TCEC Division 3 results simulator

[chrisw]

it takes my version of LC0 (about three weeks old weights, I think)
33 iterations to find c6+
after c6+ Ka7 then LC0 takes 18 iterations to find Rd3
So, this blunder oversight is quite easy to test for the mirror effect, just count iterations ...

What do you mean by "iterations"?

good question. I meant the second number in the main line, which increments up by one each time. But I think that's not actually an iteration counter (mapping from AB), but a main line depth. I was random babbling, probably. Trying to find a measure of how long LC0 took to find lines.

[AndrewGrant]

Leela did not dominate in div 4, check the score. Div 3 partial results are just a bit worse than expected. Div 4 had 2-3 engines not even running properly, and were mostly weaker even in ideal conditions. TCEC conditions are to blame, as I stated in another thread. GPUs are underpowered price-wise.

GPUs under-powered price-wise is a load of bs. It is well known that the TCEC CPUs are extremely overpriced for their value.

https://www.phoronix.com/scan.php?page= ... 90wx&num=5

As I predicted very recently, a single Ryzen2990WX has even performance with TCEC CPUs -- at only a 5th of the cost.

And, taking the 1080TIs as $900, perfectly equal in price. (I know the price has dropped, but at worst, $1450 vs $1900)

Leela and NN fans alike need to fix their problems -- instead of asking for more and more and more hardware to make up for their shortcomings.

[Laskos]

Leela did not dominate in div 4, check the score. Div 3 partial results are just a bit worse than expected. Div 4 had 2-3 engines not even running properly, and were mostly weaker even in ideal conditions. TCEC conditions are to blame, as I stated in another thread. GPUs are underpowered price-wise.

GPUs under-powered price-wise is a load of bs. It is well known that the TCEC CPUs are extremely overpriced for their value.

https://www.phoronix.com/scan.php?page= ... 90wx&num=5

As I predicted very recently, a single Ryzen2990WX has even performance with TCEC CPUs -- at only a 5th of the cost.

And, taking the 1080TIs as $900, perfectly equal in price. (I know the price has dropped, but at worst, $1450 vs $1900)

Leela and NN fans alike need to fix their problems -- instead of asking for more and more and more hardware to make up for their shortcomings.

Then wait for the next Nvidia line, it will be out soon.
But I agree that something odd happens with loads of hardware and time control.
OTOH, in blitz on home reasonable, balanced CPU/GPU hardware, Lc0 with the latest nets on GPU are pretty much the level of Andscacs or above on full CPU. And the progress in strength is fairly steady. This progress shouldn't be neglected, as the approach is completely different, completely untuned and malleable to huge improvements. One doesn't need to be a fanboy to feel a bit excited about the developments, although I had my disappointments (like the extreme specialization of DCNN approach).

[chrisw]

Nemerino, not to be underestimated, defeated ZeusX in final game of Round 19, Sim promotion predictions:

Ethereal 100%
Pedone 43%
DeusX 28%
Arasan 26%
LC0 1%

LC0 draw rate about 74%, 5 results in 19 games

Code: Select all

 1 Ethereal 10.81    3176 14.5 19 126.50 0 +141 76.3 ··· === 1== 1=1 1== 11  11= 11 
 2 Pedone 1.8        3104 11.0 19  93.25 0  +92 57.9 === ··· 100 ==  ==  0== 111 =11
 3 DeusX 1.0         3200 10.5 19  92.25 0  -21 55.3 0== 011 ··· 10= ==  =10 ==  1=1
 4 Arasan TCEC13     3142 10.5 19  91.75 0  +47 55.3 0=0 ==  01= ··· ==1 1=1 === 1= 
 5 lc0 16.10520      3219  9.0 19  79.25 0  -82 47.4 0== ==  ==  ==0 ··· =01 === 1==
 6 Nemorino 5.01     3104  9.0 19  75.75 2  +28 47.4 00  1== =01 0=0 =10 ··· 1=  011
 7 Hannibal 20180806 3193  7.5 19  61.50 0 -105 39.5 00= 000 ==  === === 0=  ··· 11=
 8 Bobcat 8          3072  4.0 19  37.75 0  -99 21.1 00  =00 0=0 0=  0== 100 00= ···

[chrisw]

Couple of losses, including to LC0 and Sim reduces DeusX promotion chances to 14%

[Laskos]

Couple of losses, including to LC0 and Sim reduces DeusX promotion chances to 14%

With these standings after 88/112 games:

Code: Select all

 N Engine            Rtng  Pts Gm     SB X  Elo Perf Et   Ar   Pe   Lc   De   Ha   Ne   Bo  

 1 Ethereal 10.81    3176 17.5 22 171.00 0 +182 79.5 ···· 1=1  ===  1==  1==1 11=  111  111 
 2 Arasan TCEC13     3142 12.5 22 123.75 0  +56 56.8 0=0  ···· ==1  ==1  01=  ===  1=1= 1== 
 3 Pedone 1.8        3104 12.0 22 117.00 0  +81 54.5 ===  ==0  ···· ==0  100  111  0==  =111
 4 lc0 16.10520      3219 11.0 22 114.00 1  -74 50.0 0==  ==0  ==1  ···· ==1  ===0 =01  1== 
 5 DeusX 1.0         3200 10.5 22 107.25 0  -69 47.7 0==0 10=  011  ==0  ···· ==0  =10  1=1 
 6 Hannibal 20180806 3193 10.5 22 101.50 0  -60 47.7 00=  ===  000  ===1 ==1  ···· 0=1  11= 
 7 Nemorino 5.01     3104  9.5 22  93.50 2   +5 43.2 000  0=0= 1==  =10  =01  1=0  ···· 011 
 8 Bobcat 8          3072  4.5 22  49.50 0 -122 20.5 000  0==  =000 0==  0=0  00=  100  ····

Common sense would say that Lc0 still has some 10%, Deus some 7%. Did they have some loses on time/disconnect? I saw something strange, but was not really awake .

[JJJ]

The game leela disconnected might be replayed if it is a server problem. And now Leela is winning bobcat.

[Laskos]

The game leela disconnected might be replayed if it is a server problem. And now Leela is winning bobcat.

Leela - Hannibal, right? If replayed, might even win this one, and in this case would change things seriously.

[chrisw]

Common sense would say that Lc0 still has some 10%, Deus some 7%. Did they have some loses on time/disconnect? I saw something strange, but was not really awake .

Sorry to say, the Sim fairy didn't grant your wish. LC0 now 7% and DX on 5%

Arasan 60% well done John Dart
Pedone 25%
and Ethereal still a dead cert. Ethereal is interesting phenomenon. What is it with these young programmers who arrive seemingly from nowhere and overtake a mass of established programmers who've been at it for years? It's no new phenomenon, it's a long list of illustrious names over the years. Well done Ethereal programmer also.

Code: Select all

Engine,   Tournament Elo, Initial Elo, First, First two
Ethereal       3333         3300       1.000  1.000
Arasan         3169         3124       0.000  0.606
Pedone         3149         3092       0.000  0.251
lc0            3138         3100       0.000  0.076
DeusX          3121         3130       0.000  0.054
Hannibal       3117         3158       0.000  0.012
Nemorino       3083         3099       0.000  0.000
Bobcat         2919         3029       0.000  0.000

Code: Select all

1 Ethereal 10.81    3176 17.5 22 173.00 0 +182 79.5 ···· 1=1  ===  1==  1==1 11=  111  111 
 2 Arasan TCEC13     3142 12.5 22 125.75 0  +56 56.8 0=0  ···· ==1  ==1  01=  ===  1=1= 1== 
 3 Pedone 1.8        3104 12.0 22 118.00 0  +81 54.5 ===  ==0  ···· ==0  100  111  0==  =111
 4 lc0 16.10520      3219 12.0 23 118.50 1  -64 52.2 0==  ==0  ==1  ···· ==1  ===0 =01  1==1
 5 DeusX 1.0         3200 10.5 22 108.25 0  -69 47.7 0==0 10=  011  ==0  ···· ==0  =10  1=1 
 6 Hannibal 20180806 3193 10.5 22 104.00 0  -60 47.7 00=  ===  000  ===1 ==1  ···· 0=1  11= 
 7 Nemorino 5.01     3104  9.5 22  95.00 2   +5 43.2 000  0=0= 1==  =10  =01  1=0  ···· 011 
 8 Bobcat 8          3072  4.5 23  50.50 0 -132 19.6 000  0==  =000 0==0 0=0  00=  100  ····

[chrisw]

The game leela disconnected might be replayed if it is a server problem. And now Leela is winning bobcat.

Leela - Hannibal, right? If replayed, might even win this one, and in this case would change things seriously.

Oh wow, did LC0 disconnect? That would indeed changes things if replayed, would be temporarily in second place in cross table.

Affects the Sim prediction, LC0 jumps to 28% promotion chance

90 games played ...

Code: Select all

Engine,   Tournament Elo, Initial Elo, First, First two
Ethereal       3333         3300       1.000  1.000
Arasan         3169         3124       0.000  0.494
lc0            3166         3100       0.000  0.287
Pedone         3149         3092       0.000  0.185
DeusX          3121         3130       0.000  0.034
Hannibal       3089         3158       0.000  0.000
Nemorino       3083         3099       0.000  0.000
Bobcat         2919         3029       0.000  0.000