lc0 memorizes patterns and was trained first with random move so probably each starting position even if of course she plays more and more the same openings with more training. But she can plays every positions thanks to that ( patterns and all opening trained ).
I call CCCC fair condition for Leela because she has enough power this time and the ratio for her seems good enough to compete.
And for your theory, you need 3 tests to be sure.
1 test with book for both engine
1 test without book for both engine
1 test with no book for Leela and a book for engine
And then you can compare if there is a big difference in elo. It might, but we can't know for sure until tested right ?
CCCC Rapid Rumble results simulator
Moderators: hgm, Rebel, chrisw
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: CCCC Rapid Rumble results simulator
Seems a nice idea.Laskos wrote: ↑Sat Sep 08, 2018 11:16 amSure, the rating in these particular conditions could be estimated as sayJJJ wrote: ↑Sat Sep 08, 2018 9:43 am Lc0 is now on part with Shredder for #4 spot and I think it will take this spot. After all, many test have shown Lc0 being the true number 4 with fair condition and in this CCCC it is also the case.
So lc0 for the #4 spot ! Your simulator is really innacurate. CCRL elo matters way more than the rating in during the tournament, except if the tournament reach his ends.
0.5*x*TP + (1-0.5*x)*CCRL
x is the fraction of games played out 46
TP is tournament performance
CCRL is the CCRL rating
This way the rating is always more determined by CCRL rating, from say 3 times more importance in the middle (23/46) of the tournament, to equal contribution of TP and CCRL at the end. The rating is then applied to remaining games, and sure starting with the current points in tournament.
This is just an example. I am countryside on vacation right now, so I can do no simulations, just write here mostly on my phone.
Lc0 estimates are very hard, the conditions matter a lot, CCRL is irrelevant, but as you observed, CCCC seem to be a fair tournament for Lc0 (pity it doesn't scale much beyond 2 GPUs), and in my fair conditions Lc0 is the fourth engine (at worst fifth, behind the last Fire), which is close to what we observe here. TCEC conditions were awful for Lc0.
-
- Posts: 10279
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: CCCC Rapid Rumble results simulator
Good conditions for Leela may be unfair for other programs.JJJ wrote: ↑Sat Sep 08, 2018 12:31 pm lc0 memorizes patterns and was trained first with random move so probably each starting position even if of course she plays more and more the same openings with more training. But she can plays every positions thanks to that ( patterns and all opening trained ).
I call CCCC fair condition for Leela because she has enough power this time and the ratio for her seems good enough to compete.
And for your theory, you need 3 tests to be sure.
1 test with book for both engine
1 test without book for both engine
1 test with no book for Leela and a book for engine
And then you can compare if there is a big difference in elo. It might, but we can't know for sure until tested right ?
It is not clear what is the meaning of fair when there is no equal hardware.
-
- Posts: 708
- Joined: Mon Jan 16, 2012 6:34 am
Re: CCCC Rapid Rumble results simulator
All competitions doent need to be fair!Uri Blass wrote: ↑Sat Sep 08, 2018 12:54 pmGood conditions for Leela may be unfair for other programs.JJJ wrote: ↑Sat Sep 08, 2018 12:31 pm lc0 memorizes patterns and was trained first with random move so probably each starting position even if of course she plays more and more the same openings with more training. But she can plays every positions thanks to that ( patterns and all opening trained ).
I call CCCC fair condition for Leela because she has enough power this time and the ratio for her seems good enough to compete.
And for your theory, you need 3 tests to be sure.
1 test with book for both engine
1 test without book for both engine
1 test with no book for Leela and a book for engine
And then you can compare if there is a big difference in elo. It might, but we can't know for sure until tested right ?
It is not clear what is the meaning of fair when there is no equal hardware.
In top gaming competitions, the conditions are not always fair but people still enjoy (one player will randomly sprawns in bad location and the other one get good location, random weapons etc)
Also, in Formula one competition, all driver uses different cars. For fairness, " All drivers must use exactly the same car, must have exactly the same body weight, exactly the same wind effects and ground condition etc." whereas those are theroetically and practically impossible.
In general, "when unbiased general public see the conditions are fair, it can be assumed as fair" .
The same in Courts, when most judges/juries see it is fair, the outcome is fair"
-
- Posts: 4313
- Joined: Tue Apr 03, 2012 4:28 pm
Re: CCCC Rapid Rumble results simulator
290 rounds so far
"broken" - Milos
"LOL" - Laskos
"really inaccurate" - JJJ
"ridiculous" - Milos
"broken" - Milos
"LOL" - Laskos
"really inaccurate" - JJJ
"ridiculous" - Milos
Code: Select all
Engine Tournament Init 1st 2nd 3rd 4th 5th 6th 7th ....
Houdini 3452 3400 0.51 0.31 0.16 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Stockfish 3451 3439 0.34 0.40 0.23 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Komodo 3436 3404 0.14 0.28 0.51 0.05 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Shredder 3346 3287 0.00 0.00 0.04 0.34 0.29 0.20 0.09 0.03 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Lc0 3346 3300 0.00 0.00 0.03 0.32 0.28 0.22 0.10 0.04 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Fire 3348 3326 0.00 0.00 0.02 0.22 0.28 0.28 0.13 0.05 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Booot 3314 3276 0.00 0.00 0.00 0.04 0.09 0.18 0.35 0.22 0.09 0.02 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Ethereal 3304 3283 0.00 0.00 0.00 0.01 0.04 0.09 0.22 0.37 0.18 0.06 0.02 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Andscacs 3268 3244 0.00 0.00 0.00 0.00 0.01 0.03 0.08 0.21 0.41 0.16 0.06 0.03 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Fritz 3231 3200 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.04 0.12 0.26 0.23 0.15 0.09 0.05 0.03 0.02 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Xiphos 3218 3179 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.03 0.10 0.25 0.23 0.16 0.10 0.06 0.04 0.02 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Texel 3204 3144 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.03 0.10 0.16 0.19 0.17 0.13 0.09 0.06 0.03 0.02 0.01 0.00 0.00 0.00 0.00 0.00
Pedone 3194 3090 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.07 0.12 0.16 0.18 0.15 0.12 0.08 0.05 0.03 0.01 0.00 0.00 0.00 0.00 0.00
Gull 3200 3184 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.04 0.07 0.11 0.14 0.16 0.16 0.12 0.09 0.05 0.02 0.01 0.00 0.00 0.00 0.00
Vajolet 3177 3101 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.03 0.06 0.10 0.15 0.17 0.16 0.13 0.09 0.05 0.03 0.01 0.00 0.00 0.00 0.00
Fizbo 3185 3259 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.02 0.04 0.07 0.10 0.14 0.18 0.16 0.13 0.08 0.04 0.02 0.00 0.00 0.00
Laser 3167 3226 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.03 0.05 0.07 0.11 0.15 0.18 0.18 0.12 0.07 0.03 0.00 0.00 0.00
Arasan 3152 3123 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.02 0.04 0.06 0.10 0.14 0.19 0.19 0.14 0.08 0.04 0.00 0.00 0.00
Nemorino 3123 3099 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.01 0.02 0.04 0.07 0.12 0.19 0.25 0.28 0.01 0.00 0.00
Wasp 3112 3041 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.01 0.03 0.06 0.11 0.21 0.27 0.30 0.01 0.00 0.00
Ivanhoe 3116 3115 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.02 0.03 0.06 0.11 0.19 0.26 0.30 0.01 0.00 0.00
Senpai 3028 3112 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.76 0.17 0.04
Nirvana 2998 3186 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.15 0.53 0.32
Crafty 2961 3013 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.07 0.29 0.64
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: CCCC Rapid Rumble results simulator
Lc0 is now 4th and will probably stay at this rank. Told you he was favorite for it. And the top 3 will stay in top 3. Stockfish still favorite for spot 1. Fire will probably catch up Shredder anytime soon as well and finished 5 and shredder will be 6. After that I can't say
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: CCCC Rapid Rumble results simulator
It is easy to predict even with a short tournament , because each engine is ranked with a good margin over the other, except Komodo / Houdini. So Stockfish is number 1 with a good margin, Lc0 number 4 with a good margin, then Fire, then Shredder, so the probability to have the number in the good order is pretty high. Higher than your simulator who needs to take in account the elo much more than the starting ranking of a tournament.
-
- Posts: 1470
- Joined: Mon Apr 23, 2018 7:54 am
Re: CCCC Rapid Rumble results simulator
Yeah, the public/jury has to be unbiased or the outcome won't be fair. So jury selection, etc...Nay Lin Tun wrote: ↑Sat Sep 08, 2018 2:24 pm In general, "when unbiased general public see the conditions are fair, it can be assumed as fair" .
The same in Courts, when most judges/juries see it is fair, the outcome is fair"
The public/jury needs to know the facts too.
I know from the other thread that lc0 CCCC hardware is 60 times more expensive than hardware for the other engines. I don't think if they knew the public would think that is fair conditions.
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: CCCC Rapid Rumble results simulator
I am not sure about these prices, but the conditions seem fine for what will appear in just few several months. NPS Lc0 shows are not that impressive, and would be equivalent to 4 or so 1080Ti, if Lc0 scaled well above 2 GPUs (which it does not). In 2-3 months we will see the following, equivalent to CCCC conditions:jp wrote: ↑Sun Sep 09, 2018 12:28 pmYeah, the public/jury has to be unbiased or the outcome won't be fair. So jury selection, etc...Nay Lin Tun wrote: ↑Sat Sep 08, 2018 2:24 pm In general, "when unbiased general public see the conditions are fair, it can be assumed as fair" .
The same in Courts, when most judges/juries see it is fair, the outcome is fair"
The public/jury needs to know the facts too.
I know from the other thread that lc0 CCCC hardware is 60 times more expensive than hardware for the other engines. I don't think if they knew the public would think that is fair conditions.
CPU: AMD Threadripper 2990X (32 cores at 3.4-4.0 GHz) ---- $1700
GPU: 2 of Nvidia Turing RTX 2080Ti --- 2x $1200 ---- $2400
RTX 2080Ti will probably have the full fp16 support, so it will be much faster than 1080Ti for Lc0, and 2 of these will probably be even more than what CCCC shows for Lc0 (it doesn't scale well to 4 GPUs).
So, I think pretty fair price-wise conditions in the frontier hardware appearing very soon. Also, the tournament performance is similar to that shown in pretty fair price-wise conditions at much smaller scale, say GTX 1060 versus 4 core i7.
Last edited by Laskos on Sun Sep 09, 2018 1:46 pm, edited 1 time in total.
-
- Posts: 1796
- Joined: Thu Sep 18, 2008 10:24 pm
Re: CCCC Rapid Rumble results simulator
2080 Ti will be much faster than 1080 Ti for the reasons you give.
But I don't think 2x 2080 Ti will be faster than 2 x Titan V (or V100) according to the specs.
The Titan V will still be about 10% faster.