User tournaments

Anyway, congratulations lazyplayer! I was afraid leela is like a 2900 player now, but following this game I guess maybe it's only 2400 or so. This means there will be some real competition in the next championship. Exciting!

shalev, i've chatted with author of leela_bot. He told me that leela was fully aware a12 was horribly strong but there is a bug in the swap code. He'll fix that bug and then we'll have to win without an handicap. In that game he was basically playing with a small handicap due to the first stone. ;)

The game you lost with leela also is due to a small handicap that you gave yourself with the first stone :D

shalev, I did know a12 was strong but I didn't know that leela was aware of it. I though it had some secret plan to refute my obvious attempts. Instead the obvious attempts worked. Very surprising. I would say, the correct mindset is this, if a position is theoretically winning, then we've to win. Aim at perfect play because with leela you may very well need it.

Hmm, it's still impressive that you beat leela. That's a handicap, yes, but not a gigantic one. I'm pretty sure I can beat a 2200 player with such a handicap without too much trouble. It's probably worth like 50 elo, or maybe 100 max. I'm assuming you're still better than me (though we didn't play in a while), but I'd imagine I can beat you 1 in 3 times without a handicap or 1 in 4 times with a12. In other words it's not THAT bad.

I agree with shalev, I don't think it a12 is that strong. lazyplayer should give himself more credit for winning! Congrats! I might leave the "bug" in place a little longer, it certainly makes for more varied/interesting opening play :)

Me and lazyplayer we've reached a position in live play tinight, which is very equal and also is played more / less properly from the start. We're unable to solve it.

In case anybody wants to see for himself, here it is

Especially if bot owners are kind enough to analyse it, we would be very grateful :)

Or this

Looks like white is ahead to me.

Why does black refuse to play josekis? Is that you lazyplayer? Play more josekis :P

It's me, I think josekis can and should be delayed.

Anyway precisely now the black will play d4 joseki to almost any white move.

White might be ahead but if he follows a typical joseki then black wins.

Well, it was late in the night and we were both tired. I missed h6 like a stupid, and he had previously missed g7 like a stupid. But with both g7 and h6 in place, it's probably white win again. We swapped colors several times during that game, so both white and black have stones suggested by both.

Overall it was an experiment to test more central play like d8-g7 couple. It didn't work well for black. White refuted black without complications.

laz, g7 is move #5 so you may have meant something else

anyway I don't agree with the conclusions and it's not even proven to be white win ;)

Arek, no i mean move #5 is experimental, obviously josekis would be worth consideration at #5...

What you do after h6? I think it's game over?

Here you go - 65% at move 35 and 83% at move 54 for white (with some bias).  Not sure it helps.

Thanks @shalev, will dig into why it thinks previous game analysis was a win for white... hopefully it was my fault, and not the bot's.

gzero, in fact we already tried 38 g5, and it's already refuted. If bot would play 38 g5, then it would simply lose to 39 d4 (if I recall correctly).

Thanks for sharing your data anyway.

leela seems to be able to untangle endgame messes, let's see if he says something on this.

gzero, your bot is written from scratch? leela has much easier job because it's just copy and paste from Go!

gzero is written from scratch.  It isn't a hex bot as such ... it is - in theory - game agnostic.  I didn't write any hex specific code (someone else wrote the hex rules in a very high level language for 9x9, I tweaked a few lines to make it 11x11 and 13x13)...  of course I had to train a neural network from scratch via self play, which means turning it on and leaving it for a couple of week or so!

I am considering starting a fresh network (so from zero again), but with different training parameters.

gzero, which technologies are you using? tensorflow?

lazyplayer, his code is available at https://github.com/richemslie/ggp-zero so you can see for yourself. Based on a quick look, it looks to be Keras only, no Tensorflow.

Wait, lazyplayer, are you sure? You think black wins after 38. g5 39. d4? I'll take white and play 40. d5

What's black next move? White seems totally winning to me.

To be honest I'm not sure because I've already forgot, but probably this was the problem.

But this is white win again, and yet we had tested this g3 stuff... maybe Arek will remember, he had found an amazing refutation, but was 2 AM so maybe I was just sleeping.

lazyplayer, leela doesn't like 38.h6. It thinks c7 or d4 have about a 70% chance of winning for black:

  c7 -> 1874762 (V: 71.67%) (N:  5.44%) PV: c7 c9 e8 h4 d4 d5 f6 e7 i5 f7 g6 h7 d8 f12 b6 c5 b5 c4 g12 d7 c8 j2 i4 f2 b4 c2 c3 e1 f3 g13 f13 d9 e2 f1 f5 h5 b9 c11 b12 d12 b11 e3 d2 d1 d6 b10 e9 e6 e5 g3 g2 f4 h3 g4 i3 i1 h2 g1 h1

  d4 ->  120294 (V: 69.63%) (N: 10.94%) PV: d4 d5 l4 k3 m2 k6 g5 h4 k4 j5 l6 m3 l3 l7 i5 h3 f4 h5 f7 g2 f3 f1 d2 e2 c5 e3 c6 d7 d6 f6 g6

  f4 ->     908 (V: 54.72%) (N: 13.60%) PV: f4 h5 k4 j5 d7 e7 h7 f6 e5 d8 b9 c8 b8 c7 b7 d5 c6 d6

  f6 ->     858 (V: 57.00%) (N: 11.11%) PV: f6 g5 k4 k3 l3 l2 m2 j5 d4 e3 c3 f5 l4

  l4 ->     800 (V: 53.75%) (N: 12.68%) PV: l4 k3 d4 e3 c3 h3 f5 d6 d5 e5 m2 k6 g3 g4 f4 f12

  k4 ->     536 (V: 60.64%) (N:  5.20%) PV: k4 k3 d4 e3 c3 h4 d5 d7 e6 d3 c4 d2

  d5 ->     525 (V: 59.33%) (N:  5.70%) PV: d5 d4 c4 d3 c5 c3 b4 b3 k4 k3 h4 h5

Maybe this , well i just go sleep bye.

leela, well if h6 loses, then what else?

shalev after your 38. g5 39. d4, 40. d5 leela likes c7

  c7 ->  108461 (V: 60.89%) (N: 12.03%) PV: c7 f7 f5 g3 b6 c3 c5 e2 e3 f2 k4 k3 h3 j2 b3 b4 l3 j5 i4 i5 g2 f3 g6 h5 g4 f6

Maybe 38 f6? :)

Ehehe no 38 f6 loses too, well i should shut up to avoid further stupid proposals :D

I suppose white is just losing at 38, indeed h6 doesn't work due 39 g5 , and white can resign.

leela seems confused in the original position, sometimes it likes g5

  g5 -> 1929451 (V: 46.99%) (N: 24.07%) PV: g5 d5 f12 c4 c2 g12 d3 f5 g3 l4 k6 f3 g2 h4 h5 i4 j3 f2 f4 k4 j5 b3 c3 b4 b2 d4 e3 i3 i5 g4 f7 e7 d9 c9 d8 f6 g6 b8 c8 b9 d6 e5

(before it switched to d5 as a response to g5 it had d4:

PV: g5 d4 h3 f5 e3 k4 k3 c3 f7 e7 g3 h4 g4 f4 c2 b3 b12 e10 d2 b2 h6 g6 f6 h5 i5 c12 d11 c11 f10 e6 d6 c8 d8 f3 e8 )

Sometimes it likes the wrong looking h3:

  h3 -> 1979539 (V: 44.56%) (N:  7.36%) PV: h3 f5 g3 c4 f12 d5 c3 b4 b3 f3 f4 d3 f2 e1 d2 g2 d7 e6 f7 e7 d9 c9 d6 c8 b7 e5 d8 e9 e8 g12 g5 g6 f6 l4 k3 m2 k6 k4 j5 h4

but it doesn't see a win for white. Looks like black might (barely) have it.

Ok, given that resign is never good, I'll take my chances with 38 f6, see you tomorrow ;]

Ehehe leela also doesn't know, nice, it's not that strong yet ;)

Leela, you realize white wins in that c7 line you suggested for black, right? The final position is this, from which white win is immediate.

Lazyplayer, OK, admittedly I didn't see some of your lines, but I still think white wins after 40.d5. I suggest 44.c9

This is maybe what Arek had found. It's same b6 stuff but played right.

Also, leela, in that c7 line white has an easier win; just 42. f3. Here is a sample line. I'd be curious if leela thinks black has any chance at all after 42.f3, because I can prove white wins there.

Shalev, sorry, i take back that. We'll have to wait for Arek (or for me to be in better shape) to show us the way (if it exists).

Shalev, i guess in our analysis we just missed 44 c9, my fault because I was playing white and it wasn't that hard to find. Congrats then.

actually leela is correct, d5 is the correct response to g5, however I still cannot solve it

1)

2)

3)

Arek, white wins in your second link: (2)

Ok...here we go! 26 new games on the radar!

This simultaneous double game format has a potential flaw in it. I hope no one copies the opponent's moves back to them in the other game. Maybe I should calm down?

That's always been a potential danger; people do it even when it's not a double-game format, they just copy the moves from all their white games into their black games, and vice-versa. It guarantees a 50% winrate no matter what, whether it's the same opponent or different ones.

shalev, yes indeed, thanks, game over here, d5 does good job but eventually white c9 wins in all cases (often as the only winning move)

But!!!

What if c10 is just a piece of garbage? Let's cheat a little bit and look at this

hmm, no, b12 and even b11 in place of c10 are equally garbage this time, game over

thanks leela for the analysis!

wait a minute... now I think lazyplayer was eventually correct by proposing f10 in place of c10

white cannot play d12

whereas counter-intuitive e11 leads to quite a diffrent game

no, no, no, sorry for the spam, I'm just talking nonsense, white plays d12 and wins, f10 discarded

Arek, moral of the story, we should go sleep early.

Idea of f10 was to keep choices, but there is no need to keep choice, black should go for the best. The best probably is this