FGRL - rating lists June 20th 2020

[fastgm]

Updates since June 14th:

16 cores
Lc0 0.25.1 703042 (New: 3406)

60 min + 15 sec
Winter 0.8 (New: 3013)
chess22k 1.14 (+6 to chess22k 1.13)

10 min + 6 sec
RubiChess 1.7.3 (-7 to RubiChess 1.7.2)
Wasp 4.00 (+38 to Wasp 3.75)

60 sec + 0.6 sec
Igel 2.5.0 (+17 to Igel 2.4.0)
Weiss 1.0 (+55 to Weiss 0.1)
Ethereal 12.25 (-7 to Ethereal 12.00)
Bagatur 2.2 (-14 to Bagatur 2.0)
SlowChess Blitz Classic 2.2 (+94 to SlowChess Blitz Classic 2.1)
Wasp 4.00 (+55 to Wasp 3.75)
RubiChess 1.7.3 (-1 to RubiChess 1.7.2)
GreKo 2020.03 (+6 to GreKo 2020.01)
Monolith 2.01 (New: 2830)
Francesca MAD 0.29 (New: 2699)

[Vinvin]

16 cores

What does this "16 cores" means ?

[AndrewGrant]

Ethereal 12.25 (-7 to Ethereal 12.00)

Seems you have the first rating list with enough weak opponents that the contempt from V12.00 that is no longer default in V12.25 is a big deal.

Should be far better results in your more exclusive lists. If not, something is wrong. Are you using Linux, or the Windows binaries I provided? I'm compiling Windows binaries from a new machine, so that could be a factor.

[fastgm]

16 cores

What does this "16 cores" means ?

16 cores is the name of the rating list.
This means the NN-engines (NVidia GeForce RTX 2070) are playing against an a/b engines on 16 cores (Dual Intel Xeon E5-2670v3).

Code: Select all

Playing conditions

CPU:                 Dual Intel Xeon E5-2670v3 @ 2.6 GHz, 24 Cores
GPU:                 NVidia GeForce RTX 2070
OS:                  Windows 10 64-Bit
Tool:                Cutechess-Cli
Leela Ratio:         ~ 1.0
A/B-Engines:         16 Cores, 64 Bit PEXT/BMI2, default settings
NN-Engines:          default settings
Hash-Table:          A/B-Engines 512 MB
NN-Cache:            default settings
Lc0-backend:         cudnn-fp16
Time control:        60 seconds + 0.6 seconds
Tablebases:          No, but tablebase adjudication with 6 pieces
Openings:            Hert_250_lowdraws.epd, changing colors
Ponder:              Off
Learning:            Off
Large Memory Pages:  Off

[fastgm]

Ethereal 12.25 (-7 to Ethereal 12.00)

Seems you have the first rating list with enough weak opponents that the contempt from V12.00 that is no longer default in V12.25 is a big deal.

Should be far better results in your more exclusive lists. If not, something is wrong. Are you using Linux, or the Windows binaries I provided? I'm compiling Windows binaries from a new machine, so that could be a factor.

Hi Andrew,

I'm using the windows binaries.

I am really looking forward to the results of the 10 and 60 minute rating lists.

Andreas

[Raphexon]

Ethereal 12.25 (-7 to Ethereal 12.00)

Seems you have the first rating list with enough weak opponents that the contempt from V12.00 that is no longer default in V12.25 is a big deal.

Should be far better results in your more exclusive lists. If not, something is wrong. Are you using Linux, or the Windows binaries I provided? I'm compiling Windows binaries from a new machine, so that could be a factor.

I think there's something wrong with the elo calculation:
See the head 2 head results.
Ethereal 12.25 scores better than 12.00, but is listed as having less elo.

http://www.fastgm.de/h2h60.html

H6 vs Eth:

Code: Select all

    
     Ethereal 12.00                    :    250 (   108,   120,   22),  67.2 :   +125,    3,  100.0
     Ethereal 12.25                    :    250 (    89,   139,   22),  63.4 :   +132,    4,  100.0

SF11 vs Eth:

Code: Select all

 Ethereal 12.00                    :    250 (   164,   82,   4),  82.0 :   +259,    4,  100.0
     Ethereal 12.25                    :    250 (   147,  101,   2),  79.0 :   +267,    5,  100.0

Kmcts14 vs Eth:

Code: Select all

     Ethereal 12.00                    :    250 (    64,  129,   57),  51.4 :     +9,    4,   98.3
     Ethereal 12.25                    :    250 (    55,  143,   52),  50.6 :    +16,    4,  100.0

Xiphos vs Eth:
Xiphos is listed as 27 elo weaker than Eth 12.00, scores much worse vs Eth 12.25 but is somehow only 19 elo weaker.

Code: Select all

     Ethereal 12.00                    :    250 (    56,  132,   62),  48.8 :    -27,    3,    0.0
     Ethereal 12.25                    :    250 (    28,  151,   71),  41.4 :    -19,    4,    0.0

@fastgm, eth12.25 scores better across the board, yet the rating list has Eth12.25 below 12.00 even with only 1 common opponent.

[AndrewGrant]

I think there's something wrong with the elo calculation:

Unless I'm not understanding something, I think you are right.

[fastgm]

Ethereal 12.25 (-7 to Ethereal 12.00)

Seems you have the first rating list with enough weak opponents that the contempt from V12.00 that is no longer default in V12.25 is a big deal.

Should be far better results in your more exclusive lists. If not, something is wrong. Are you using Linux, or the Windows binaries I provided? I'm compiling Windows binaries from a new machine, so that could be a factor.

I think there's something wrong with the elo calculation:
See the head 2 head results.
Ethereal 12.25 scores better than 12.00, but is listed as having less elo.

http://www.fastgm.de/h2h60.html

H6 vs Eth:

Code: Select all

    
     Ethereal 12.00                    :    250 (   108,   120,   22),  67.2 :   +125,    3,  100.0
     Ethereal 12.25                    :    250 (    89,   139,   22),  63.4 :   +132,    4,  100.0

SF11 vs Eth:

Code: Select all

 Ethereal 12.00                    :    250 (   164,   82,   4),  82.0 :   +259,    4,  100.0
     Ethereal 12.25                    :    250 (   147,  101,   2),  79.0 :   +267,    5,  100.0

Kmcts14 vs Eth:

Code: Select all

     Ethereal 12.00                    :    250 (    64,  129,   57),  51.4 :     +9,    4,   98.3
     Ethereal 12.25                    :    250 (    55,  143,   52),  50.6 :    +16,    4,  100.0

Xiphos vs Eth:
Xiphos is listed as 27 elo weaker than Eth 12.00, scores much worse vs Eth 12.25 but is somehow only 19 elo weaker.

Code: Select all

     Ethereal 12.00                    :    250 (    56,  132,   62),  48.8 :    -27,    3,    0.0
     Ethereal 12.25                    :    250 (    28,  151,   71),  41.4 :    -19,    4,    0.0

@fastgm, eth12.25 scores better across the board, yet the rating list has Eth12.25 below 12.00 even with only 1 common opponent.

No, that's not correct. See the results against the weaker opponents.
Otherwise the calculation with Ordo 1.2.6 would be wrong and I do not assume that.

[AndrewGrant]

No, that's not correct. See the results against the weaker opponents.

I suppose the + values next to each H2H are NOT the H2H elo difference for the matchup then

[jonkr]

Pretty happy to get a result so high in the ranking list, usually I expect rating list results won't be as good as in my testing, and this looks about the same, slightly better even. So my goal of passing Xiphos 0.6 in at least one list is met.

My understanding of the H2H list is the elo difference values are from the engine elos on the rating list, not the specific H2H matchup. For Ethereal it looks pretty strongly like a case of better results against stronger engines, but lower/more drawish against the large amount of weaker engines.