We'll keep on fighting 'til the end.
March 22, 2017 4:29 PM   Subscribe

The World Go Championship is underway in Japan. Organized by the Nihon Ki-in, it is the first international tournament that will include both professional human players and an artificial intelligence Go-playing program. 1st-day coverage is up on youtube, with commentary from Michael Redmond and Antti Tourmanen (commentary begins at 2:53).

Neither the top-rated (by one estimate and some public opinion) human player nor what is widely believed to be the strongest AI will be participating.

Since AlphaGo demonstrated it was possible last year, a number of other organizations have developed strong go programs using neural networks. Several of them competed in the recent UEC Cup, with Fine Art (aka JueYi, from Tencent) being the winner of that computer Go contest.

Among the computers that have been turning up on internet Go servers recently, one stood out as undefeated in 60 quick online games against various strong players. Known as "master" or "magister", it was subsequently revealed to be a version or descendant of AlphaGo. The record of those 60 games is now the largest collection of game records we have from the Deepmind AI, so even though they are played at very short time settings -- some 30 seconds per move, some only 20 -- these games have been the subjects of much commentary. Some games are viewable on eidogo via reddit, or you can download all 60 games in sgf format. There is also a new AGA video review series. Some other video commentaries from the usual suspects are listed at Sensei's Library. Extrapolating from its rapid rate of improvement between its games in 2015 and those in 2016, and considering its undefeated record, it is difficult to guess how good at this game AlphaGo might be by now. It may be completely unbeatable already, or its performance might suffer at time settings more suited to human professionals.

The free programs available for download to play go against home computers have also become much stronger in the past year. Leela (easier to use) and Ray (open source) both appear to be capable of play at amateur mid-dan ranks on a fairly average PC. That is not good enough to challenge a professional, but it is suddenly much closer than a year or two ago.
posted by sfenders (11 comments total) 19 users marked this as a favorite
 
There was about a ten year period where chess competition between top grandmasters and elite programs was very dramatic and interesting. Then programs became, more or less, unbeatable by human players. In 1998, I watched a strong chess player just crush Fritz 5.32, the best commercial PC program at the time; within about 3 years, none of the stuff he did to it would work against the next generation. A lot of that was faster hardware, not programming breakthroughs, although they did continue to find better search techniques and evaluation heuristics.

As I understand it, the neural net Go programs work completely differently than the alpha-beta search tree pruning that works for chess, so I don't know how much you can draw analogies in predicting the course of their abilities. I hope there's a nice period where human-AI Go competition is meaningful.
posted by thelonius at 4:47 PM on March 22


I hope there's a nice period where human-AI Go competition is meaningful.

It's not now, really. If Alpha Go was released in its current state it could easily beat any pro player. But presumably there are ways to cripple it and bring it down to human level (I don't know whether it would require completely new training to be useful at multiple difficulties).
posted by dilaudid at 4:52 PM on March 22


As a fan of Go, I find the conversation about computers becoming more powerful to be a little dispiriting. It seems like such a challenge, a game even, for programmers to play at, where defeating humans is the objective. I don't condemn it at all - progress marches on, after all, and I'm sure useful or useless innovation will come out of this. But, it seems like we automate and speculate so much of our lives away. I would much prefer global commentary to be delivered in such a way that more people started playing Go, in parks and cafes and casual leagues , instead of focusing on how now a single program is more powerful than every human on the planet.
posted by rebent at 8:04 PM on March 22


Honestly, I feel like the AlphaGo/Sedol matches probably did a lot to raise the mainstream profile of Go in the west, although something like Hikaru no Go might have inspired more people to actually pick up the game, since it's more inherently instructive to new players.
posted by Wandering Idiot at 8:22 PM on March 22 [1 favorite]


Thanks for the post, I have been following the developments pretty closely and it's all getting very interesting. I have always been interested in go, but the AlphaGo vs Lee Sedol match pretty much pushed me to actively learning to play. I still haven't managed to muster the nerve to play humans yet, just the SmartGo app (anyone wants to play? OGS?). Master/Magister is surely a reference to Herman Hesse's The Glass Bead Game?

As a fan of Go, I find the conversation about computers becoming more powerful to be a little dispiriting.
I find it extremely interesting. As far as I can tell, AlphaGo plays moves that humans would normally discard, I think mostly because there are established ideas on what is a good move and what is not. AlphaGo is showing us a different way to play, and that is super interesting. Also there's no denying that the profile of Go has been raised worldwide, and that there are more people interested in playing than before. So ironically, it may actually lead to more people playing Go in parks etc.
posted by dhruva at 10:08 PM on March 22 [1 favorite]


But presumably there are ways to cripple it and bring it down to human level

Not really necessary, Go has a handicap system to even out the playing field.
posted by dhruva at 10:09 PM on March 22


anyone wants to play?

I'll play you, or anyone who wants to, assuming that our timezones intersect somewhere in a convenient way. I was raised by computers too. Memail for KGS or OGS info.
posted by sfenders at 5:26 AM on March 23 [1 favorite]


It's not now, really. If Alpha Go was released in its current state it could easily beat any pro player. But presumably there are ways to cripple it and bring it down to human level (I don't know whether it would require completely new training to be useful at multiple difficulties).

I suspect that your desktop computer won't put up much of a fight for a few years. From this paper (see table 6) AlphaGo is superhuman when you throw more than 1000 CPUs at it (in the table, "Distributed AlphaGo"). With tens of CPUs, it's good but beatable ("AlphaGo"). With just the one or two CPUs you've got on your desktop, who knows?
posted by Jpfed at 7:27 AM on March 23


more than 1000 CPUs

Well, the info in last year's paper describing it may already have been out of date by the time AlphaGo played Lee Sedol. It played more strongly than would be implied by the ELO rating in table 6 there, and we don't know how much of that is due to them throwing more CPUs and GPUs onto the fire, how much due to bug fixing and design improvements, or whether they simply used the entire power output of several small countries for a couple of months to better train the neural networks. What kind of hardware Master might have been running on in January is even more entirely unknown to the public. Perhaps they're seeing how it scales up to 100,000 CPUs. Or maybe it's gotten so much more efficient that it runs on someone's iPhone. Or maybe they were testing out human-computer team play.

Hideki Kato has apparently claimed that DeepZen could be only one stone weaker than its current top pro equivalent level on an "eight-core Intel with GTX-1080". That would make it perhaps two stones weaker than Master, on a machine that could be set up for $2000 or so. I don't know how likely that is in reality, it sounds too good to be true. At least the new Zen is sure to eventually be released as a commercial product, so we'll find out soon.

Zen did well in the tournament and in other games recently, but it seems that the program suffered from some already-identified bugs that hindered its performance, which sound as if they should be possible to fix. Its problem with Japanese scoring and komi (as opposed to Chinese which is easier for computers) may have made it stronger when playing black: The value network misunderestimating the komi could make it lose close games.

Park Jeong-hwan, official world champion Go player, was quoted as saying "My toughest match was against DeepZenGo. I imagine that humans will only be able to contend with AI for another five years or so." It may be so, but I wonder what new plays the humans might come up with once more than a few have the chance to make more than a handful of attempts.
posted by sfenders at 6:15 PM on March 24


I want to see professional human commentary on top level Computer vs Computer games.
posted by Theta States at 7:17 AM on March 31


Exploring the mysteries of Go with AlphaGo and China's top players: That’s why we’re so excited to announce AlphaGo’s next step: a five-day festival of Go and artificial intelligence in the game's birthplace, China
posted by dhruva at 2:19 AM on April 10 [1 favorite]


« Older “Alger is to America what Homer was to the Greeks"   |   But you didn't say what type of beans are on the... Newer »


This thread has been archived and is closed to new comments