the forth or fifth game?
    I think it is still possible for Lee Sedol to beat AlphaGo in the forth or fifth game if he relax and play with a calm and normal attitude and treat the AlphaGo as normal player. The chance for for Lee Sedol to win the forth or fifth game is more than 50% (51:49).
    My judgement is based on the following fact
    As Lee Sedol said from the new conference, Althrough AlphaGo is very strong but AlphaGo is not perfact. That mean if Lee Sedol play near perfectly (it is very difficult but possible ) he still can win.
    Apart from the third game on which Lee Sedol do not have any chance to win,
    In both the first and second game Lee Sedol had some chance to lead the game. If he play better and take the chance he could already win a game
    The third game does not represent the real level of Lee Sedol.. Lee Sedol had never lost a game in such a way on which he was behind in the very early stage game and struggled hopelessly with every effort and used his all ability but did not work at all. It is partly because that AlphaGO is strong but the main reason is Lee Sedol had not played on his best level due to the pressure and stress. Also as some professional said Lee Sedol may had think too much during the game.
    I believe if Lee Sedol play his best level he still has chance to win.
    The question if it is possible and how can Lee Sedol play his best level?
    I believe that it is possible for Lee Sedol to play his level. Because Lee Sedol has noting to lose after the third game and he should be free from pressure and stress in the fourth and fifth game
    It will help Lee Sedol to play with his best level if he do the following things:

    It is very important for Lee Sedol be fully relaxed before and during the game and do not think too much.
    Every move is count.It is very important for Lee Sedol to concentrate on every move he palyed is the best move and avoid any mistake from the begining to the end of game. Since AlphaGo do not always play best move the possibility for Lee Sedol to win the game if every move he palyed is the best move and avoid any mistake from the begining to the end of game. I believe Lee Sedol is capable on this if he is totally relaxed and concentrated.
    Assume that AlphaGO may have a better ability to foresee the game and AlphsGO may has better ability to calculation (reading and judge) on the position, it is important for Lee Sedol to avoid wristwrestling again AlphaGO.
    It may be beneficial to review the principle of go ten tactics and try to apply them during the game:
    1, Don’t be greedy;
    2, move cautiously into contested areas.
    3, Offend while defending;
    4, sacrifice to take the lead.
    5, Release the small and seize the big;
    6, throw away the ones at risk.
    7, Don’t be careless and hasty;
    8, coordinate your moves.
    9, Be moderate in coping with a strong opponent;
    10 aim for parity when in trouble.
    In summary, it is still possible for Lee Sedol to beat AlphaGo in the forth or fifth game if he relax and play with a calm and normal attitude and treat the AlphaGo as normal player. The chance for for Lee Sedol to win the forth or fifth game is more than 50% (51:49).
    Mark Baoliang Zhang

    [2] My view on ‘Can AlphaGo beat Lee Sedol?’

    I do not think AlphaGO can beat Lee Sedol

    Demis Hassabis assume the victory of AlphaGo againgest Lee Sedol based on his jugement of AlphaGo
    is now going beyond -hopeful eventiually -what even best humans in this area can do and it has the capabilities to pick things it self.

    But I believe since the natury of the Go Game The top professional go players have the some ability which AlphGo can not learn on this stage.

    That is imagination , creativeity and telent as a humen. As David Silver mentioned ‘HUmans inevitably have alot of tricks up their sleeve that we are not able to train against.’

    I think Lee Sedol will prapare the new opening and new Joseki to lead the game into an area which AlphGo can not easily ‘understand’.

    and the game would be decisided on the middle game stage. Lee Sedol will prapare two types /styles of game to avoid AlphGo to Pick up his style

    after the second he will change the style in the third game. In this way he can at least beat AlphGo in 4:1 if on fifth Gamme AlphaoGo leared
    and improved from the former four games then AlphaoGo may have some Chance at the last game .

    Althrough some people claim that AlphaGo has only used the Game records of amateur players to training itself before he played againest the professional player Mr Fan (2P)

    (If this is true ,) then the AlphGo may have ablity to beat the top professional player if the game records of top professional player been used on training.

    I think this will be true if AlphaGO play with the professional player whose rank is below top 20 in the world.

    But with the professional player whose rant is above the top 5 in the world AlphaGo still is not going behond their ablity. They have more ability on creativity move

    Expscially Lee Sedol is very good on unusual, creavity moves which will suprise AlphGo

    If AlphaGo beat Lee Sedol in the March that will be a big suprise and that will indicate that AI can have some ability

    to be creative or Lee Sedol has not use his ablity of creativity

    I look forward to seeing game between AlphGo wit Lee Sedol

    Mark Baoliang Zhang

  45. Baoliang Zhang says:

    I do believe Lee Sedol had found some weakness of AlphaGo.

    I do not care of David Cantrell’s word of ‘foolish’ but I do believe Lee Sedol had found some weakness of AlphaGo. David Cantrell can not deny the following factor no mater how clever he is:

    Lee Sedol had won the fourth game just because he had attacked the weakness of AlphaGO

    Also Lee Sedol had played much better in the last game compare with first three game

    Lee started the last game strongly, taking advantage of an early mistake by AlphaGo. But in the end, Lee was unable to hold off a comeback by his opponent, which won a narrow victory.

    Regard what David Cantrell’statement We can only learn what Alphago’s true weaknesses are if Google let it play lots of games in public. I hope they do, unlike what IBM

    deep with Deep Blue. I think in fact there are many Alphago’s game records which have not been open to public. If thease game records can be provided to Lee Sedol for him to do full research the Lee Sedol can find more weakness of AlphoGO and possible can win all five games

    Verbal Violence and Abusive language

    I suggest BGA write a guideline to ban the Verbal Violence and Abusive language in Gotalk

    I think the email from David Cantrell is abusive.

    I do not happy to be called foolish

    I think David Cantrell also will not be happy if some one point on his nose and say in the public that David Cantrell ‘s idela is foolish stupid or silly

    I am very sorry for the word ‘foolish’the email from David Cantrell

    I hope I am blind and can not read that word which made me very sad……..

    Mark Baoliang Zhang

    Message: 1

    Date: Tue, 15 Mar 2016 14:23:17 +0000

    From: David Cantrell

    Subject: Re: [Gotalk] Can Lee Sedol win the fifth/last game AlphaGO

    To: [email protected]

    Message-ID:

    Content-Type: text/plain; charset=us-ascii

    On Mon, Mar 14, 2016 at 10:34:11PM +0000, Baoliang Zhang wrote:

    > Lee has find the weakness of AlphaGo …

    I saw lots of comments like this online but I think it’s foolish to go

    from a single win to he has found *the* weakness. Or even a significant

    weakness. We can only learn what Alphago’s true weaknesses are if Google

    let it play lots of games in public. I hope they do, unlike what IBM

    deep with Deep Blue.

    David Cantrell | semi-evolved ape-thing

    Blessed are the pessimists, for they test their bac

Speak your mind