Page 2 of 2
Re: More AlphaGo games to be played in 2017
Posted: Mon Nov 07, 2016 11:07 am
by Bonobo
To expand on Uberdude’s comment, Aja Huang’s comment on
his — edited — post on FB:
Aja Huang wrote:一位團隊的同事說"significantly stronger"不足以表達AlphaGo進步幅度之大，我已根據他的建議改成"massively stronger"。
Automatic translation on FB makes this of it:
A team of colleagues say " significantly stronger " is not sufficient to express go progress magnitude, I have been under his proposal to replace " massively stronger".
So… something like …
geologist’s hammer vs.
steam hammer?
Re: More AlphaGo games to be played in 2017
Posted: Mon Nov 07, 2016 11:22 am
by DrStraw
Uberdude wrote:Edit: and Aja, on the advice of his colleagues, just edited his post from "significantly stronger" to "massively stronger".
I have to wonder in what basic this claim is made. There isn't anyone massively stronger to compare it to. Did it play and earlier version of itself to come to this conclusion?
Posted: Mon Nov 07, 2016 11:27 am
by EdLee
There isn't anyone massively stronger to compare it to. Did it play and earlier version of itself to come to this conclusion?
This part seems relatively trivial: handicap stones. Either against previous versions, or against human pro(s).
Re:
Posted: Mon Nov 07, 2016 11:29 am
by DrStraw
EdLee wrote:There isn't anyone massively stronger to compare it to. Did it play and earlier version of itself to come to this conclusion?
This part seems relatively trivial: handicap stones. Either against previous versions, or against human pro(s).
I was actually assuming that had it played again another pro we would have heard about it, but perhaps that is not so.
Re: More AlphaGo games to be played in 2017
Posted: Mon Nov 07, 2016 11:32 am
by uPWarrior
(Good) Scientists tend to avoid the word "significantly" in publications as it might be read as "statistical significant", which requires significantly more work to show (pun intended).
But I think Aja's comment means something else, as it was not changed as if fixing a typo.
Posted: Mon Nov 07, 2016 11:33 am
by EdLee
I was actually assuming that had it played again another pro we would have heard about it, but perhaps that is not so.
Remember that they had worked with a pro for a few months before the public announcement last year.
They could do the same thing: either with the same pro(s), or with different pro(s), and new experiments (handicap stones).
Because of non-disclosure agreements, the public would not have heard anything about it until their official announcements.
Re: More AlphaGo games to be played in 2017
Posted: Mon Nov 07, 2016 12:04 pm
by Gotraskhalana
I would certainly be interested to see a game with a lot of handicap stones (even between alpha go and a previous version of itself) because it is an interesting question how many stones a "perfect player" could give the current top players. Four stones already seems a lot.
Re: More AlphaGo games to be played in 2017
Posted: Mon Nov 07, 2016 1:20 pm
by yoyoma
So it's time to debate the definitions of "significant" and "massive"? Fun!
I guess they are basing this on games vs older versions of itself.
Significant: 80%-90% winrate against previous version
Massive: 95+% winrate against previous version
Source: me
Poll for science!
What is the minimum winrate for "Massive Improvement"?
http://www.strawpoll.me/11597646
Re:
Posted: Mon Nov 07, 2016 5:51 pm
by pookpooi
EdLee wrote:I was actually assuming that had it played again another pro we would have heard about it, but perhaps that is not so.
Remember that they had worked with a pro for a few months before the public announcement last year.
Also Fan Hui is in DeepMind team too.
And Aja Huang is also 7d KGS, if AlphaGo can give him six stones and win then it's really a massive improvement.
Re: More AlphaGo games to be played in 2017
Posted: Mon Nov 07, 2016 6:13 pm
by John Fairbairn
Yasunaga Hajime was a 4d pro. At one stage Go Seigen beat him down to 4 stones (in quickplay). There are other similar examples.
Re: More AlphaGo games to be played in 2017
Posted: Tue Nov 08, 2016 12:48 am
by pookpooi
Aja Huang is changed into just 'stronger'.
The details that are coming soon are not soon enough for me lol.
Posted: Tue Nov 08, 2016 1:01 am
by EdLee
So it's time to debate the definitions of "significant" and "massive"?
My guess is they didn't want 'significant' to be confused with 'statistically significant'.
Suppose the new version is 5% better.
People may ask: well, is the 5% just noise, or is it statiscally significant ?
Reply: it's not just noise; the 5% is statiscally significant.
They want to differentiate between:
- a "small, but nonetheless statiscally significant" improvement ;
- a huge improvement ( in everyday language ) ;
They want "massive" to convey the latter sense.
Just my guess, though.
Re: More AlphaGo games to be played in 2017
Posted: Tue Nov 08, 2016 1:21 am
by Cassandra
pookpooi wrote:Aja Huang is changed into just 'stronger'.
Too many rumors on the net ?
Now, his statement contains almost no information.
According to Demis Hassabis, they have "been
hard at work improving AG" (emphasis mine), so it should be self-evident that AlphaGo made some progress.
Re: More AlphaGo games to be played in 2017
Posted: Tue Nov 08, 2016 4:27 am
by pookpooi
Cassandra wrote:pookpooi wrote:Aja Huang is changed into just 'stronger'.
Too many rumors on the net ?
Now, his statement contains almost no information.
According to Demis Hassabis, they have "been
hard at work improving AG" (emphasis mine), so it should be self-evident that AlphaGo made some progress.
Also AlphaGo team is bigger even after Lee Sedol match
https://www.reddit.com/r/baduk/comments ... _deepmind/
Re: More AlphaGo games to be played in 2017
Posted: Tue Dec 13, 2016 5:23 pm
by Bonobo
http://www.chinadaily.com.cn/bizchina/t ... 659466.htm (2016-12-13)
Google bosses meet top Go players in Beijing
Google's top executives have met with China's best Go players, including the current world champion.
Google CEO Sundar Pichai, and co-founder Sergey Brin, who is also president of Alphabet, met some of the board game's best players on Tuesday at the headquarters of the Chinese Go Association.
The weibo account of Google posted several pictures of the meeting that included world champion Nie Weiping and Yu Bin, head of the Chinese Go team.
The visit puts Google's computer program AlphaGo back in the spotlight, and follows an announcement last month from Demis Hassabis, the father of the computer program AlphaGo, that AlphaGo will resume playing early next year.
Hassabis didn't reveal who and where AlphaGo will play.
[..]