TCEC Division 3 results simulator

[chrisw]

"Yes the game will be replayed. The logs made the reason very clear: 47519116 <lc0 16.10520(22): FATAL ERROR: Network error: Software caused connection abort"

Taken from the leela forum, in answer to the question will the hannibal v leela game be replayed.

That's from a TCEC admin, so is probably reliable. He said later that the game would be replayed at the end, but suggested he was going to try and fiddle around with the results data to see if he couldn't insert the replay sooner. So maybe sooner than the end.

[frankp]

Reading the leela forum seems to indicate that the two gpu box TCEC are renting has insufficient cooling.
Talk of setting both gpu to 85% and 65%. One is blowing hot air on the other.
Looks a mess. Not sure of the impact on leela and leela-DX if gpu suddenly throttle during a game; and whether this in any way explains the poor performance. (Leela-dx has a smaller NN so may not stress the gpu as much.).

Apparently the gpu drivers are beta too.

To be fair to tcec, they did say it was an experimental tournament, so glitches are to be expected I suppose.

[chrisw]

AFter 95 games, and (I think?) the LC0 Hannibal replay, promotion chances are according the Sim

Ethereal 100% Dead Cert
Arasan 77%
LC0 16%
Pedone 6%

Code: Select all

Engine,   Tournament Elo, Initial Elo, First, First two
Ethereal       3314         3300       1.000  1.000
Arasan         3155         3124       0.000  0.771
lc0            3130         3100       0.000  0.160
Pedone         3117         3092       0.000  0.067
DeusX          3120         3130       0.000  0.001
Hannibal       3119         3158       0.000  0.000
Bobcat         2982         3029       0.000  0.000
Nemorino       3092         3099       0.000  0.000

Code: Select all

 1 Ethereal 10.81    3176 19.0 24 197.00 0 +201 79.2 ···· 1=1  1=== ===  1==1 111  11=1 111 
 2 Arasan TCEC13     3142 14.5 24 150.50 0  +93 60.4 0=0  ···· ==1  ==1  01=1 1=1= ===1 1== 
 3 lc0 16.10520      3219 13.5 24 141.25 0  -34 56.3 0=== ==0  ···· ==1  ==1  =01  ===1 1==1
 4 Pedone 1.8        3104 12.5 23 126.25 0  +81 54.3 ===  ==0  ==0  ···· 100  0=== 111  =111
 5 DeusX 1.0         3200 10.5 23 115.00 0  -88 45.7 0==0 10=0 ==0  011  ···· =10  ==0  1=1 
 6 Nemorino 5.01     3104 10.0 23 105.00 2   +5 43.5 000  0=0= =10  1=== =01  ···· 1=0  011 
 7 Hannibal 20180806 3193  9.5 24  98.75 0 -127 39.6 00=0 ===0 ===0 000  ==1  0=1  ···· 11= 
 8 Bobcat 8          3072  4.5 23  54.25 0 -132 19.6 000  0==  0==0 =000 0=0  100  00=  ····

[chrisw]

If LC0 wins against Arasan and they end with same pts, the tie break works as:
Direct Encounter: that would be equal
Win count: probably Arasan wins (at the moment Arasan has 8 wins, and LC0 6, add 1 for beating Arasan = 7, not enough)

If LC0 draws or loses to Arasan but still manages to get to tie break, Arasan wins on Direct Encounter score.

So, if I got that right, LC0 will always lose tie break to Arasan, so therefore needs 0.5 pts more to promote. That's a pretty tall order. Beat Arasan AND get more pts overall.

Very likely, Ethereal and Arasan through to Div 2

[chrisw]

LC0 wins against Nemorino, sim promotion chances:

Ethereal 100%
Arasan 72%
LC0 25%

[JJJ]

LC0 wins against Nemorino, sim promotion chances:

Ethereal 100%
Arasan 72%
LC0 25%

And he did !

[CMCanavessi]

So after the "non-optimal" fix they applied in the 2nd TCEC server (the 2xgpu one) leela is 5.5/6, where's all the people that were saying that the level of play that leela was showing at the beginning was real and the ones that knew that something was wrong were just "fanboys"? I understand variance of outcome, but this is too extreme. From being humiliated and not winning a single game for what, 20 games (?) to getting 5.5/6 is a long way...

There are rumors now that the previous NN games that happened before the ugly fix will be replayed (Leela and DeusX). That's gonna create another flamewar in itself

[chrisw]

ignoring the next two games, if ...

Arasan vs Ethereal 1-0 then chance promotion Arasan 90% LC0 9%
Arasan vs Ethereal Draw then Arasan 78% LC0 20%
Arasan vs Ethereal 0-1 then Arasan 63% LC0 33%

[CMCanavessi]

And what if

Arasan 0-1 Ethereal
and
lc0 1-0 Arasan

That's the interesting case. Also considering Arasan 0.5-0.5 Ethereal.

[Uri Blass]

So after the "non-optimal" fix they applied in the 2nd TCEC server (the 2xgpu one) leela is 5.5/6, where's all the people that were saying that the level of play that leela was showing at the beginning was real and the ones that knew that something was wrong were just "fanboys"? I understand variance of outcome, but this is too extreme. From being humiliated and not winning a single game for what, 20 games (?) to getting 5.5/6 is a long way...

There are rumors now that the previous NN games that happened before the ugly fix will be replayed (Leela and DeusX). That's gonna create another flamewar in itself

Fixing the hardware setting for lc0 in division 3 is not fair when tcec did not fix the fact that ivanhoe used only one core in division 4 because of rule that they fix nothing unless the engine crash(even if it is not the exe file but some setting).

Replaying previous NN games is not fair unless you go back to division 4 and replay all Ivanhoe games and if Ivanhoe or Wasp promote instead of Deus as a result also give Ivanhoe or Wasp a chance to compete in division 3 instead of Deus.