Official announcement from DeepMind
https://twitter.com/demishassabis/statu ... 3282954240

Thank you!

Embedding the image from his tweet:

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

Here are some videos in which Master(P) won against Park Jeonghwan, Ke Jie, Iyama Yuta and Shin Jinseo:

Might be a good time to update the thread title to something more informative? So people don't continue to accidentally skip the thread. Maybe something like: "AlphaGo/Master(P) undefeated in 60 online blitz games vs top pros".

OP or a moderator, interested in changing the thread title?

My impression is that how AlphaGo is able to play in Japanese rule with 6.5 komi

Are they force AlphaGo? And degrading performance is acceptable since it's already way ahead of top pro.
Or they train with new data to play specifically with Japanese rule.
Or they do something, like unsupervised training, self-train from scratch, that can be flexible in rule and komi?

Regarding Game 55 - Chen Yaoye, there's a funny anecdote to it.
It was in fact Chen Yaoye and Meng Tailing playing both against Master (although the decision was always Chen's to make) during a live session at Weiqi TV.

Here's the first part where Chen Yaoye and Meng Tailing talk about their respective games and their feelings when waiting for Nie Weiping's game to finish :
https://www.youtube.com/watch?v=WUFZLxj_u70

And here's the backstage as the match was going on :
https://www.youtube.com/watch?v=BknZMjTpqmY

During the game, around minute 8, there is a funny part that translates like this after Master plays the keima in response to the pincer :

Meng Tailing : and what about the kosumi ? is it playable (shi bu shi qi = (literally) is it go) ?
Chen Yaoye : (muttering) no, that's not playable
Meng Tailing : (after Chen Yaoye protects the corner) and now, if white presses you down, do you play it ?
Chen Yaoye : i don't know

and later in the game (forgot where) :
Meng Tailing : (again) why not play this move ?
Chen Yaoye : no
Meng Tailing : what's your refutation ? how does white respond ?
Chen Yaoye : i don't know... doesn't feel like the right move...
Meng Tailing : (laughing) oh sh** !

German IT news site “Golem” links to this thread in L19 in an article about AlphaGo

http://www.golem.de/news/deep-mind-alph ... 25415.html

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

So if you linked back to it then we have a circular link. Will the internet crash?

_________________
Still officially AGA 5d but I play so irregularly these days that I am probably only 3d or 4d over the board (but hopefully still 5d in terms of knowledge, theory and the ability to contribute).

Meng Tailing reviewed his first game against Master:



Expect him to review the rest in future episodes, although he did say he performed the best in this first game and attributed part of that to not knowing it's a computer AI yet. The exchange in the upper-right was a highlight, where he believed he got much closer after falling behind in the first 50 moves. Seeing what happened at the 33:19 mark was pretty crazy too.

According to CrazyStone's analysis, this game was one of a very few "untypical" ones of the series, as it was "undecided" until late into the game (approx. move 180).

Usually, the "point of no return" is much earlier in the games (please refer to viewtopic.php?p=215477#p215477 for diagrams).

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

Please find attached CS's result of its analysis with "depth=high".

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

Which game are you referring to?

"This" game, which is the basis for the diagrams above, is game #9.

Please find attached a diagram that shows the averages of "P (B wins)" for AlphaGo as Black and as White.

EDIT: Adjusted the diagram following ez4u's suggestion below.

"1-White" is the mirrored curve for "AlphaGo as White". It is above "AlphaGo as Black", which can be interpreted that AlphaGo plays better with White than with Black (this was also said at the Lee-Sedol-time).

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

Yet another IT Web site links to this thread—this time it’s Ars Technica!

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

I think you should invert White rather than Black to make your comparison. You would then have lines for the probability AG wins as Black versus the probability AG wins as White.

In any case, we see that the winning probability is higher playing White from the very beginning. The immediate interpretation would seem to be that komi is too large and gives White an advantage. This is consistent with many players' opinion based on human vs human play. The lines seem very consistent over the first 150+ plays. If AG were better playing White, the gap between the lines should diverge move by move. This does not seem to happen to any significant extent.

How should we interpret the changes that occur around move 165? Is it AG changing its evaluation process as the winning probability grows too large? Is it desperation appearing in the opponents' play?

Anyway, very interesting information. Well done!

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

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.

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

...

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

For those who are interested in, please find attached the exports of CS's analysis of the five games mentioned above.

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)

The only thing we can conclude is that Crazy Stone prefers white. We know nothing about AlphaGo's perceived winning probability.

Yes, indeed, it seems that there is an offset. I did not consider this.

_________________
The really most difficult Go problem ever: https://igohatsuyoron120.de/index.htm
Igo Hatsuyōron #120 (really solved by KataGo)