AlphaGo vs Lee Sedol, who will win?

[seigenblues]

This question might be partially answered in the paper itself.

According to this interesting figure:


...

There's a lot of folks assuming the AG will get much better w/ 6 months of extra work. The history of go AI has been large jumps interspersed with long plateaus of nominal improvement. Why do people assume this is different and that 6 months will be 6 months of linear progress?

If anything, look at how quickly the scaling improvements seem to level out. 280 GPUS is only slightly stronger than 112.

Fan Hui's rating on goratings is 2900. Lee Sedol is 600 points higher, 3500! By their own chart AG distributed is ~3100 -- and the paper mentions they anchor their Elo scale to Fan Hui.

Lastly, i have to observe that all computer programs have bugs. If AG has no bugs, the Google press release should've been about how Deepmind figured out how to write bug-free software instead

[Solomon]

This question might be partially answered in the paper itself.

According to this interesting figure:


...

There's a lot of folks assuming the AG will get much better w/ 6 months of extra work. The history of go AI has been large jumps interspersed with long plateaus of nominal improvement. Why do people assume this is different and that 6 months will be 6 months of linear progress?

If anything, look at how quickly the scaling improvements seem to level out. 280 GPUS is only slightly stronger than 112.

Fan Hui's rating on goratings is 2900. Lee Sedol is 600 points higher, 3500! By their own chart AG distributed is ~3100 -- and the paper mentions they anchor their Elo scale to Fan Hui.

Lastly, i have to observe that all computer programs have bugs. If AG has no bugs, the Google press release should've been about how Deepmind figured out how to write bug-free software instead

I don't think it will get *much* better (well okay, maybe the way I phrased it in my prediction post earlier says otherwise...), but that it only needs just the slightest boost to get there. However, I do really think some people are blowing up the difference between Fan Hui's level and Lee Sedol's level, one factor being the large ELO point difference on goratings.org which I don't consider to be such a reliable measure of skill difference (Andy Liu 300 points higher than Fan Hui, and higher than Yoda Norimoto, Nie Weiping, and Zhou Zunxun? Yuki Satoshi higher than Lee Changho, Cho U, Chang Hao, and Kong Jie?).

[hyperpape]

Cho U is in much worse form lately. GoRatings and mamumamu0413 agree. Yuki Satoshi probably is stronger (and go searching for John F's posts on Yuki).

Andy Liu benefits from having four recent wins and only five games, which is pretty awesome, but his rating is wrong for predictable reasons that don't discredit the whole system.

Now, you might wonder whether Fan has many games? He doesn't, but you could take a basket approach. Look at all the Western players: Andy Liu, Dinerstein, Gansheng Shi, Lisy, Surma, Shikshin, Yongfei Ge, Eric Lui, Daehyuk Ko, Jabarin, Huiren Yang.... and a few others I've missed. Andy is the only one to have a rating above 3000.

From what we know in European competition, Fan could be the strongest of those players. I'd give that more credence than his ratings on the goratings site. But that still wouldn't suggest a rating significantly above 3000. That's a smaller distance behind Lee, but still a huge one.

[Uberdude]

@Calvin (I see no point in hiding)

1. I don't think the funders of the Google Deepmind project have any intention of giving Lee a million dollars, no matter how much they'd like to be VIPs at the bar he will certainly open when he retires (again).

Google is very rich. They splashed $400 million on buying Deepmind, and even more on a fancy themostat (Nest). $1m is pocket change. This announcement has already gained them >$1m worth of press coverage and helped cement the image of Google as a bunch of clever clogs and will encourage smart people to want to work there.

I agree with your other 2 points though, but this lack of information isn't enough to make me think Lee will lose, though I wouldn't want to bet a lot on it.

[Calvin Clark]

@Calvin (I see no point in hiding)

1. I don't think the funders of the Google Deepmind project have any intention of giving Lee a million dollars, no matter how much they'd like to be VIPs at the bar he will certainly open when he retires (again).

Google is very rich. They splashed $400 million on buying Deepmind, and even more on a fancy themostat (Nest). $1m is pocket change. This announcement has already gained them >$1m worth of press coverage and helped cement the image of Google as a bunch of clever clogs and will encourage smart people to want to work there.

I agree with your other 2 points though, but this lack of information isn't enough to make me think Lee will lose, though I wouldn't want to bet a lot on it.

I can't discuss marketing behavior of multinationals without a stiff drink, and I am out of scotch, so you win.

[pookpooi]

We've weak evidence (from Demis Hassabis's go teacher) that AlphaGo is around british 3D (maybe 4D KGS) around February, we now know that it's around 9D in October. What we don't know is the AlphaGo growth rate, it maybe constant with time, exponential (scary) or just plateau out in the end. So the March match is very important to keep track of what's going on, and to predict what will come in the future.

But I'll guess this just for fun

Exponential Rate --> unlikely, but if it does happen, more than 14D in March, it's immeasurable using the traditional ranking system
Constant Rate --> around 13D, enough to crush Lee Sedol to 5-0 though
Plateau --> around 11D, the outcome might go either way but I'll bet AlphaGo win at 3-2

[Charles Matthews]

But I'll guess this just for fun

Yes, extrapolation is fun - not so much more.

I have just been watching the AGA video with a Korean 9p saying AlphaGo plays with a "Japanese" style. If this is attributable to the training set of games being skewed towards Japanese pros playing online ... then from some points of view there is a potential increase in strength by redoing all that with a more broadly based set of games.

Well, one could see why this would be an obvious move, for a Korean pro!

Looking at it another way, versions of the AI trained on different sets might show fine differences of level. Which could perhaps be calibrated by reference to pro ranks.

So, only a small amount of external consultancy might have suggested some path or paths to retraining, and I don't suppose we can guess the resulting increment. Traditionally, a third of a stone means plenty in terms of pro levels.

Charles

[Mike Novack]

We also don't know if we have any valid data on which to base that "growth rate".

The machine(s) used for training (what the program is running on for training and testing) not necessarily the same throughout that time. We don't that 3-4D British in February and 9D in October was on the same machine.

Go is played with time controls, time management is part of playing go, and machine power trades with time. We mustn't confuse "program" with "program + machine". Google is willing and able to throw money at the problem, and when running in the October contest used MUCH MORE machine than we have seen the other programs using in the periodic computer program vs computer program contests or for the bots playing on KGS. They can afford the "iron", the others can't.

When we say "this PROGRAM is stronger that that PROGRAM" we should really be adding "when both have comparable computer resources available to them".

I do think this new program IS stronger than the existing MCTS progams. But I think the "how much stronger" is being distorted by the hardware disparity. Thus while I will be watching the outcome of the upcoming match with great interest, I would actually rather see a contest between this program and say AYAMC5 both running on say a single i7-4790 machine. Better? How many stones better? << that would be a machine significantly more powerful than most of our home computers, but a serious gamer might have something like that >>

[mhlepore]

I have just been watching the AGA video with a Korean 9p saying AlphaGo plays with a "Japanese" style. If this is attributable to the training set of games being skewed towards Japanese pros playing online ... then from some points of view there is a potential increase in strength by redoing all that with a more broadly based set of games.

Charles

Fan Hui said AlphaGo has a peaceful style - I guess that means the same thing.

I wonder if there is something inherent about a strong AI program that causes it to be more peaceful. Possibilities:

- Perhaps what we view as messy just isn't that messy to AlphaGo, and it simply chooses the most straightforward path to a good outcome.

- Perhaps playing to its strengths means keeping a small lead is good enough, because it is probably unflappable in the endgame.

[Uberdude]

I think there is a bit of confusion over what the training games for the network was. It seems to me it was KGS 6-9d games, rather than pro games. That strikes me as a bit odd as I'd have thought the higher quality of play in pro games in GoGoD, go4go, some bulk download of Tygem 9d(P) games etc would lead to a stronger AI but perhaps there was not enough quantity. In the AlphaGo paper it says:

We trained a 13 layer policy network, which we call the SL policy network, from 30 million positions from the KGS Go Server. The network predicted expert moves with an accuracy of 57.0% on a held out test set, using all input features, and 55.7% using only raw board position and move history as inputs, compared to the state-of-the-art from other research groups of 44.4% at date of submission.

Most KGS 6-9d will be amateurs, though perhaps at the 9d level more are pro. But I don't think there are many Japanese players there, in fact most of the known top KGS players are European/American/Chinese (rapyuta was a Japanese insei iirc). It could be that KGS has a more 'Japanese' style than Tygem though.
Myungwan Kim and Andrew Jackson in their review (https://www.youtube.com/watch?v=NHRHUHW6HQE very interesting, great job guys!) seemed to think it was trained on GoGoD pro games, which would indeed probably have more Japanese pro games, but they may be mistaken.

macelee tells us he gave them a go4go dump in December 2015, so it's possible that the version that beat Fan Hui was only trained on the relatively weak KGS 6-9d games and the one that will face Lee Sedol in March will be trained on a higher-quality diet of professional games and be even stronger. Gosh.

I wonder how different the network in AlphaGo is from the one a few of the DeepMind folks described in their earlier paper:
http://www.cs.toronto.edu/~cmaddis/pubs/deepgo.pdf
When we discussed that paper at end 2014 / early 2015 there was some scepticism over what a move prediction accuracy of 55% meant when it came to playing strength. Perhaps back then their policy network plus MCTS was around 3d and then the development of the value network (plus refinements) since then has given it the boost up Fan Hui beating strength.

[pookpooi]

Most programmers know their own plateau point of their program (where throwing a better hardware doesn't make it stronger), which is why in all computer Go tournament (as I know) they prefer to bring their own hardware. I'm aware that in chess there's a single system competition though.
In their research, they give Zen 8 cores and Crazystone 32 cores, which is fair enough because they can only acquire the commercial version which only scale at that point.
Mogo use to run on up to 3200 cores supercomputer, but it became a solid evidence that MCT is not scale well on high performance computer.
If AlphaGo happen to scale well on supercomputer, then it's a program strength too! And I've no problem with Google throwing their computer farm at it, because they can. However, according to the paper, the alphaGo reach the plateau as well (they give 700 more cpu core and 100 more gpu core but the ELO only increase from 3140 to 3168)

[uPWarrior]

I don't understand all this scalability non-sense. It is clear in the paper, and in the previous page of this topic, that single-machine AlphaGo plays at Fan Hui level, winning the majority of 4 stone handicap games against Zen, etc.

This is not a matter of throwing more machines at the problem.

[Mike Novack]

I don't understand all this scalability non-sense. It is clear in the paper, and in the previous page of this topic, that single-machine AlphaGo plays at Fan Hui level, winning the majority of 4 stone handicap games against Zen, etc.

This is not a matter of throwing more machines at the problem.

A Boeing 747 and a Piper Cub are both "single aircraft". But there is a vast difference in their load carrying capability. That's a very big "one machine" compared to what everybody else is using.

By scalable, means a more or less linear change in performance with changes in computation power. Probably neither is very scalable. Nor do I know how close Zen is to the "diminishing returns" point. But I do know for one of the other MCTS programs. Running on an old Lenovo T61, MFOG12 plays at about 1 dan KGS. Give it about 10 times more crunch power like what the bot on KGS has and is is 2-3 dan. However make that another 10 times, maybe would go up only half a stone (that's what I mean by diminishing returns).

[thirdfogie]

I voted for Lee Sedol by 3-2, though I hope he wins 5-0.

This opinion is worth little, since I am 4k at Go and it is 25 years since
I worked with a simple neural net. And as many have observed, we can only
guess at how much the program and its computer resources will improve by March.

There are two reasons I favour the human this time. Perhaps the human
must be good enough to find one or more valid tesuji or joseki innovations
that were not in the program's training set. Lee Sedol may be capable of that,
Fan Hui less so. Even world-level human players make mistakes (in the judgement
of their peers, or find them when reflecting on their games), which is why
the human may need to play more than one unprecedented masterstroke per game
to beat a program which never blunders.

The other reason is that analysis published by the British Go Association
http://www.britgo.org/files/2016/deepmind/BGJ174-AlphaGo.pdf criticises
several of AlphGo's moves, so it is not in fact perfect either. Comments by Younggil An
also identify mistakes by the program in Game 5, while also noting that it played
the endgame perfectly.

On the other hand, Lee Hajin has pointed out that AlphaGo may only have played
as strongly as was necessary to win, and it might automatically play better
against a stronger human.

So, 3-2.

I do hope that the match terms are clear and are clearly followed, so we do
not get the kind of bad taste that was left after Kasparov complained about
his loss to Deep Blue. For example, the terms should allow the human to rest
or sleep, and the computers to be restarted if they crash or the power fails, but
they should not allow the program to be adjusted on the fly by a secret cabal
of experts. (Whether such adjustments actually happened against Kasparov or
would even be possible this time is beside the point.)

[Tumtumtum]

I think Lee should go for a very weird opening right from the start. The engine wouldn't be able to deal with it with the current data set. I voted 3-2 for AlphaGo.