Zen vs Yonil Ha 6p (and you can play against it in KGS now)

[Shenoute]

Sai lost a second game (third one is going on right now).

[Uberdude]

So Blackie played as BIBA and Zen won again!:

[Shenoute]

Move 14 (p3) in game 2 and the hane at f2 (move 16) in the last game make me feel like Kim was trying very (too?) hard. Or are these possible tactics (I'm not very well up up to date on professional josekis)?

[Uberdude]

Move 14 (p3) in game 2 and the hane at f2 (move 16) in the last game make me feel like Kim was trying very (too?) hard. Or are these possible tactics (I'm not very well up up to date on professional josekis)?

I certainly agree they are trying hard rather than playing a leisurely game (more so the hane) but both have been played professionally before. In fact I studied the first shape as it appeared it one of my recent games: http://www.lifein19x19.com/forum/viewto ... 07#p211307 in which I made the leisurely 2-space extension which allowed the shimari. There were only a handful of similar sides in pro games and approach was a bit more common. I can see the logic in approaching as allowing black 2 big solid corners does feel a bit soft. As for the counter hane in game 3, the solid wall building is much more common but you do see that move too so it's not such an obvious overplay that pros don't do it. Probably helps if white already has a stone on the left side, but check out this identical pro game, black defends the corner though instead of fighting like Zen: http://ps.waltheri.net/database/game/44989/.

[Shenoute]

Thanks! I'll have a look at the games.

[pookpooi]

Blackie finally won after the fifth try, he commented that the game would be very close if Zen play normal yose
But all can guess what Zen did in that game Now BIBA team will rest for a bit...

[Uberdude]

That game was Blackie's 4th and first win (2 loses as sai, one as BIBA). He also won the fifth (below) with a comeback confusing Zen in the endgame with double ko and seki so is 2 wins 3 loses vs Zen now. I wonder if he is learning better how to beat it, despite the amazing achievements of (computer) neural networks they still have a way to go to match humans' ability to learn from just a few games instead of thousands/millions.

[Uberdude]

Another win for Blackie as BIBA (his son Hyunbin played most of and lost one). So that's 2 loses as sai, 3 wins 1 loss as BIBA. Contrary to what I said elsewhere, Zen played the large avalanche this game and did some nice attacking to build the right side moyo. Blackie made a big ko and Zen's 3-3 threat wasn't so amazing, and then Blackie killed it in a semeai on the top side that perhaps Zen didn't realise it was dead (it had eye but lost versus huge no-eye group). After that it played rather poorly.

[Uberdude]

A few more interesting games from Zen showing how it can be beaten. First zenbieland got a lead in the first 'joseki' when Zen played a cut which would be a nice probe if it were sente, but it was too slow and the mystery human captured the cutting stones with a classic wedgy squeezy ladder tesuji that Zen really ought to have seen. I think old MCTS without DCNN 5d Zen wouldn't have made this mistake, is it the neural networks giving it blind spots of devastating tesujis not seen in pro games because pros don't get into those situations (similar to CrazyStone's blunder vs Haylee, and even a bit like AlphaGo's mistake against Lee Sedol's famous wedge, though that did need some deeper reading). Zen seemed to recover somewhat but when it tried and failed to kill white's centre group it looked to be game over, but then suddenly it killed white on the right side because he had neglected to make some forcing exchanges before to ensure life. It was then very close, but Zen ignored a sente move for a kyu level life and death problem and game over!


Next in the mystery guest list was Godmoves, who beat Zen twice in whole-board fighting whilst blitzing, one starting at tengen. These games it feels Zen was just outplayed by a stronger player, rather than losing because of some blunder.

[Shenoute]

A few more interesting games from Zen showing how it can be beaten. First zenbieland got a lead in the first 'joseki' when Zen played a cut which would be a nice probe if it were sente, but it was too slow and the mystery human captured the cutting stones with a classic wedgy squeezy ladder tesuji that Zen really ought to have seen. I think old MCTS without DCNN 5d Zen wouldn't have made this mistake, is it the neural networks giving it blind spots of devastating tesujis not seen in pro games because pros don't get into those situations (similar to CrazyStone's blunder vs Haylee, and even a bit like AlphaGo's mistake against Lee Sedol's famous wedge, though that did need some deeper reading). (...)

I think the "joseki" first appeared in a game played yesterday between Zen and kidnap (I think it was him), with both players missing the tesuji. Blackie was kibitzing and pointed at it. It seems his comment did not go unheeded

Edit. It was in this game with wvxxe

[Mike Novack]

I want to point something out.

This isn't quite a matter of opponents learning how to defeat the PROGRAM but already understanding that the TRAINING of the neural net has mainly been from the database of exisitng games, so their chances are better going off in directions unlikely to lead to the sorts of positions that would have been in the training database.

The "fix" (from the program's point of view) isn't changes in the program but more training on different data. This is not unrelated to poor performance realtive to large handicaps. That too not yet in the training.

[Uberdude]

Playing one stone on tengen wasn't giving enough of a handicap for GodMoves so he did a one point jump around it and still won (blitizing of course). It's got the attention of Ohashi Hirofumi 6p who linked this video (with comical music!) of the first 2 games on youtube: https://www.youtube.com/watch?v=A2u6wJy9xls. 3rd game:



P.S. Blackie lost 2 games today.

[pookpooi]

Rumor that GodMoves is AlphaGo is spreading like fire for example

https://translate.googleusercontent.com ... 3FiyFpOnUg

But Hideki tweet that he personally doesn't believe that GodMoves is AlphaGo.

[Uberdude]

Some people on KGS said that but it's easy to say bollocks. Highly unlikely IMO, without even looking at the games the idea of AlphaGo secretly playing on KGS goes against DeepMind's pattern of behaviour. But look at the games, particularly the opening! Clearly a strong human taking the piss. I don't think AlphaGo has developed a sense of humour yet.

[dust]

Language, Timothy..