Hi,

I am very happy to announce that Hajin Lee will play a live commented game against Crazy Stone on Sunday, at 8PM Korean time. The game will take place on KGS, and she will make live comments on her youtube channel. That will be very exciting. I am looking forward to this event.

Haylee's youtube:
https://www.youtube.com/c/HayleesWorldofGoBaduk

Rémi

We did not discuss time settings yet. I will probably use an Amazon server (18 cores). Crazy Stone does not use the GPU.

Cool! Will this be an even game?

P.S. Are you aware of any plans for another edition of the Code Centric Go Challenge (https://go.codecentric.de/)? For the unaware in October 2014 FJ Dickhut German 6d played CrazyStone and won 3-1, and then in October 2015 played Zen and won 3-1 again. I suspect the new deep learning CrazyStone could win the match now.

Yes, even game.

I am not aware of a new edition of the Code Centric Go Challenge.

Well, if CrazyStone can win against Hajin Lee in the even game then there's no need for another go centric challenge

My prediction would be for Hajin to win this one, though perhaps the distraction of live commenting will reduce her level and give CrazyStone a chance. And even if CrazyStone does win I'd still like to watch more serious human vs computer matches.

Yeah, more human vs. Computer game is always welcome by me, remi once stated that he doesn't fond of it very much though.

I don't thinking making a live commentary is too distracting, assuming the player talks to a microphone rather than type. On China's WeiqiTV, Meng Tailing 6-dan has been doing this for ages, living commenting hundreds of his internet games.

Well, CrazyStone got crushed in a rather disappointing match. In the first corner it did the unusual (common is 1970s Japan) attach on top for move 18 which Hajin didn't seem to know, but she quickly found the usual continuation of hane and iron pillar. CS then played r6 to cover the cut which is unusual, normal is either q7 (aims at r7 taking lib for q9) or descend to r2 key point, which allows black to cut, but she is reluctant to do so as it makes r8 a stupid empty triangle so s5 slide is normal.
p12 counter hane was a nice move that impressed Hajin, and then in the top left CS played the c12 hane kyu players' favourite, but she also went for the simple punishment of c18 descend rather than d18 which disappointed me somewhat. Then CS choose the complicated thickness joseki variation but crazily played b3, presumably because it missed the e8 clamp tesuji and the game was over by move 65. I would expect a MCTS bot of its level to not miss this (even on only 20 seconds a move) so my guess is the neural networky bits may have been too prominent, perhaps because if d11 was at d10 the clamp wouldn't work and it got confused. The rest was nonsense, including playing itself into a ladder.

Thanks for your comments.

I was also very disappointed by the way Crazy Stone played. I had played plenty of games on KGS to calibrate levels before the release, and it did reach 7d with a 4-core machine at 1 second per move. I expected a 18-core server with 20 seconds would be considerably stronger.

From my quick analysis, it seems that the neural network makes the search too selective, so searching deeper does not help. You correctly analyzed that it did not anticipate B E8. I will tune selectivity at longer time controls, to see if I can fix it. I had done some testing and tuning before the release, but mostly in self-play. Playing humans reveals interesting weaknesses that are not visible otherwise. I don't have strong enough gtp opponents besides my previous version, unfortunately.

Anyway, I will try hard to fix these problems.

Remi, you usually said that you are not fond of human vs computer game as you found that computer vs computer is more valuable. May be this is the perfect time to appreciate human vs computer game

You could always ask Haylee to play CS in private so that you could find the best setup.

E8 is a standard technique to capture opponent stones, next to ladder, net, etc.

Actually I liked how CrazyStone played, other than the blunder. It was similar to AlphaGo's blunder in game 4 with Lee Sedol, but on a different scale of course. Nevertheless, it's about missing a move in fight.

I think CrazyStone played solidly, and it impressed Hajin in 3 different instances, I think that is very good!
After the blunder, it played a few MCTS-style nonsense moves, but we learned from AlphaGo that that is state of the art

Someone wrote on computer-go mailing list that the commercial version of CrazyGo does not make this blunder in the same position - Remi (or somebody who already bought it), can you please clarify/confirm if that is the case?

_________________
Sorin - 361points.com

I, too, was curious about this and had CrazyStone analyze the position after move 63 (right before the blunder). At first b3 is only 5th in its move list, quickly rising up the ranks until it's solidly cemented in 1st place after a few seconds. If you let it analyze further however b3 will eventually drop down again in favor of c9. So it must have made its move in the time frame (~10-20s) when it was still convinced that b3 was best.

- Kap

Edit: Although it appears to favor b3 again after about 500.000 playouts. And even after ~3mio playouts it's convinced that P(B wins) < 43%. It's only after you manually enter e8 as response to b3 that CS realizes its mistake and P(B wins) immediately jumps to about 54%.

Setting up alternate boards with this joseki does not usually favor this weird descent, but there's no question that crazystone does not see this clamp; though once the clamp is made, it can see the disaster. It sees the clamp after the atari at G6, but this seems like a bad exchange.

What's funny is that it saw the right combination in game 4 of LSD-AlphaGo. and generally CS has the similar estimate compared to the estimate AlphaGo made! (if you have those leaked graphs)

There is one more joseki I noticed in which it makes a mistake very often. It's a pity that Hajin played one of 2 josekis unknown to CS. Also, CS first edition also needs more thinking time for the endgame, but besides that, opening and everything else seems legit and reasonably strong even for 6d KGS:)

I've given the game an 8h analysis on my 4790k in CS:DL. It really does seem like the parameters were a bit off for the match, as this self-analysis confirms the blunders.

Did you manually make the .sgf with the analysis information or is there a way to get CS:DL to create an .sgf with the analysis info in it? I've tried finding a way but no luck so far.

I've adapted this Python script to my liking: viewtopic.php?p=204523#p204523