The future is here: A professional-level go AI

Thursday, January 28 2016, 7:30 AM in Tokyo, Japan.

I wake up to my alarm clock as usual, preparing to go to the Nihon Ki-in headquarters in Ichigaya for my first day as an official game transcriber. Before my morning shower, I turn on my computer and check the news headlines of the day:

>Game Over? AlphaGo Beats Pro 5-0 in Major AI Advance

“…”

“Maybe it was a low-handicap match?” *reads a bit* “Wait, seems that’s not it, they were even games by Chinese rules.”

“…No, it’s not April 1st yet.”

“Maybe it’s a marketing ruse?”

>Nature: Go players react to computer defeat

“No, that’s not it either, they have too many important people commenting on the match.”

“…”

“Wow, I really am living in the future.”

Just a few weeks ago, I gave a television interview to Yle, Finland’s national public-broadcasting company, confidently saying that go-playing AI would definitely catch up to top human players, but that it might yet take dozens of years. It seems I cannot help but correct my statement.

Fan Hui, a Chinese professional 2 dan now living in France, is perhaps not at the very top of the go world, but he is reasonably high up there. A go-playing AI defeating him 5-0 surely implies a playing strength comparable to professional players. In Fan Hui’s defense, too, the five games played were perhaps not the best show of his skill. Apparently five unofficial matches were also played, which Google’s AlphaGo won 3-2.

In March 2016, AlphaGo is set to play against Lee Sedol, widely considered the current strongest human player. My personal assessment is that Lee Sedol will still win the match, with 4-1 or 5-0 sounding like plausible outcomes. However, AlphaGo’s match with Fan Hui was played in October 2015, and they say the AI is constantly getting stronger, so I might yet have to correct myself about this prediction as well… And even if Lee Sedol wins the match, in a year or two the situation may get reversed anyway.

The news was noticed in Japanese professional circles as well. When I was setting up the computer for transcribing, I overheard two professionals talking along the lines of “Did you read the news?” “Yes, it seems it was a Chinese professional, I wonder how strong he is.”

Finally, in case the reader is interested in what I thought of AlphaGo’s play, here are some comments jotted down by me about the first game in the five-game series! Enjoy!

 

Download SGF

 

PS. Personally, I’m all for go-playing AI getting stronger than humans. Once the AI get strong enough that they don’t copy human tactics anymore, we’re really getting to find out what the game is about!

25 thoughts on “The future is here: A professional-level go AI”

  1. Good to hear your thoughts about this. I agree with your postscript: when we get a go AI that’s much stronger than the top professionals, we will start seeing things like new computer-discovered joseki.

    Still, the matches with Fan Hui and Lee Sedol are just a stuttering beginning. For now, the hardware requirements of AlphaGo are quite stunning (170 GPU’s, 1200 CPU’s), and as far as I know, all currents methods for scaling down the complexity of such “deep ANN’s” incur a significant loss in performance. As Moore’s law is now fading away, this cannot be fixed just by waiting.

    The “dozens of years” you mentioned might still be a realistic guess for how long it will take before you can run a pro-level go AI on your own hardware. :)

    1. Yes and no, software advances still beat hardware, and I’m pretty sure there is a *lot* to be improved in ANN techniques from what AlphaGo uses, not to mention adding compact heuristics that are discovered from experience with AI players.

      Someone in an IRC channel is testing a part champion Fritz against modern player Stockfish in chess, and Stockfish can beat Fritz with three pawn handicap 10 games out of 10, using tenth of the processing power.

    2. > 170 GPU’s, 1200 CPU’s

      Where’s that number from? I believe the running one against Fan Hui was 8 GPUs/48 CPUs (about $2/hour on Amazon EC2, incidentally). Training does require a lot more computing power, of course.

      > as far as I know, all currents methods for scaling down the complexity of such “deep ANN’s” incur a significant loss in performance.

      Cite? My impression was the opposite, that model compression of a final trained model offered staggering savings, like easily 80% of parameters/bits with no noticeable loss of accuracy, and this has been known for quite a while now (eg the 2006 paper http://www.cs.cornell.edu/~caruana/compression.kdd06.pdf ). Consider the savings reported in one of the most recent papers http://arxiv.org/abs/1510.00149 : “reduce the storage requirement of neural networks by 35x to 49x without affecting their accuracy…On the ImageNet dataset, our method reduced the storage required by AlexNet by 35x, from 240MB to 6.9MB, without loss of accuracy. Our method reduced the size of VGG-16 by 49x from 552MB to 11.3MB, again with no loss of accuracy. This allows fitting the model into on-chip SRAM cache rather than off-chip DRAM memory. Our compression method also facilitates the use of complex neural networks in mobile applications where application size and download bandwidth are constrained. Benchmarked on CPU, GPU and mobile GPU, compressed network has 3x to 4x layerwise speedup and 3x to 7x better energy efficiency.” I don’t know any particular reason that the NNs used in AlphaGo would not compress well, although I admit I’ve never seen anyone specifically establish that RL NNs can compress as well as more common tasks like image recognition.

    3. > the hardware requirements of AlphaGo are quite stunning (170 GPU’s, 1200 CPU’s)

      That’s the distributed version of AlphaGo. The one who won against Fan Hui is the single machine version: 48 CPU, 8 GPU.

  2. I was particularly interested in your assessment of the game, so thank you. It would be nice to hear your comments on the last four games as well. Judging from Fan Hui’s resignations and the way the games went and my limited understanding of the game, it looked a bit like Fan Hui had trouble orienting himself to perform on top level.

  3. Here are some comments from top chinese players:
    1. Ke Jie (world champ) – limited strength…but still amazing… Less than 5% chance against Lee Sedol now. But as it can go stronger, who knows its future…
    2. Mi Yuting (world champ) – appears to be a ‘chong-duan-shao-nian (kids on the path to pros)’, ~high-level amateur.
    3, Li Jie (former national team player) – appears to be pro-level. one of the games is almost perfect (for AlphaGo)

    1. Ke Jie is probably accurate. Based on estimated ELO ratings for Fan Hui and Lee Sedol, the chance of the former winning is about 3.1%. The question is really how much stronger than Fan Hui the gobot is.

  4. I agree, once the computer is stronger than human players, then we can start to learn from the computer on how to make our go better. Since go in more abstract than chess, then it will be easier to take away notions and concepts than the simple brute force strength that computers can offer in chess.

  5. Where can I find information about the inofficial matches? I have not read about them anywhere else. Where they with the same time constraints? Also one game per day?

    1. The inofficial games were played with 3×30 second byoyomi, without main time. There was one inofficial game per day, and AlphaGo won 3 out of 5. I think it’s interesting that the program clearly had more difficulties with the shoter time settings. It perhaps indicates that AlphaGo uses it’s thinking time very efficiently.

  6. What’s your opinion on Black 29? After Black 27, I thought black would invade the top to split white’s formation. Allowing white to link two groups with L17 seems too easy on white.

  7. Is there a 3 faction variant of weiqi?(black vs gray vs white) with same rule set as Google was following (on android would even be better ?

    1. Nami foo:

      Yes, there is. At least there is on iOS. I’m not sure about Android. On iOS, the app is called Go Toucher. It has a 3-sided variant like you describe which it calls “Gochaos.” I don’t know if it follows Google’s rule set though.

  8. Thank you for this review!

    With regard to its playing strength being only “low pro level”, it is important to note that AlphaGo _oblitarated_ the strongest bots even at 4 stone handicap, which even Yoda Norimoto 9p couldn’t do. Yes, Yoda Norimoto 9p is not Lee Sedol, but it’s another indication that the program could be stronger than it looks. And as everyone has been saying, that was in October.

    Also the AlphaGo Nature paper says the Neural Network (NN) for estimating the most probable next move was build using a high level KGS game database. Which might be why the feeling of “strong level amateur” might come from. I am eager to see if it would show even more brilliance if the NN was trained on the best pro kifus.

    I just love this achievement for all the possibilities it offers. And not just for the understanding of the game. You could teach AlphaGo different playing styles based on the kifu databases used for the NN. Hikaru no Go raised the question of “how would Honinbou Shusaku fair against modern pros?”. Well now we might know!

  9. Moore’s law may be broken, but that’s not relevant to this problem in my opinion. While it may not be possible to create more powerful CPU’s as we have done in the past, it’s possible to use more cores. I think this problem should be suitable for such parallel calculation. I could be wrong of course, as the details of AlphaGo’s algorithms are not clear to me.

    In any case, I believe this is a major breakthrough and something I definitely didn’t expect to happen so soon. As for the new joseki, I pose a question: how many new moves and insights have modern super strong chess computers revealed? It may take a long time for a computer to develop the game in a manner human experts have developed go strategy. Or it might happen fast, we’ll see in a few years :-)

    1. In chess provided very little insight into game. Some details like g2-g4 might playable more often. Biggest insight is that concrete variations are far more important than general principles.

      Obviously opening theory advances faster now as just about anyone can do research on them.

      In Go it might be hard to extract even that much. New Joseki -dunno , joseki is joseki only if it matches surroundings? And new move by MonteCarlo/DCNN could be just artefact of concrete variations and hard to evaluate by human.

      but as training tool, tool for preparing agains some particular setup, sure

Leave a Reply

Your email address will not be published. Required fields are marked *

Please confirm you are a human by solving this: