New AlphaGo prototype has 50-0 record against top Go players

[Laskos]

In "Nature News",
http://www.nature.com/news/google-revea ... rs-1.21253

I would be curious about hardware used.

[CheckersGuy]

My guess would be that they used smiliar hardware which was used in the match against Lee SeDol. I think a crushing defeat is awaiting the top professionals that want to play Alpha-Go this year.

https://en.wikipedia.org/wiki/AlphaGo#Hardware

[Laskos]

My guess would be that they used smiliar hardware which was used in the match against Lee SeDol. I think a crushing defeat is awaiting the top professionals that want to play Alpha-Go this year.

https://en.wikipedia.org/wiki/AlphaGo#Hardware

If the machine is similar to Lee Sedol match, then it's pretty hard to guess its stregnth for regular home use. MCTS Go scales very well to many cores, say a 24 core, 2 GPU system might be 400-500 ELO points weaker than this distributed machine. And it's again hard to transform ELO points to dan levels, it is probably worth 2-3 KGS amateur dans (stones), but Pro dans are different. 50-0 against top pro Go players might indicate though that AlphaGo reached top Pro level for home use. It seem it is stronger than Crazy Stone and Zen for this purpose, which are not yet level with top Pro level on home machine.

[Isaac]

If the machine is similar to Lee Sedol match, then it's pretty hard to guess its stregnth for regular home use. MCTS Go scales very well to many cores, say a 24 core, 2 GPU system might be 400-500 ELO points weaker than this distributed machine. And it's again hard to transform ELO points to dan levels, it is probably worth 2-3 KGS amateur dans (stones), but Pro dans are different. 50-0 against top pro Go players might indicate though that AlphaGo reached top Pro level for home use. It seem it is stronger than Crazy Stone and Zen for this purpose, which are not yet level with top Pro level on home machine.

I think zen reached pro level on home machine (top or close to top KGS player on a small server : https://www.gokgs.com/top100.jsp and here is a ranking list on home machine : http://www.yss-aya.com/cgos/19x19/bayes.html). Crazystone on the other hand lags a bit behind zen unless Rémi Coulom has a much improved new version.

[Laskos]

If the machine is similar to Lee Sedol match, then it's pretty hard to guess its stregnth for regular home use. MCTS Go scales very well to many cores, say a 24 core, 2 GPU system might be 400-500 ELO points weaker than this distributed machine. And it's again hard to transform ELO points to dan levels, it is probably worth 2-3 KGS amateur dans (stones), but Pro dans are different. 50-0 against top pro Go players might indicate though that AlphaGo reached top Pro level for home use. It seem it is stronger than Crazy Stone and Zen for this purpose, which are not yet level with top Pro level on home machine.

I think zen reached pro level on home machine (top or close to top KGS player on a small server : https://www.gokgs.com/top100.jsp and here is a ranking list on home machine : http://www.yss-aya.com/cgos/19x19/bayes.html). Crazystone on the other hand lags a bit behind zen unless Rémi Coulom has a much improved new version.

Thanks for the second list, I was unaware of it. I was saying that Crazy Stone and Zen are not _top_ Pro level on home machine, but AlphaGo probably is. The Bayeselo ratings of Pro Go players vary from 2700 to 3600 in "Go ratings", and from top 100 KGS my guess was that Zen and Crazy Stone are probably low to average Pro level on home machine, maybe 3000. On the other hand, I didn't know that Zen distanced itself from Crazy Stone, and might be at stronger Pro level on home machine.

[matthewlai]

My guess would be that they used smiliar hardware which was used in the match against Lee SeDol. I think a crushing defeat is awaiting the top professionals that want to play Alpha-Go this year.

https://en.wikipedia.org/wiki/AlphaGo#Hardware

If the machine is similar to Lee Sedol match, then it's pretty hard to guess its stregnth for regular home use. MCTS Go scales very well to many cores, say a 24 core, 2 GPU system might be 400-500 ELO points weaker than this distributed machine. And it's again hard to transform ELO points to dan levels, it is probably worth 2-3 KGS amateur dans (stones), but Pro dans are different. 50-0 against top pro Go players might indicate though that AlphaGo reached top Pro level for home use. It seem it is stronger than Crazy Stone and Zen for this purpose, which are not yet level with top Pro level on home machine.

I can't say much, but from the Nature paper, the distributed version only wins 77% against single machine version . That's ~200 Elo.

MCTS does scale very well, but scaling across machines is much more difficult.

[Laskos]

My guess would be that they used smiliar hardware which was used in the match against Lee SeDol. I think a crushing defeat is awaiting the top professionals that want to play Alpha-Go this year.

https://en.wikipedia.org/wiki/AlphaGo#Hardware

If the machine is similar to Lee Sedol match, then it's pretty hard to guess its stregnth for regular home use. MCTS Go scales very well to many cores, say a 24 core, 2 GPU system might be 400-500 ELO points weaker than this distributed machine. And it's again hard to transform ELO points to dan levels, it is probably worth 2-3 KGS amateur dans (stones), but Pro dans are different. 50-0 against top pro Go players might indicate though that AlphaGo reached top Pro level for home use. It seem it is stronger than Crazy Stone and Zen for this purpose, which are not yet level with top Pro level on home machine.

I can't say much, but from the Nature paper, the distributed version only wins 77% against single machine version . That's ~200 Elo.

MCTS does scale very well, but scaling across machines is much more difficult.

Well, that single machine is pretty much a monster, 48 core 8 GPU. I was saying that a reasonable 24 core 2 GPU system might be 400-500 ELO points weaker than the full distributed 2000 core machine used against Lee Sedol. Still, 50-0 record indicates that even a home machine with AlphaGo might be a _top_ Pro level (say 3500 in BayesElo "Go Ratings", where Ke Jie is at 3600). That is way above Zen and Crazy Stone on home machine (although the post from Isaak indicates that Zen progressed a lot).

[jshriver]

I've been following this on the mailing list. Go is really hitting Deep Blue stages now and it's been exciting to read.

Google created some beautiful Neural-Network based ASIC for their own use, and also used it for their own Go engine which played about a year ago. Was amazing. Since then NN has been insanely a buzz on the mailing list.

If anyone is interested in this I really recommend joining the mailing list. Great people like this community.

I just wish Arimaa had such a following lol

-Josh

[duncan]

I can't say much, but from the Nature paper, the distributed version only wins 77% against single machine version . That's ~200 Elo.

an insider quoting a paper positively is saying much.

[duncan]

I would like to see the results of fourth game.

https://qz.com/878503/ke-jie-the-worlds ... lphago-ai/

But at least one human hasn’t given up hope. Ke Jie, the reigning top-ranked Go player, claimed that he still has “one last move” to defeat AlphaGo, after he lost three games to the AI during the test.

“If I were not in hospital, I would have used the one last move I prepared for a week… It’s a little shame,” Ke wrote Jan. 4 on Weibo, China’s version of Twitter, shortly after AlphaGo revealed its identity. The 19-year-old said he had already known that “Master,” the user name for the AI on the gaming platforms, was AlphaGo in disguise even before its identity was confirmed by Google—and that’s why he “hoped so much” that humans could win at least one game.

Ke revealed that he was taken to hospital, for the first time in his life, after his three losses to AlphaGo, but didn’t provide specific reasons. His latest Weibo message suggested that he would like to take on AlphaGo for a fourth time.



https://www.goratings.org/
1 Ke Jie ♂ cn 3628
2 AlphaGo gb 3599