The 1st World AI Go Open 2017

kris · #1

It will start tomorrow in China.
12 Go engines is participating:

Yi-Tianrang CHINA
AQ JAPAN
DolBaram KOREA
OracleWQ CHINA
CGI TAIWAN
MuGo USA
DeepZenGo JAPAN
Leela BELGIUM
FineArt CHINA
Rayn JAPAN
Golois FRANCE

Which one is your favourite?

jeromie · #2

From what I've seen / read, FineArt is probably the expected winner in this competition. I'd love to see Leela do well, though, since I can actually play with it. :-)

Vargo · #3

Very interesting event, I'm really looking forward to following it !

The URL seems to be :
http://www.intergofed.org/igf-news-feed/2017-citic-securities-cup-the-1st-world-ai-go-open.html

and what do you think of this rule ?

1...
2...
3...
4. Human vs machine match
The champion of the 1st World AI Go Open will play
against World Champion KongJie 9p. The winner gets
70000 RMB while the loser receives 30000 RMB. For this
match, KongJie will select from one of the participating AI
programs as his “assistant”. A stipend of 10000 RMB
（before tax）is awarded to the AI team.

tchan001 · #4

Vargo wrote:

4. Human vs machine match
The champion of the 1st World AI Go Open will play
against World Champion KongJie 9p. The winner gets
70000 RMB while the loser receives 30000 RMB. For this
match, KongJie will select from one of the participating AI
programs as his “assistant”. A stipend of 10000 RMB
（before tax）is awarded to the AI team.

It looks very much like the application of what Ke Jie once wrote before he played a game against Master on the FoxGo server.

Check out this quote from the book "Invisible - The Games of AlphaGo" page 42:

Quote:

The last opponent of the day, Ke Jie, wrote a note on Weibo before his game that blended resignation with hope: ‘Humans have spent thousands of years playing go, practising and evolving, but computers now tell us everything we know is wrong… But I believe that going forward, the union of human and computer players will usher in a new era. A storm is coming, and I will apply all of my knowledge to it… Together, man and AI can find the truth of go. Only AI can lead the way in this endeavour. Throughout this process, the model should be one of mutual enlightenment and encouragement.’

RobertJasiek · #5

"Only AI can lead the way in this endeavour."

This overlooks mathematical theorems of go theory, which currently are only created by humans. Theorems have a potential much greater than AI and human play combined but require very much mathematical research. In the (far?) future, GAI or AI specifically designed to successfully do mathematical research might also create theorems. Currently, all the research in mathematical theorems of go theory is human. Any computer contribution to that so far is just occasional and assistive but not creatively researching.

Neural net / Monte Carlo AI and human play are subject to making mistakes. Theorems are without mistakes.

Uberdude · #6

Results of preliminary rounds (5 rounds to narrow 12 entrants down to 8 for knockout finals);

Round 1. FineArt probably the favourite, DeepZen next? CGI apparently improved a lot. DolBaram from Korea was on of the top bots pre AlphaGo a bit behind Zen/CrazyStone, came 2nd in a UEC cup, 3-4 handis from pro back then. AQ and RayN are both fairly new Japanese deep learning bots. I don't know much about Abacus or Tianrang or Oracle from China, are any of these the other strong Chinese bots that were playing on Fox? (LiLong was an alias of FineArt, dunno about others such as Xing Tian from forum/viewtopic.php?f=10&t=13921).
FineArt beat Leela
DeepZenGo beat Abacus
Rayn beat Golois
Tianrang beat AQ
Dolbaram beat Oracle
CGI beat Mugo

Round 2.
CGI beat DeepZenGo! Notable result, by resign or half point on board if it's Japanese 6.5 komi
AQ beat Oracle
Abacus beat MuGo
FineArt beat Rayn
Tianrang beat Dolbaram
Leela beat Golois

Round 3. FineArt pushed through the table shape in magic sword joseki, like AlphaGo does. It's also played quite a few 3-3 points so maybe shares some value of that with AlphaGo, but doesn't do the early 3-3 invasions under 4-4s. AQ put up a decent fight. Oracle is comically weak, >20 kyu, keeps dying on 2nd line. Why is it even competing, maybe expenses are paid, guaranteed prizemoney or they just wanted to meet other go bot programmers?
FineArt beat AQ.
Rayn beat Oracle
Dolbaram beat Abacus
CGI beat Tianrang
DeepZenGo beat Leela
Golois beat MuGo

Round 4.
CGI beat FineArt!!!!
Rayn beat Dolbaram
DeepZenGo beat Tianrang
Golois beat Abacus
AQ beat Leela
MuGo beat Oracle

Round 5.
CGI vs Rayn
FineArt vs DeepZenGo
Tianrang vs Golois
AQ vs Dolbaram
Abacus vs Oracle
Leela vs MuGo

Standings after round 5 (SOS in brackets):
5 wins: CGI (14)
4 wins: FineArt (15)
3 wins: DeepZenGo (16), Tianrang (15), Rayn (14), Dolbaram (13)
2 wins: Leela (12), AQ (12), Golois (11), Abacus (9)
1 win: MuGo (11)
0 wins: Oracle (11)
So MuGo and Oracle out. Golois beat Abacus and MuGo so SOS 3, AQ beat Leela and Oracle so SOS 2, Abacus beat Oracle and MuGo so SOS 1, Leela beat Golois and MuGo so SOS 3. So probably AQ and Abacus out too. I didn't look much at Golois's games, but I feel a bit sorry for AQ, it seemed decent and the last game with DolBaram was spectacular with loads of dead groups. Quite a lot came down to the luck of who got drawn with the idiotic Oracle, GnuGo would have been a tougher competitor!

Some game records, more at http://weiqi.qq.com/qipu.html (Chinese). Search for 绝艺 to find FineArt in titles.

Cyan · #7

Uberdude wrote:

I don't know much about Abacus or Tianrang or Oracle from China, are any of these the other strong Chinese bots that were playing on Fox? (LiLong was an alias of FineArt, dunno about others such as Xing Tian from forum/viewtopic.php?f=10&t=13921).

Xing Tian is also FineArt.

Cyan · #8

Uberdude wrote:

So MuGo and Oracle out. Golois beat Abacus and MuGo so SOS 3, AQ beat Leela and Oracle so SOS 2, Abacus beat Oracle and MuGo so SOS 1, Leela beat Golois and MuGo so SOS 3. So probably AQ and Abacus out too. I didn't look much at Golois's games, but I feel a bit sorry for AQ, it seemed decent and the last game with DolBaram was spectacular with loads of dead groups. Quite a lot came down to the luck of who got drawn with the idiotic Oracle, GnuGo would have been a tougher competitor!

That's SODOS, not SOS. SOS is the sum of the scores of all opponents that a player played against in a tournament. In general SOS is preferable to SODOS. There are several variations of SOS:

Quote:

SOS-1 has two distinct variations
drop the numerically smallest opponent's score from the sum. (Known to Chess as Buchholz Cut 1)
drop the first round opponent's score from the sum. The first round differs from all following rounds in that first round is randomly paired. This method was used in the 1st World Mind Sports Games. [ext] source
SOS-2 has two distinct variations which proceed directly from the definitions of SOS-1
drop the numerically smallest and the next smallest opponents' scores from the sum. (Known to Chess as Buchholz Cut 2)
drop the both the first and second round's opponents score from the sum.
SOS-n has two distinct variations which proceed directly from the definitions of SOS-1
drop the numerically smallest n scores from the sum.
drop the both the first n round's opponents score from the sum.

So probably Golois and Abacus are out.

Uberdude · #9

Cyan wrote:

That's SODOS, not SOS. SOS is the sum of the scores of all opponents that a player played against in a tournament. In general SOS is preferable to SODOS. There are several variations of SOS:

Good point! So my feeling of AQ being hard-done-by is perhaps due to bad tiebreaker. So basic SOS was
Leela: FineArt 4 + Golois 2 + Zen 3 + AQ 2 + MuGo 1 = 12
Golois: Rayn 3 + Leela 2 + MuGo 1 + Abacus 2 + Tianrang 3 = 11
AQ: Tianrang 3 + Oracle 0 + FineArt 4 + Leela 2 + Dolbaram 3 = 12
Abacus: Zen 3 + Mugo 1 + Dolbaram 3 + Golois 2 + Oracle 0 = 9

Cyan · **#10**

Uberdude wrote:

Cyan wrote:

That's SODOS, not SOS. SOS is the sum of the scores of all opponents that a player played against in a tournament. In general SOS is preferable to SODOS. There are several variations of SOS:

Good point! So my feeling of AQ being hard-done-by is perhaps due to bad tiebreaker. So basic SOS was
Leela: FineArt 4 + Golois 2 + Zen 3 + AQ 2 + MuGo 1 = 12
Golois: Rayn 3 + Leela 2 + MuGo 1 + Abacus 2 + Tianrang 3 = 11
AQ: Tianrang 3 + Oracle 0 + FineArt 4 + Leela 2 + Dolbaram 3 = 12
Abacus: Zen 3 + Mugo 1 + Dolbaram 3 + Golois 2 + Oracle 0 = 9

Our predictions are right, Leela and AQ entered the final 8. Here are the pairings for tomorrow:
http://i.imgur.com/YttfbCg.jpg

Uberdude · **#11**

Knockout quarterfinal results:

Tianrang beat Rayn
CGI beat Leela
DeepZen beat Dolbaram
FineArt beat AQ

Dolbaram played AlphaGo-esque early 3-3 invasion, crawled an extra time with no hane connect and then peeped the wall (but does AG peep before crawl?). Has it independently discovered this strategy too? Nice play from both, Dolbaram took all 4 corners and Zen used aji nicely to attack reduction group but then Dol inexplicably played an atari tenuki which let Zen's dead group live and then kill more stuff and it all went downhill.

CGI and Leela went their own way in the opening, cgi made a boxy territory which Leela invaded too late and died without compensation, probably over then.

FineArt killed AQ in the first joseki fight in exchange for losing 2 stones in a ladder. Fairly smooth win from there and even pulled out the stones in exchange for a group dying which was a nice kill by AQ but it later messed up allowing it to become seki.

Tianrang played a cute tesuji on the 2-1 point which meant a group that looked safe could later end up as a dead big eye. Seems like that confused rayn which didn't play well and tianrang even let it live but got enough elsewhere to win by resign before endgame.

Semifinals:

CGI beat Tianrang
DeepZen beat FineArt !!!

Final:

DeepZen beat CGI !!

Quite a reversal from the prelims where CGI beat Zen and FineArt, and Zen also lost to FineArt, now beats both when it counts. Congrats team DeepZen!

pookpooi · **#12**

DeepZen win against FineArt! Sweet revenge.
Now waiting for final match.

gamesorry · **#13**

CGI has also developed world-champion-level Connect6 programs (https://www.littlegolem.net/jsp/info/pl ... plid=20818)

pookpooi · **#14**

Result
1. DeepZenGo
Congrats! Remind me of this year computer olympiad when Zen won gold medal while CGI had silver in 19x19. Now the commercial version can have the tagline 'World AI Go Champion'
2. CGI
very surprised, and if more games are played instead of round-robin it might become a winner. I hope CGI play in go server for a while.
3. FineArt
again, it's very close between these three programs, more games are needed to determine more accurate ranking
4. Tianrang

Uberdude · **#15**

Record for Zen beating FineArt in the semi final below with Chinese comments from Fox; the key moment seems to be when FineArt playing black pulled out the 2 stones. There is a lot of aji and the marked white stones are rather weak so my feeling at the time was a kill would be hard, but Zen went for it. The rest of the game was the life or death fight over these stones, and Zen succeeded in killing them in classic Zen centre-kill style. Perhaps given white's thickness at the right a little more preparation was prudent (e.g. something like lean at a (black did actually get this exchange for block on 3rd line later) and/or attach at b and if white answers then maybe black can pull out the 2 stones, so if white plays honte to (almost-)complete their capture then black gets some follow-up profit. This actually reminds me a bit of the famous Karigane-Shusai killing game where one of the players should have made some outside exchange before starting the sequence.

Click Here To Show Diagram Code: [go]$$B FineArt (black) vs DeepZenGo $$ --------------------------------------- $$ | . . . X . O . . . . . . . . . . . . . | $$ | . . O X O . O . . . . . . . O X . . . | $$ | . O X X X O . X X . . . . . . . . . . | $$ | . . O X X . . . . , . . . . O X X . . | $$ | . . . X . . X . . . . . . . O O X . . | $$ | . . . . . . . . . . . . . . . X O X . | $$ | . X X . . . . . . . . . . . . O O X . | $$ | . O O O . . . . . . . . . . . O X O . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . X , O . . . . , . . . . . , O . . | $$ | . . . . . . . . . . . . . . . a . . . | $$ | . . . . . . . . . . O . . . . . . . . | $$ | . O . . . . . . . . . . . . b O . . . | $$ | . . O O O . . . . 1 X . . . . . . . . | $$ | . O X X O B . . . O X W W . . . . . . | $$ | O X X X X . O . O , O X . X X , X . . | $$ | . O O X . X . . . . O X . . . . . . . | $$ | . . . X . . . X . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Here's black's dead centre 100 moves later, despite capturing 2 stones at crossed points. White's outside groups had some problems but all managed to survive (lower left group reduced to only 1 eye but wins semeai with black group inside).

Click Here To Show Diagram Code: [go]$$B Move 179-180 $$ --------------------------------------- $$ | . . . X . O . . . . . . . . . . . . . | $$ | . . O X O . O . . . . . . . O X . . . | $$ | . O X X X O . X X . . . . . . . . . . | $$ | . . O X X . . . . , . . . . O X X . . | $$ | . . . X . . X . . . . . . . O O X . . | $$ | . . . . . . . . . . . . . . X X O X . | $$ | . X X . . . . . . . . X O O O O O X . | $$ | . O O O . O O O O O . O X X . O X O . | $$ | . . . . X O . O X X . O O X O . . . . | $$ | . X X X O O O X X X . O X X X O O . . | $$ | . X O O X X X M M X . O X . O X O . . | $$ | X X O X X O . X X O O O X . . X X O . | $$ | O O O . . O X . X X . X X O O O 2 . . | $$ | O . O O O O X O O X X . O O X . . . . | $$ | . O X X O X O X . O X O O X . O . . . | $$ | O X X X X X O O O , O X . X X X X . . | $$ | O O O X . X X O O O O X . . . . . . . | $$ | X 1 . X . . . X . . X . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Other semi-final: CGI beats TianRang: interesting to see CGI do the fast-paced semi-sacrifice of corner in lower left like humans do recently, a nice fight in lower right corner which CGI then added a move to to prevent white linking up to corner on 1st line or centre sentes, and a cute 2nd line attachment to live in top right. CGI does seem to have a habit of surviving with small groups.

Uberdude · **#16**

Final: DeepZen vs CGI. The fighting over Zen's reduction of CGI's moyo was the focus of this game. Zen allowed itself to get cut and then settled the lower half and ran out with the top, only to sacrifice a large part of it for a some lower left territory. More fighting followed but Zen held on to a decent chunk of the top side without anything dying. But perhaps black 203 was a mistake in a leading position and it should have played the big reverse sente block at a7? Perhaps this was due to incorrectly thinking it was larger than it really based on the follow-up cut of 233 capturing 3 stones when it fact it didn't due to a snapback?

pookpooi · **#17**

What surprised me the most (but may not for other people) is Kong Jie choose CGI as his partner tomorrow. I thought he gonna choose FineArt for sure even if it placed the third.

Vargo · **#18**

Do we know the hardware used by DeepZen, FineArt, CGI, etc ?

Gomoto · **#19**

I wanna buy it. NOW! Deep Zen that is.

alphaville · **#20**

Gomoto wrote:

I wanna buy it. NOW! Deep Zen that is.

How much would you pay for it?

The 1st World AI Go Open 2017

Who is online