Leela zero Go forum
15 replies. Last post: 2018-03-18Reply to this topic Return to forum
Carroll at 2018-01-31
There is a nice project at designing a neural network to mimic what Deepmind has done to produce AlphaGo zero.
See http://zero-test.sjeng.org/ and https://github.com/gcp/leela-zero for a description and current evolution.
This is a smaller network that the one used in AlphaGo because it would take too much resources to train a big neural network.
But if you have a recent GPU and want to give some of its power to help the project you can let it play games against itself for best release of the Neural Net, by using the included autogtp tool on your computer.
If enough people play games, it may reach a professional level.
I have it running on my laptop (CPU only) on latest network ; if you want to test its strength (2733 Elo on CGOS, ~ 1dan ?), on 19x19, please invite me to a game.
Here is a spreadsheet trying to evaluate its current strength: https://docs.google.com/spreadsheets/d/1QhYWYQvXBhI5eHJm45fvf97CaV2rRmgrATwVgBFGKoU/edit#gid=806323235
Carroll at 2018-02-06
If you want to follow Leelaz games here against a 5.4 dan and make predictions, here they are:
https://www.littlegolem.net/jsp/game/game.jsp?gid=1923424#1923424 Carroll vs. eastknight1999
Move: 52Leelaz-last network
https://www.littlegolem.net/jsp/game/game.jsp?gid=1923424#1923425 Carroll vs. eastknight1999
Move: 50Leelaz-last network
Unfortunately, Eastknight does not speak English, so no in game reactions.
Carroll at 2018-03-02
Leela seems to have trouble to read this position with Kos and thinks it will win.
EastKnight says the lower left black corner is dead as 4-stones.
Can someone explain if there are enough ko threats ?
urmaul at 2018-03-02
Game link: https://littlegolem.net/jsp/game/game.jsp?gid=1923424 With Japanese rules, the lower left corner is just dead, like eastknight says. White doesn’t have to play a Ko in the lower left. That’s because in the evaluation of life&death in the end of the game, nonlocal Ko threats may be ignored. Maybe https://senseis.xmp.net/?BentFourInTheCornerIsDead helps. With Chinese rules, the lower left corner is also just dead, because white could – without losing points – remove all ko threats, before starting the Ko.
urmaul at 2018-03-02
With Japanese rules, the score should be 88.5 for white (white needs a move like a13), 86 for black (black needs a move like f7). With Chinese rules white has also won. As far as I know you’re playing with Japanese rules. Surely Leela Zero knows only Chinese rules. Though white has won in both rulesets, the correct way to prove the black corner to be dead in Chinese rules is suboptimal play in Japanese rules and may lead to a loss. I suggest that you ignore Leela Zero and either resign or play on, by filling f7 and, if you like, neutral points, and then counting the black group as dead.
Carroll at 2018-03-02#1923424 Carroll vs. eastknight1999
Leela zero plays according to Tromp-Taylor rule, near to NZ rules, an easy for computer rule-set, nearer to Chinese rules: https://senseis.xmp.net/?TrompTaylorRules
It says: "The only rule on scoring is rule 9, therefore both players should capture all dead stones before passing", that is why Leela is trying to continue to play and don’t think sacrificing stones is costly.
But it also was trained with a komi of 7.5 which should help keep a margin of 1.
I will keep playing and will report when it will change its mind about the outcome... and will then pass.
urmaul at 2018-03-02
I’m afraid it might take hundreds of moves until LZ sees it has lost. It might also pass at some point, confident it has won.
Carroll at 2018-03-05
Before move 300, Leela Zero has eventually managed to see that it is a lost game with win probability dropping to 5% (from 52% and 28% in last two moves).
Congrats to Eastknight for this fight where Leela consistently reported winning probability greater than 95%.
The Network has recently been trained with 10 blocks, so it should be able to better read ladders and this sort of configurations.
We are still far from AlphaGo 20-blocks and 40-blocks versions.
urmaul at 2018-03-06
It wasn’t a problem, because, since black played inside it’s own territory, white could afford to do the same, and white didn’t care about the winning margin. In the end LZ could see that the ko was unavoidable and there were no ko threats.
With help of Leela (not Zero), I have tried to analyze when LZ lost the game. I think black 155 at d15 was a big mistake, and black should just have protected at d2. In the resulting position Leela actually wants to resign for white. Black isn’t in a hurry to play d15 (for white, to play d15 and go for a ko with the sequence d15 e15 e14 seems too risky). I think black should also have played a1 at some point before white played there, so it’s a ko instead of just dead. After white played a1 to kill the black group, I believe LZ was still slightly ahead, but didn’t play it’s best, confused by the lower left corner.
Anyway, I think this LZ (still the old version from a month ago?) is already very strong, and I wouldn’t stand a chance. Did LZ use only 1600 playouts per move?
Carroll at 2018-03-06
Thank you for your analysis.
Yes for move 155, Eastknight said the same thing so I looked at the evaluation and d2 was the second best choice, just behind, and a4 is a possibility too:
F14 -> 2997 (V: 98.57%) (N: 9.50%) PV: F14 F13 G14 G13 H15 E14 D15 F15 C15 B13 B16 A16 B18 C16 S13 R8 //
D2 -> 1279 (V: 97.20%) (N: 28.15%) PV: D2 R8 S7 R2 S2 Q2 S13 R10 S9 Q10 F14 H11 H10 //
H15 -> 493 (V: 98.14%) (N: 3.84%) PV: H15 G14 F13 G13 H11 H12 S13 R8 S7 R10 S9 Q10 Q2 P2 //
A4 -> 416 (V: 97.53%) (N: 7.11%) PV: A4 D2 C3 R8 S7 R2 S2 Q2 F14 F13 G14 .
Yes it is with 1600 playouts, except for the last moves where I wanted to let MCTS help the network to “see”, but I updated the network weights to the best one, which has the drawback that the new network loses the analysis in the tree and has to rediscover the position.
Carroll at 2018-03-17
Haylee(9P) go player has played Leela Zero and commented on Youtube https://www.youtube.com/watch?v=_qVlpJ_fWzo&feature=youtu.be
The hardware consists in 4 GPU recent cards, the time is quite short.
Carroll at 2018-03-18
Erratum, Hajin Lee is a retired 4P, see https://en.wikipedia.org/wiki/Lee_Ha-jin.
apetresc at 2018-03-18
And now, just a few hours ago, Leela also won against Zhang Li, rated #52 in China, and Zhang Tao: