Go master Cho wins best-of-three series against Japan-made AI

http://www.japantimes.co.jp/news/2016/11/24/national/go-master-cho-wins-best-three-series-japan-made-ai/#.WDdQUUnhXqA

by happy-go-lucky • 10 years ago

130 77 10 years ago

japantimes.co.jp

conanbatt3• 10 years ago

This news has flewn way under the radar in comparison to the previous Match.

Cho Chikun is a legendary player ,and was top of the world in the 70's. From that era, along with Takemiya Masaki, he was one of the only ones that kept playing and kept winning titles even in his old age.

The style of 70's players was artistic and not as sports-like as koreans were when they rose in the 90's. They always showed a higher level of philosophy and art in game, but it was ultimately crushed by the "machine-like" tactical capacity of korean players. Cho chikun has very little chance to beat Lee sedol. But he was confident on beating Zen, and he did something unspeakable in the LeeSedol Matches: he came back from a losing game, from a machine that looked that would never relent its edges.

As a retired semi-professional Go player that learnt by replaying games from Cho Chikun, it is exciting to see the Japenese artistic and humane vision dominate the computer over the Korean Tactical sports-minded way.

espadrine2• 10 years ago

Isn't it plausible that DeepZenGo simply isn't at the level of play of AlphaGo? Facebook's DarkForest is also a competent AI, but even it does not compete yet.

gcp6• 10 years ago

It's hard to compare, as even Deep Zen runs on much more...modest...hardware than the match version of AlphaGo did. It's conceivable that AlphaGo would get clobbered on Zen's hardware, or that Zen would clobber AlphaGo if that team had Google-eseque resources available. But there's no reason to believe we'd ever see such a fair comparison. It makes no sense for either team!

Zen is commercial software sold to people with a normal PC, so obviously reaching AlphaGo performance on AlphaGo hardware does their customers no good either.

It's for sure stronger than Darkforest. Some free PC programs like Leela have reached Darkforest levels on normal PC hardware (Darkforest used 44 NVIDIA Tesla cards, Leela uses a single normal "gaming" GPU), and Zen is several stones stronger than Leela.

tasuki• 10 years ago

Not just plausible, it's pretty damn sure.

Still, major props to Hideki Kato, who's developing cutting edge AI at the age of 62. I wish to have his drive at that age!

thomasahle• 10 years ago

Computers losing to Machines in Go is not yet that surprising. The article doesn't really give any reasons to why we should have expected something else. Perhaps the main news worthy part was that the machine actually won a game. But then we don't really know how strong its opponent is today.

gcp6• 10 years ago

we don't really know how strong its opponent is today

He's not exactly retired, you know: https://www.goratings.org/players/98.html

gcp6• 10 years ago

The news flies under the radar because for most people Go is now solved (compare to what happened after Deep Blue's win). There's a small blemish in that Ke Jie is still rated higher than AlphaGo. Perhaps they will rectify that, perhaps not.

Cho Chikun had the large advantage over Lee Sedol that his opponent (Zen) and similar programs (Crazy Stone, Leela) are simply available for him to practice against and probe for weaknesses.

Then of course there's also a significant reduction in the speed of the hardware.

Both of those matter a lot. I won't make a firm statement how relevant the style argument is you mention, but really, don't underestimate those two.

msbarnett• 10 years ago

The news flies under the radar because for most people Go is now solved (compare to what happened after Deep Blue's win).

"Machines reaching a level of play that is consistently better than humans" and "the game is solved" are two completely different statements and it isn't particularly useful to conflate them.

We will likely never be able to solve Go in the formal sense.

hellbanner2• 10 years ago

Speed of the hardware is significant to discuss. A human, given sufficent notebook paper & time, could think ahead for thousands of years and thousands of turns -- this is effectively what a machine is doing.

With much tighter time constraints, I don't believe machines would do as well because they are using full-breadth search (to my knowledge) -- searching every possibliity. Whereas humans use visual pattern recognition, and we can run this programs quite fast.

YeGoblynQueenne2• 10 years ago

> Cho Chikun had the large advantage over Lee Sedol that his opponent (Zen) and similar programs (Crazy Stone, Leela) are simply available for him to practice against and probe for weaknesses.

That's an argument against Go having been solved by AlphaGo, not for. If Lee Sedol could beat AlphaGo given the opportunity to practice against it, then nothing is solved.

Also, for a system to dominate against humans in a game it must be able to beat any opponent every time. AlphaGo hasn't demonstrated this yet, so while in the popular opinion AlphaGo "dominates", in the technical sense it does not, and Go won't be "solved" conclusively until it has.

YeGoblynQueenne2• 10 years ago

Something else- today you can run programs that play chess at superhuman levels on your phone. At the time of DeepBlue's win against Kasparov the same kind of system could run on off-the-shelve hardware [1]. Very soon after AI programs running on ordinary laptops demonstrated dominance against human masters (that's in Russel & Norvig, somewhere, I can dig it up if you like).

So the fact that Cho Chikun won against a less powerful machine is not necessarily an indication that the entire system is weaker than DeepMind's one. To know that you'd have to run Zen against AlphaGo.

______________

[1] A big hint of why it didn't is that, at the time, IBM was still selling more hardware than software.

brilliantcode2• 10 years ago

There's a small blemish in that Ke Jie is still rated higher than AlphaGo.

This is false. Ke Jie is far far below Lee Sedol.

Lee Sedol has no equal. Well, AlphaGo beat him but we are talking about player ranks.

nv-vn• 10 years ago

This sounds a lot like video games like StarCraft and Overwatch today. Korean competitive gamers have achieved the same kind of machine-like skill. In the Overwatch World Cup this year, not only did the Korean team go undefeated, but they massively crushed their opponents (even in the final game!).

taejavu• 10 years ago

There's even a joke that the currency of prize money for all international StarCraft tournaments should be the Korean Won.

convolvatron• 10 years ago

i play go, nowhere near the retired semi-professional level

my backwater go community doesnt really care about computers because we're too focussed on trying not to be so lousy, and on the conversation. the light chase.

i'm super curious...when you see alpha against sedol, or whatever this is against chikun. do you read something there? is there a character?

conanbatt3• 10 years ago

About 4 years ago, a friend i know used a 3-4d level computer against me for a fraction of his game. I noticed something was off, i actually thought he was being substituted, but it didnt cross my mind it was a bot. However, it was a very tactical situation.

Looking at alpha, something similar but at higher level happens: it looks so profoundly tactical that its hard to catch character from it. I can tell how Lee felt however: like i felt playing people close to his level, like you play a wall that you cant read.

chunsj• 10 years ago

Cho is a korean as well. Though I'm not sure on his current nationality. Yes, there might be some difference in the style of playing Go, which I do not know the detail.

conanbatt3• 10 years ago

For the purposes of Go, he is japanese. He learnt Go from Kitani Minoru from age 5-6, and he became a professional player at 11~ breaking records for his time (if I recall correctly). Back in the 70's japan was everything for Go.

fiatjaf• 10 years ago

What happened to chess enthusiasts when chess was "solved"?

exolymph• 10 years ago

They kept playing chess with each other.

cybertronic2• 10 years ago

and chess was not "solved"

jm547ster• 10 years ago

Computer aided chess became a game in and of itself

olalonde3• 10 years ago

The best computers now consistently beat the best humans but "solved" has a precise definition in game theory and it isn't completely solved under that definition[0]. With regards to your question, I can't say for sure but my impression is that interest in chess has declined a lot.

[0] https://en.wikipedia.org/wiki/Solving_chess

andrepd• 10 years ago

With regards to your question, I can't say for sure but my impression is that interest in chess has declined a lot.

It hasn't, in the slightest. The fact that computers playing at full strength can beat top humans makes no difference in the enjoyability of the game for people.

espadrine2• 10 years ago

my impression is that interest in chess has declined a lot.

I don't think it has. Much like horseback riding has not made olympic races obsolete, human competition remained fascinating. I for one very much followed the rise of Magnus Carlsen.

imaginenore• 10 years ago

That's an understatement.

The best computer algos beat humans at chess by a huge margin, to the point a regular PC beats Carlsen (and I heard even the best smartphone apps beat him).

Carlsen's ELO rating = 2853

Top chess program ELO rating = 3381

The number of DIFFERENT programs stronger than Carlsen = 59

http://www.computerchess.org.uk/ccrl/4040/

https://ratings.fide.com/top.phtml?list=men

partycoder2• 10 years ago

Solving chess would mean having the perfect move as a response to every move. In that sense, chess is far from being solved.

brilliantcode2• 10 years ago

Cho won a close game 1,[28] lost game 2 when his invasion into enemy territory was killed,[29] and won game 3. Zen uses neural network techniques similar to AlphaGo,[30] however ran on more modest hardware during the match.

not sure if suggests any comparison between AlphaGo and Zen. I'd like to see Cho play against AlphaGo.

iopq• 10 years ago

Cho Chikun would not stand a chance against AlphaGo. AlphaGo has shown itself to be stronger than Lee Sedol, who is much stronger than Cho.

partycoder2• 10 years ago

It depends on many things... such as the game data used to train it and the hardware in which it ran, the time settings of the game, the implementation of the algorithm, etc.

But so far commentators such as Ke Jie (higher rating than Lee Sedol), said that Deep Zen Go's yose (end-game) is very weak compared to Alpha Go.

Another factor is that Cho Chikun had the advantage of studying more MCTS/DL bot games. Today you can find MCTS/DL bots in servers such as KGS that you can play against.

Cho Chikun had more clear expectations of the bot playing level. e.g: Lee Sedol greatly underestimated Alpha Go initially, then he was in shock and did not know what to expect. Your playing strategy can vary depending of strengths and weaknesses of your opponent.

deegles2• 10 years ago

What bothered me about AlphaGo was the sheer amount of power that was dedicated to training and running it vs. the relatively puny power budget a human has.

i.e. "Lee Sedol used about 20 Watts of power to operate. By contrast, AlphaGo runs on a whopping 1920 CPUs and another 280 GPUs for an estimated power consumption of approximately 1 MW (200 W per CPU and 200 W per GPU). That’s 50,000 times as much power as the amount of power that Lee Sedol’s brain uses and the two are not quite evenly matched but it is close enough to use for comparison."[0]

I mean, sure, it's apples to oranges, but it feels like a pyrrhic victory to have to use 1MW of computing power to beat a 20w brain.

I want there to be a series of AI vs. AI games where everyone has an equal power budget for their AI. Everyone has to bring their own equipment and they receive exactly X kwh from a calibrated power supply to train their model. Training data could possibly also be standardized.

Each team would received a standardized power supply for the actual competition (e.g 20w). There could be bonuses for using less energy for training or during gameplay, or for faster response times between moves.

I believe this would drive innovation in building special purpose hardware, more efficient training models, and so on.

[0] http://jacquesmattheij.com/another-way-of-looking-at-lee-sed...

rcthompson• 10 years ago

Everyone has to bring their own equipment and they receive exactly X kwh from a calibrated power supply to train their model

What's the advantage of this over simply providing every team with identical hardware and equal computing time? It seems like providing a fixed power supply would favor whoever can afford to buy more efficient computing equipment.

meric• 10 years ago

Measuring by power supply allows human participants if the limit is 20W.

deegles2• 10 years ago

Developing special purpose machine learning hardware is part of the goal. I think of it like Formula 1, where the R&D eventually trickles down to consumer hardware.

Also, there's no reason why you couldn't have different competition categories where you're limited to using X dollars worth of time on a specific cloud provider, for example.

TwoBit• 10 years ago

Identical hardware limits you.

arcticfox• 10 years ago

why optimize for power? You can't create a megawatt human, that's our limitation that is being shown in these matches. power would just be a weird artificial constraint.

monix2• 10 years ago

Even though an argument can be made about computers being able to scale, energy is always the only constraint that matters.

monix2• 10 years ago

On the other hand AlphaGo proves that computers can scale, even if they are inefficient, whereas the only way to grow that 20 Watts brain would be to use multiple brains and that's not going to be any good in Go matches.

argonaut• 10 years ago

On the contrary, I think there is a real question over whether AlphaGo would be able to beat a group of 3-5 tier-1 Go players (e.g. ~top 10 rankings players) who were allowed to write things down on paper, discuss, use whiteboards, use separate game boards to play variations, etc, even with same time controls.

mrob• 10 years ago

Lee Sedol did not use 20W, because he is not a disembodied brain. A non-exercising human is usually approximated as using 100W.

shard• 10 years ago

Another difference between the human brain and a computer is the number of round of hardware optimization. Human brains have millions of years and millions of search paths explored, while computer hardware, although advancing at a much faster pace, is still a relatively new technology.

Radim• 10 years ago

"Meh" news because we had a higher profile match just recently: AlphaGo vs Lee Sedol (both much stronger than DeepZen or Cho).

I mean, it's mildly interesting that DeepZen uses fewer resources (just ~40 CPUs and 4 GPU, compared to AlphaGo's ~2k CPU and ~200 GPUs, IIRC). But that's relevant only to hardcore Go enthusiasts who would want to run such AIs on their own hardware.

In contrast, the next interesting milestones in Go AI are:

* AI beating a team of cooperating top pros, say top-10 humans vs AI (easy)

* solving Go (hard)

* finding ways to translate insights gained from observing strong Go AI into real-life situations relevant outside of Go (my personal favourite; Go has that fascinating simple-meets-complex abstracting property [1])

[1] https://rare-technologies.com/go_games_life/

emptybits• 10 years ago

* solving Go (hard)

1. Perfect play from any game position.

2. Proof of perfect play.

#1 may be achievable with a sufficiently clever algorithm. (We may suspect it is perfect and may only later prove it.) #2 may be much much harder. Neither are suitable for brute force approaches to play or proof as some problem spaces are.

adrianN• 10 years ago

I think the next interesting milestone would be beating top rated humans using commodity hardware. In chess for example, programs running on cell phone processors have Elo rating around 2800. Using weaker hardware means that the algorithms need to have more insight into the game and can't rely as much on exploring the game tree.

qznc• 10 years ago

The next interesting non-Go milestone seems to be Starcraft ;)

pelario• 10 years ago

* AI beating a team of cooperating top pros, say top-10 humans vs AI (easy)

It is not clear at all that the cooperating team is more difficult to beatä

* solving Go (hard) Not only hard, but 'solving Go' it is beyond AI, and probably cannot be done.

Here I fully agree, as conan noted in another thread, stylistic factors are quite important, and the art-like approach is a thing in Go, so I would love to know "what the AI thinks about it" :)

bluetwo2• 10 years ago

OK, can't we just have Zen and AlphaGo play each other?

gcp6• 10 years ago

What would the upsides of such a match be for AlphaGo?

How many matches with other computers did Deep Blue play after its victory over Kasparov? (zero)

How many world championships did Deep Blue win? (zero, it played one in 1995 and lost to Fritz)

How many people tought Deep Blue was the strongest computer in the world in 1996? (All of them, because IBM said so and conveniently forgot about the lost world championship)

Google has already said AlphaGo is much stronger than other programs. There's no point in playing a match then, is there?

bluetwo2• 10 years ago

Errrrrr.... It would shut up Zen go or prove them right? That seems like a reason.

Go master Cho wins best-of-three series against Japan-made AI

Related Stories