Leela has a new version

[pookpooi]

Can you produce more game between them? Like 9 more matches so that we can measure how good leela is against other bot on the same hardware

[Gomoto]

It is fun to watch strong players (Zenith and Leela). But I am not quite sure it is that much fun.

[Pippen]

Leela lacks a self-play modus.

[pookpooi]

It is fun to watch strong players (Zenith and Leela). But I am not quite sure it is that much fun.

Leela 0.9 is now on cgos so we'll know tomorrow if it's stronger than zen or not

[yharzu]

Leela 0.9 is now on cgos so we'll know tomorrow if it's stronger than zen or not

ah, cool. Can you keep us up to date?

[Garf]

Leela lacks a self-play modus.

gogui can do this I believe - shame it's not developed any more.

[Garf]

Leela 0.9 is now on cgos so we'll know tomorrow if it's stronger than zen or not

Are CGOS ratings reliable?

3 Zen-13.1-1c0g 3791 88 88 252 94% 3143 0%
4 Zen-13.3-1c0g 3762 51 51 500 89% 3159 0%
...
11 Zen-13.7-1c0g 3676 49 49 582 90% 3002 0%
13 Zen-13.9-1c0g 3645 43 43 486 74% 3219 0%
14 Zen-13.8-1c0g 3637 70 70 259 88% 3001 0%

Zen's authors should stop developing it, every new version is getting weaker

72 Leela-0.8.0-4c1g 3162 76 76 217 60% 2906 0%
73 Leela-0.7.0-4c 3154 79 79 213 71% 2642 0%

No real difference and older version wasn't even using GPU. Leela's author should also stop "improving" it

29 Rn.3.6-4c 3439 70 70 215 71% 2968 0%
30 Rn.3.7-4c 3430 45 45 523 70% 2985 0%
32 Rn.3.3-4c 3426 29 29 1200 65% 3011 0%
33 Rn.3.5-4c 3411 68 68 204 68% 3006 0%
43 Rn.3.4.1-4c 3360 65 65 201 59% 3113 0%
49 Rn.3.8-4c 3295 71 71 229 65% 2889 0%

"Rn" author seems to have found the trick after a lot of trying because Rn.3.9 was finally rated higher.

I don't know what causes this but either CGOS is not reliable or authors of Go bots have no idea what they are doing.

[pookpooi]

My interpretation is that Zen developers failed to improve after 13.3, that's why they push 13.3 1c1g (to try to see if gpu implement correctly) and will probably use this version in March big matches.
There's currently 3 Zen version running on cgos now, 13.3 1c1g, 13.9 1c1g (just in case if gpu implement will be super effective that will surpass 13.3, which sadly it isn't) and 12.0 1c0g as anchor.

[Garf]

I think your Leela just beat the very latest Zen, who's also using GPU acceleration?

[Garf]

My interpretation is that Zen developers failed to improve after 13.3.

12.5 is actually the highest rated one. Very sad that Zen didn't improve, then, but I do wonder if it's really true.

Code: Select all

1       Zen-12.5n-1c1g          3815 	147 	147 	188 	97% 	2768 	0%
2       Zen-13.1-0.5-1c0g       3799 	52 	52 	881 	95% 	3069 	0%
3       Zen-13.1-1c0g           3791 	88 	88 	252 	94% 	3143 	0%
4       Zen-13.3-1c0g           3762    51      51      500     89%     3159    0%

[pookpooi]

13.3 1c1g is not in BayesElo page yet, you can find it here

http://www.yss-aya.com/cgos/19x19/standings.html

I think it'll appear tomorrow.

[johnsmith]

0.9.0 is stronger than Zen in my personal opinion.

One interesting thing: neural net gives around 36-37% at 3 handicap stones and 0.5 komi, but if you increase komi to 10.5, % will be much lower. The same thing happens in even game with/without komi. Why is that?

[Uberdude]

0.9.0 is stronger than Zen in my personal opinion.

Do you mean the latest commercially available version (about 6d ama strength) or the latest on KGS/Tygem (about mid pro strength)?

[pookpooi]

We'll know it tomorrow, but the comparison is not quite fair cause Leela 0.9.0 is 1c1g while Zen 11.4 (aka Zenith Go 6) is only 1c (gpu not supported yet).
From what I look like now it's close.

[johnsmith]

0.9.0 is stronger than Zen in my personal opinion.

Do you mean the latest commercially available version (about 6d ama strength) or the latest on KGS/Tygem (about mid pro strength)?

I meant Zen6 (7 dan). Because in a couple of games Zen was saying it was even (around 50%), while Leela's neural net % was extremely high (80%) and MC % was like Zen's. In 10-15 moves, Zen's estimate dropped down to 35%.