Life In 19x19

Uberdude wrote:Whilst I don't think Master is AlphaGo*, that last game against Ke Jie with the white top left invading group eventually turning the tables and counter-attacking black did rather remind me of AlphaGo vs Lee Sedol game 3.

* because of the weirder opening style, though quite possibly since LSD AlphaGo has gotten stronger and free-er so changed style, maybe newly trained from self-play rather than humans.

Chatting in KGS and someone said that Tencent guy said that 'this dog is not that dog' (dog = alphago's nickname?)
Anyway, I rule out Chinese and Japanese AI since it's highly unlikely that they'll use South Korean flag.
If it's South Korean AI, then it's definitely Dolbaram
If it's not, my second choice is Deepmind's AI (AlphaGo or new AI), followed by CrazyStone.

A source from West China City Daily claimed that Master is indeed a newer version of AlphaGo, currently in beta: http://sports.sina.com.cn/go/2017-01-04 ... 4149.shtml (in Chinese). Still very unofficial though. It does make sense in that AlphaGo was given a honorary 9P by Korean Baduk Association after the Lee Sedol matches.

Playing Nie Weiping now, with 3 60-seconds byoyomi to compensate for his age (so I guess there are definitely some negotiation behind the scenes).

Someone assume that Fan Hui is behind cause some of his pro friends get to play more than others.

Winrate of bots
Master 98% Tygem and Fox (due to tie game because of Chen Yaoye's disconnection)
DeepZenGo 90% Tygem
Xing Tian 84% Fox
Jueyi 70% Fox

All bots are against 9d in even match.

Master thanked Nie Weiping in traditional Chinese ("謝謝聶老師") in the end, so still think it's Aja Huang behind given he's from Taiwan.

Of course not only neural network. The main power is still Monte Carlo tree search. Neural network just decreases the number of candidate moves. So compute power plays decisive role.

Marcel Grünauer wrote:From what I've read, Master seems to play its moves within ten seconds. Does that mean that it mostly relies on the neural network(s) and that relatively few computation is involved?

The other possibility is that its machine is superfast that it calculate enough search to be that strong within 10 seconds. For example, even the single machine version of AlphaGo is 48CPU and 8GPU, about 8 times of the consumer grade hardware.

Marcel Grünauer wrote:If so, Master could have a similar strength on consumer grade hardware, given a bit more thinking time.

From AlphaGo nature paper, in 2 seconds per move game, the gap between distributed version (1202CPU/176GPU) and single machine version is about one stone, the gap between single machine version and 48CPU/1GPU version is more than three stones. Consumer grade hardware have to wait for a bit, it'll definitely happen but not that soon.

As of now, we've solid evidence since 2010 that human use extra time much more efficiency than bot. So while it's true that giving bot more time will increase its strength, human time must stay the same, and (s)he must not think while in bot's turn to play.

In chess, the span between "first win of a bot against top human" and "last win of a human against top bot" was about 10 years. In Go it is less then 1 year, which I find extremely impressive (we still have to wait for longer time controls. Humans might simply more time to cope with the creative moves of the Master But I doubt whether it makes a difference, the comments being so far along the lines of "Master too strong" and not "I should have won this".)

Is there a good reason, why the time span between first win against top human and no loss against human is so short, or was chess a special case?

bayu wrote:In chess, the span between "first win of a bot against top human" and "last win of a human against top bot" was about 10 years. In Go it is less then 1 year, which I find extremely impressive (we still have to wait for longer time controls. Humans might simply more time to cope with the creative moves of the Master But I doubt whether it makes a difference, the comments being so far along the lines of "Master too strong" and not "I should have won this".)

Is there a good reason, why the time span between first win against top human and no loss against human is so short, or was chess a special case?

Neither chess nor go are special case. A game that has luck in its element, like Scrabble, card game variation (poker?) bot can still lose to top human if it's really, really unlucky.
And English draughts/American checkers went from "first win of a bot against top human" to "solved" in 12 years, that escalated very quickly!!!

Back on topic, do you know which rule Master is playing, Japanese or Chinese?

Master (aka confirmed as AlphaGo) is playing with Gu Li right now, and it look like to be the last game for a while so no one will get prize even if Gu Li win.

Is this DeepMind's new year present? Well, my favorite part of this whole event is when Ke Jie said that 'how can a gap be this large?!!!'

I don't like this hysterical behavior of Ke Jie. Let's keep some respect to pro players and be more skeptical to bot's strength. Ok it is too strong to beat it in fast game but let pro play a long one.

Marcel Grünauer wrote:From what I've read, Master seems to play its moves within ten seconds. Does that mean that it mostly relies on the neural network(s) and that relatively few computation is involved?

If so, Master could have a similar strength on consumer grade hardware, given a bit more thinking time.

Or maybe they use 1000 GPUs. Plus 100 CPUs.

Back in the AlphaGo Nature paper they described their system and it used 2 neural nets + MCTS. There were a bunch of graphs of it's strength with all combinations of those 3 systems. All 3 together was much stronger than other combinations. And those measurements were made giving the computer 5 seconds per move.

Posted in the other thread, but will do so here as well. Ke Jie's latest tweet: http://weibo.com/u/2865101843?is_all=1

"Thanks for the shock and awe the latest version of AlphaGo provided us. As someone who knew its identity from the beginning, I'd have loved to see us human winning a blitz game. If not for being in the hospital, I'd have tried one last response that I prepared for a week...slight regret, and hope Gu Li can play to the best of our human ability in this final game of AlphaGo beta test."

Oh, and my other favorite part is when western and eastern social network working together to find who master is. Well planned, deepmind.

I saw the twitter confirmation. Was there any comment on whether they've changed the training methods, given that the new style is so unique?