Who's playing in Tygem? -- It was AlphaGo!!

[splee99]

...

I watched these two games where both human players tried new opening moves to test alphago. It appears that these 'new' moves only led to faster lose.

[negapesuo]

Please find below the results of CS's analysis (depth="high") for some of the "longer" games.
Scaling for both axis is identical in all the diagrams.

I am not strong enough to correlate this "bot view" with the judgement of a strong human player.

However, one idea might be that the diffences in the graphs are affected by the number of territories during the games ?

Please note that the "score" in game #9 did never exceed 10 points, which seems to be quite unique here (#36 was in this range for most of the game, especially in the phase before the "point of no return").
On the other hand, Gu Li's large moyo strategy in game #60 was not "praised" by CS.

Interesting how Ke Jie's winning probability was higher than other humans. I guess, he really is in a class of his own.

[xiayun]

More review of the games from Weiqitv, this time with Mi Yuting on his game against Master (featuring a brand new variation of large avalanche joseki @11:30):

[Bonobo]

Hajin Lee reviews a game between AlphaGo and Ke Jie

[xiayun]

More from Weiqi TV:

Park Junghwan vs. Master (Game 24)

Master vs. Tang Weixing (Game 30)

[Bonobo]

Anybody analyzing the Master games might be interested in this Facebook post by Aja Huang:

https://www.facebook.com/aja.huang/post ... 5658588001

(emphasis added by me)

感謝大家對AlphaGo最近的測試對局所展現出的興趣，我們團隊感到很高興。由於這60局棋被廣泛且深入的研究，我們想要澄清兩處人為的操作失誤，幫助大家更好的分析這些對局。這些是由於快棋而導致的人為操作失誤，因此沒有下出AlphaGo原本的下法。
The AlphaGo team has been delighted by the Go community's interest in AlphaGo's recent test games played by Master(P) and Magister(P). As these 60 games have become the object of deep study and investigation, we would like to clarify two situations in which, due to the short time controls, the operator accidentally entered a move other than AlphaGo's recommendation. We hope these notes will help all the Go players analyzing and learning from these games.

在第9局與HopeIdo(P)的對局中，AlphaGo原本是想下在圖中的黑1(F13)提子，實戰的K6是操作者的失誤。
In game 9 against HopeIdo(P), AlphaGo recommended capturing at F13, and K6 was an operator error.

在與jpgo01的對局中，AlphaGo原本是想下在圖中的黑1(O12)，實戰的N6是人為失誤。
In the game against jpgo01, AlphaGo suggested O12 instead of N6, which was also due to human error.

Please note that the coordinates are different from what we use here:
• A–S from left to right (INCLUDING J)
• 1–19 from top to bottom
________________

Sadly it’s not possible to include screenshots from there, but if you say it’s necessary I’d download them from there and upload here, assuming that Aja Huang wouldn’t object.

<edit>
I just asked Aja Huang there whether it’s okay to suck the screenshots, he doesn't object
</edit>

HopeIdo

hopeido.jpg (77.81 KiB) Viewed 9938 times

jpgo01

jpgo01.jpg (75.1 KiB) Viewed 9938 times

[xiayun]

Yeah, the specific move in Game 9 was especially interesting to me since in Meng Tailing was wondering why AlphaGo didn't take the stones and showed a disconnect with its previous move in his self-review on WeiqiTV. Good to know it's actually a human error. So I guess in those two cases Dr. Huang entered one of AG's backup recommendations and missed its optimal suggestion?

[Uberdude]

So I guess in those two cases Dr. Huang entered one of AG's backup recommendations and missed its optimal suggestion?

Maybe so in the Meng game (or he played AlphaGo's expectation of white's move after AlphaGo's capture move?), but for the Iyama game on facebook Aja said it was a simple misclick, but a pretty good one!

大橋拓文[Ohashi Hirofumi 6p]: I thought K6 is strange in game 9. On the other hand, N6 was good move, but it looks better O12 now with jpgo01. Was N6 a second or third suggestion by AlphaGo??
Aja Huang: No, N6 was simply a mis-click.

[Uberdude]

That would mean that AlphaGo had to make the next move even faster, doesn't it?

Not necessarily, if the average Master move time was 10 seconds on a 20 or 30 second byo-yomi then maybe after the misclicks it took 18 or whatever but no one noticed and we just talk about the average 10. Do Tygem games store move times like KGS?

[emerus]

That would mean that AlphaGo had to make the next move even faster, doesn't it?

Not necessarily, if the average Master move time was 10 seconds on a 20 or 30 second byo-yomi then maybe after the misclicks it took 18 or whatever but no one noticed and we just talk about the average 10. Do Tygem games store move times like KGS?

Master consistently made its move in 7 or 8 seconds on Tygem. Its surprising that Aja Huang was inputting moves and never using full byo-yomi. Perhaps shows just how confident the Deepmind team is.