FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

fastgm · Post by **fastgm** » Thu Mar 12, 2020 1:12 pm

New: 2. Fat Fritz 1.1
52 Elo better than Fat Fritz 1.0

   # Engine                            :    Elo  Error   Games    Points    (%)      W      D      L     D(%)   CFS
 -------------------------------------------------------------------------------------------------------------------
   1 Lc0k 0.23.2 256x20-T40-1541       :   3446     10    2750    2067.0   75.16   1503   1128    119   41.02    95
   2 Fat Fritz 1.1                     :   3433     10    2500    1902.0   76.08   1409    986    105   39.44    96
   3 Stockfish 11 C0                   :   3422      8    2750    1447.0   52.62    432   2030    288   73.82    60
   4 Lc0 0.22.0 T40B.4-160             :   3420      9    3000    2138.5   71.28   1455   1367    178   45.57    56
   5 Lc0 0.22.0 42850                  :   3419      9    3000    2135.5   71.18   1472   1327    201   44.23    65
   6 Stockfish 11                      :   3417      7    3750    2047.5   54.60    758   2579    413   68.77    99
   7 Lc0k 0.23.2 62172                 :   3402     10    2500    1827.0   73.08   1256   1142    102   45.68    51
   8 Lc0 LS-LS12.2                     :   3402     10    2750    1944.5   70.71   1299   1291    160   46.95    63
   9 Lc0 0.22.0 J13B.2-188             :   3399      9    3000    2072.5   69.08   1351   1443    206   48.10    78
  10 Lc0 LS-LS12.1                     :   3394      9    3000    2056.0   68.53   1299   1514    187   50.47    61
  11 Lc0k 0.23.2 384x30-T40-2036       :   3392     10    2750    1917.5   69.73   1312   1211    227   44.04    95
  12 Fat Fritz 1.0                     :   3381      9    3000    2012.0   67.07   1311   1402    287   46.73    88
  13 Lc0 0.22.0 32930                  :   3373      9    3000    1986.0   66.20   1249   1474    277   49.13    53
  14 Lc0 0.22.0 LS12 preview           :   3373      9    2750    1876.0   68.22   1258   1236    256   44.95    60
  15 Stockfish 10                      :   3371      8    2750    1337.5   48.64    428   1819    503   66.15    68
  16 Lc0 0.23.2 61950                  :   3368      9    3000    1969.0   65.63   1224   1490    286   49.67    58
  17 Lc0 0.23.2 58613+                 :   3367     10    2500    1750.0   70.00   1200   1100    200   44.00    72
  18 Lc0 0.22.0 384x30-T40-1573        :   3363      9    2750    1846.0   67.13   1220   1252    278   45.53   100
  19 Lc0 0.22.0 11260                  :   3335      9    2750    1755.5   63.84   1044   1423    283   51.75    97
  20 Houdini 6.03                      :   3324      6    4250    1772.5   41.71    486   2573   1191   60.54   100
  21 Komodo 13.2.5                     :   3307      7    4250    1678.5   39.49    495   2367   1388   55.69   100
  22 Scorpio 3.07 NN-ens-net-20x256    :   3276      9    2500    1466.5   58.66    855   1223    422   48.92   100
  23 Ethereal 11.75                    :   3206      7    4250    1155.0   27.18    207   1896   2147   44.61    66
  24 Xiphos 0.6                        :   3203      7    4250    1144.5   26.93    195   1899   2156   44.68    74
  25 Fire 7.1                          :   3200      8    4250    1128.5   26.55    240   1777   2233   41.81   100
  26 Laser 1.7                         :   3130      8    4250     839.0   19.74    156   1366   2728   32.14    98
  27 Andscacs 0.95                     :   3118      9    4250     792.0   18.64    156   1272   2822   29.93   100
  28 Deep Shredder 13                  :   3100      9    4250     730.0   17.18    109   1242   2899   29.22    85
  29 Booot 6.3.1                       :   3093      9    4250     706.5   16.62    112   1189   2949   27.98   ---

White advantage = 60.11 +/- 1.19
Draw rate (equal opponents) = 64.39 % +/- 0.31

Games        : 47.500 (finished)

White Wins   : 15.885 (33.4 %)
Black Wins   :  9.606 (20.2 %)
Draws        : 22.009 (46.3 %)

White Score  : 56.6 %
Black Score  : 43.4 %

http://www.fastgm.de

Albert Silver · Post by **Albert Silver** » Thu Mar 12, 2020 4:54 pm

fastgm wrote: ↑Thu Mar 12, 2020 1:12 pm New: 2. Fat Fritz 1.1
52 Elo better than Fat Fritz 1.0

Code: Select all

   # Engine                            :    Elo  Error   Games    Points    (%)      W      D      L     D(%)   CFS
 -------------------------------------------------------------------------------------------------------------------
   1 Lc0k 0.23.2 256x20-T40-1541       :   3446     10    2750    2067.0   75.16   1503   1128    119   41.02    95
   2 Fat Fritz 1.1                     :   3433     10    2500    1902.0   76.08   1409    986    105   39.44    96
   3 Stockfish 11 C0                   :   3422      8    2750    1447.0   52.62    432   2030    288   73.82    60
   4 Lc0 0.22.0 T40B.4-160             :   3420      9    3000    2138.5   71.28   1455   1367    178   45.57    56
   5 Lc0 0.22.0 42850                  :   3419      9    3000    2135.5   71.18   1472   1327    201   44.23    65
   6 Stockfish 11                      :   3417      7    3750    2047.5   54.60    758   2579    413   68.77    99
   7 Lc0k 0.23.2 62172                 :   3402     10    2500    1827.0   73.08   1256   1142    102   45.68    51
   8 Lc0 LS-LS12.2                     :   3402     10    2750    1944.5   70.71   1299   1291    160   46.95    63
   9 Lc0 0.22.0 J13B.2-188             :   3399      9    3000    2072.5   69.08   1351   1443    206   48.10    78
  10 Lc0 LS-LS12.1                     :   3394      9    3000    2056.0   68.53   1299   1514    187   50.47    61
  11 Lc0k 0.23.2 384x30-T40-2036       :   3392     10    2750    1917.5   69.73   1312   1211    227   44.04    95
  12 Fat Fritz 1.0                     :   3381      9    3000    2012.0   67.07   1311   1402    287   46.73    88
  13 Lc0 0.22.0 32930                  :   3373      9    3000    1986.0   66.20   1249   1474    277   49.13    53
  14 Lc0 0.22.0 LS12 preview           :   3373      9    2750    1876.0   68.22   1258   1236    256   44.95    60
  15 Stockfish 10                      :   3371      8    2750    1337.5   48.64    428   1819    503   66.15    68
  16 Lc0 0.23.2 61950                  :   3368      9    3000    1969.0   65.63   1224   1490    286   49.67    58
  17 Lc0 0.23.2 58613+                 :   3367     10    2500    1750.0   70.00   1200   1100    200   44.00    72
  18 Lc0 0.22.0 384x30-T40-1573        :   3363      9    2750    1846.0   67.13   1220   1252    278   45.53   100
  19 Lc0 0.22.0 11260                  :   3335      9    2750    1755.5   63.84   1044   1423    283   51.75    97
  20 Houdini 6.03                      :   3324      6    4250    1772.5   41.71    486   2573   1191   60.54   100
  21 Komodo 13.2.5                     :   3307      7    4250    1678.5   39.49    495   2367   1388   55.69   100
  22 Scorpio 3.07 NN-ens-net-20x256    :   3276      9    2500    1466.5   58.66    855   1223    422   48.92   100
  23 Ethereal 11.75                    :   3206      7    4250    1155.0   27.18    207   1896   2147   44.61    66
  24 Xiphos 0.6                        :   3203      7    4250    1144.5   26.93    195   1899   2156   44.68    74
  25 Fire 7.1                          :   3200      8    4250    1128.5   26.55    240   1777   2233   41.81   100
  26 Laser 1.7                         :   3130      8    4250     839.0   19.74    156   1366   2728   32.14    98
  27 Andscacs 0.95                     :   3118      9    4250     792.0   18.64    156   1272   2822   29.93   100
  28 Deep Shredder 13                  :   3100      9    4250     730.0   17.18    109   1242   2899   29.22    85
  29 Booot 6.3.1                       :   3093      9    4250     706.5   16.62    112   1189   2949   27.98   ---

White advantage = 60.11 +/- 1.19
Draw rate (equal opponents) = 64.39 % +/- 0.31

Games        : 47.500 (finished)

White Wins   : 15.885 (33.4 %)
Black Wins   :  9.606 (20.2 %)
Draws        : 22.009 (46.3 %)

White Score  : 56.6 %
Black Score  : 43.4 %

http://www.fastgm.de

Many thanks, and quite interesting. I suppose I could train longer to eke out the extra Elo still in it. In training terms it still has Elo under the belt to go for, which you'll just have to trust me on. But it was a deliberate limit I set due to time. If i want to develop a new and stronger net I cannot proceed that way, so this was it pretty much. Regardless, I am happy with its performance, and was happy to notice that aside from the absolute lowest-rated entries, it had the lowest draw rate by far, attesting to its aggressive pro-active play.

Ozymandias · Post by **Ozymandias** » Thu Mar 12, 2020 8:03 pm

Any plans on testing LS 14?

fastgm · Post by **fastgm** » Thu Mar 12, 2020 8:40 pm

Ozymandias wrote: ↑Thu Mar 12, 2020 8:03 pm Any plans on testing LS 14?

The test is already running and will be finished in about 5 days.

jjoshua2 · Post by **jjoshua2** » Thu Mar 12, 2020 11:45 pm

Using the kiudee settings Pohl used?
WCEC is using these similar settings:

Code: Select all

"Threads": 3,
"MinibatchSize": 512,
"Cpuct": 1.9000
"NNCacheSize": 20000000,
"maxCollisionEvents": 32
"CPuctBase": 15000,
"CPuctFactor": 2.817,
"TradePenalty": 0.0
"TradePenalty2": 0.0
"TempDecayMoves": 0,
"TempCutoffMove": 2,
"TempEndgame": 0.000000,
"TempValueCutoff": 0.420000,
"TempVisitOffset": 0.000000,
"DirichletNoise": false,
"VerboseMoveStats": false,
"FpuValueAtRoot": 1.000000,
"CacheHistoryLength": 0,
"OutOfOrderEval": true,
"KLDGainAverageInterval": 100,
"MinimumKLDGainPerNode": 0.000000,
"CertaintyPropagation": true,
"TwoFoldDrawScoring": true,
"Ponder": false,
"Backend":" roundrobin",
"BackendOptions":" (backend=cudnn-fp16,gpu=0),(backend=cudnn-fp16,gpu=1)",
"MaxPrefetch": 32,
"Temperature": 1,
"TempVisitOffset": 0.00,
"smartpruningfactor": 1.4,
"FpuStrategy": "reduction",
"FpuValue": 0.443,
"PolicyTemperature": 1.607,
"MaxCollisionVisits": 9999,
"MultiPV": 1

Noticeable difference is cpuct at 1.9 and cpuctbase at 15k and smartpruningfactor 1.4. The cpuct 1.9 change was chosen because of a fixed 800 node test. Up 16 elo after 1000 games compared to kiudee default of ~2.1.
CCC is using almost the same settings but using trade penalty in weaker divisions and 4 GPUs.

fastgm · Post by **fastgm** » Fri Mar 13, 2020 7:26 am

jjoshua2 wrote: ↑Thu Mar 12, 2020 11:45 pm Using the kiudee settings Pohl used?

Yes, I'm using Lc0 v0.24.0

option name CPuct type string default 2.147000
option name CPuctBase type string default 18368.000000
option name CPuctFactor type string default 2.815000
option name FpuValue type string default 0.443000
option name PolicyTemperature type string default 1.607000

Ozymandias · Post by **Ozymandias** » Fri Mar 13, 2020 7:35 am

What's WCEC?

jjoshua2 · Post by **jjoshua2** » Fri Mar 13, 2020 2:33 pm

Leelenstein is defending its title, and in the lead at WCEC
Its 2 RTX gpus with a strong I9. Leela ratio is a little worse than last TCEC season. If he disabled a few threads from SF it matched TCEC's ratio, but this tournament is at full bore. You can see a lot more information below the stream.

Ozymandias · Post by **Ozymandias** » Fri Mar 13, 2020 8:53 pm

If anything, SF needs more threads to keep up with the Leelas.

The 1000 games settings test you mention... can you post the final crosstable?

carldaman · Post by **carldaman** » Sat Mar 14, 2020 7:36 am

I see a couple of Trade Penalty settings, which are new, apparently.

Where can we find more about them? I wouldn't want to mess with these unless I knew what impact to expect.

FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1

Re: FGRL - 16 Cores rating list - 2. Fat Fritz 1.1