10 ms·
Go master Cho wins best-of-three series against Japan-made AI
- conanbatt 10y agoThis 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.
- espadrine 10y agoIsn'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.
- gcp 10y agoIt'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.
- sams99 10y ago"Zen is commercial software sold to people with a normal PC" Note, it looks like only version 6 is out: https://book.mynavi.jp/ec/products/detail/id=62646 https://book.mynavi.jp/ec/products/detail/id=62646 It appears that DeepZen would be a new version that is not quite yet released to the public.
- tasuki 10y agoNot 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!
- gcp 10y agoNot just plausible, it's pretty damn sure. Your certainty is based on what data?
- cybertronic 10y agohttps://www.goratings.org/ https://www.goratings.org/
- gcp 10y agoA rating based on 3 games has a pretty high (read: enormous) error margin.
- WilliamDhalgren 10y agoA rather similar hardware with a recent version of the same program plays on the KGS server. Seems to be about what would be 10d and is among the top players on the server (was 1., now seems to be 3.). You can see the graph of its playstrength; its quite impressive to see the gains Zen got from implementing ValueNet (and previously with PolicyNet) really. https://www.gokgs.com/graphPage.jsp?user=Zen19K2 https://www.gokgs.com/graphPage.jsp?user=Zen19K2 Hardware for ZenK2 was given on computer-go mailinglist: Zen19K2's information is ------------------------------------- Computer program Zen running on KURISU server provided by DWANGO. CPU: Xeon E5-2623 v3 x2 GPU: GeForce GTX TITAN X x4 http://computer-go.org/pipermail/computer-go/2016-October/009587.html http://computer-go.org/pipermail/computer-go/2016-October/00... author of Zen gave the hardware behind this match, in a post to the same mailinglist as: http://computer-go.org/pipermail/computer-go/2016-November/009612.html http://computer-go.org/pipermail/computer-go/2016-November/0... CPU: 2 x Intel Xeon E5-2699v4 (44 cores/2.2 GHz) GPU: 4 x nVidia Titan X (Pascal ) RAM: 128GB So rather similar. ~10d is way way lower than what AlphaGo had to be to play on par with one of the top players of Go. Pro ranks start around or just above the KGS 9d (which is around what a 7d rank would be for regular amateurs in most world amateur organisations). A rough approximation is that you can take 3-4 pro ranks for each dan rank/stone of handicap (well, to the extent one can even approximate strength by the rank, as a rank once won is kept for life, so as you age and young pro players pass you by, your rating is getting inflated). So apriori you could expect Zen to play around middling pro ranks, and maybe around the strength of the v13 of AlphaGo, the one that beat Fan Hui, and described in the Nature paper. And that's about what his play with Chikun indicates as well. Cho Chikun's rating is 3239, while Lee Sedol's is 3508. so around 300 elo difference, which seems a conservative estimate of the difference scale (and the results of the two matches do indicate, if weakly, that the actual difference is even larger). A 230 point gap corresponds to a 79% probability of winning. For comparison, distributed AlphaGo v13, the one that played vs Fan Hui is 250 elo points stronger than a single machine version of the same version of AlphaGo. Also if that would carry any weight with you, you could just take it on the say-so of Myungwan Kim 9p, who was doing commentary on both DeepZen and AlphaGo matches. Really the difference in strength seems rather obvious to such a strong go player, as you can see in the way he criticized some DeepZen's moves, and how impressed he was with so many of AlphaGo's.
- thomasahle 10y agoComputers 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.
- gcp 10y agowe don't really know how strong its opponent is today He's not exactly retired, you know: https://www.goratings.org/players/98.html https://www.goratings.org/players/98.html
- gcp 10y agoThe 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 10y 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.
- hkmurakami 10y agoThe key though is "for most people". We live in a post-fact society, remember? ;)
- andrewflnr 10y agoQuite. gcp is lamenting the same conflation as msbarnett.
- rwallace 10y agoI don't think there is a factual disagreement here, just a terminology issue. Mathematical jargon is not quite the same language as plain English. In the former, 'solved' by default means you have found the mathematically perfect solution. In the latter, 'solved' by default just means you have found a solution good enough for practical purposes.
- deleted 10y ago
- nv-vn 10y agoThis 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 10y agoThere's even a joke that the currency of prize money for all international StarCraft tournaments should be the Korean Won.
- convolvatron 10y agoi 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?
- conanbatt 10y agoAbout 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 10y agoCho 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.
- conanbatt 10y agoFor 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 10y agoWhat happened to chess enthusiasts when chess was "solved"?
- exolymph 10y agoThey kept playing chess with each other.
- cybertronic 10y agoand chess was not "solved"
- jm547ster 10y agoComputer aided chess became a game in and of itself
- olalonde 10y agoThe 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 https://en.wikipedia.org/wiki/Solving_chess
- andrepd 10y 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.
- espadrine 10y 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.
- brilliantcode 10y 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 10y agoCho Chikun would not stand a chance against AlphaGo. AlphaGo has shown itself to be stronger than Lee Sedol, who is much stronger than Cho.
- partycoder 10y agoIt 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.
- deegles 10y agoWhat 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-sedol-vs-alphago http://jacquesmattheij.com/another-way-of-looking-at-lee-sed...
- rcthompson 10y 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 10y agoMeasuring by power supply allows human participants if the limit is 20W.
- deegles 10y agoDeveloping 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.
- Radim 10y 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/ https://rare-technologies.com/go_games_life/
- emptybits 10y 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 10y agoI 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 10y agoThe next interesting non-Go milestone seems to be Starcraft ;)
- pelario 10y 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. * 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]) 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" :)
- bluetwo 10y agoOK, can't we just have Zen and AlphaGo play each other?
- gcp 10y agoWhat 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?
- bluetwo 10y agoErrrrrr.... It would shut up Zen go or prove them right? That seems like a reason.