LCZero progress

[majortom]

Code: Select all

   # PLAYER                   :  RATING  ERROR  POINTS  PLAYED     %
   1 Stockfish 130418 D30     :  1058.7  125.2   193.0     200    97
   2 Stockfish 130418 D15     :   472.2   70.8  1012.0    1367    74
   3 LC0 173_101eeb41 N32k    :   257.6   87.0    38.5     167    23
   4 LC0 173_101eeb41 N16k    :   223.0   79.0    39.5     200    20
   5 LC0 173_101eeb41 N8k     :   181.0   90.6    32.5     200    16
   6 LC0 173_101eeb41 N4k     :   105.8   86.9    22.5     200    11
   7 LC0 173_101eeb41 N2k     :    39.4   96.0    16.0     200     8
   8 LC0 173_101eeb41 N1k     :     0.0   ----    13.0     200     7

White advantage = 44.71 +/- 10.51
Draw rate (equal opponents) = 50.00 % +/- 0.00

[majortom]

Code: Select all

   # PLAYER                   :  RATING  ERROR  POINTS  PLAYED     %
   1 Stockfish 130418 D15     :   196.7   38.3   150.5     200    75
   2 LC0 173_101eeb41 N32k    :     0.0   ----    49.5     200    25

White advantage = 35.50 +/- 21.68
Draw rate (equal opponents) = 50.00 % +/- 0.00

Games

Code: Select all

   # PLAYER                   :  RATING  ERROR  POINTS  PLAYED     %
   1 Stockfish 130418 D30     :  1057.2  135.4   193.0     200    97
   2 Stockfish 130418 D15     :   471.5   74.6  1034.0    1400    74
   3 LC0 173_101eeb41 N32k    :   274.1   84.0    49.5     200    25
   4 LC0 173_101eeb41 N16k    :   222.8   91.8    39.5     200    20
   5 LC0 173_101eeb41 N8k     :   180.8   90.5    32.5     200    16
   6 LC0 173_101eeb41 N4k     :   105.7   95.5    22.5     200    11
   7 LC0 173_101eeb41 N2k     :    39.4  101.9    16.0     200     8
   8 LC0 173_101eeb41 N1k     :     0.0   ----    13.0     200     7

[majortom]

The new networks have 15 blocks and 192 filters, that reduce nps.
I decided calculate ratio between 192x15 and 128x10 networks nps.
For this ran 1-core test match with ID237 and ID173, then took the info strings.
Extracted 45 111 info strings.

In some strings found too high nps values, which have negative affect the average values:

Code: Select all

LC0 173_101eeb41 depth 18 nodes 14821361 nps 152863 score cp 0 time 53334 pv e8f6

Code: Select all

LC0 173_101eeb41 depth 19 nodes 12764049 nps 139217 score cp 0 time 63584 pv c4b3

And reduced them to 272 for ID173 and to 99 for ID237

Result on the picture.
rc nps is recalculated nps from nodes and time.

To reach 1000, 2000 and 32000 nodes needs 11.67 sec, 65.31 sec and 373.54 sec for ID173 (85.67 average nps for such time).
There are not big sample of the info strings with time, where more than 50 sec and therefore range was reduced.

In this range the average nps for ID237 is 30.62.
85.668 nps / 30.621 nps = 2.798
The average nps ratio in this range is 2.799.

So I decided to use the ratio = 2.8 to calculate number of nodes for 192x15 networks.

Code: Select all

128x10	192x15

 1,000	   357
 2,000	   714
 4,000	 1,429
 8,000	 2,857
16,000	 5,714
32,000	11,429

[majortom]

Between ID173 and ID239 10 days:

It's time for the new tests.

[majortom]

LCZero ID239 (714 nodes per move) vs. ID125 (2000 nodes per move):

Code: Select all

   # PLAYER                  :  RATING  ERROR  POINTS  PLAYED   (%)
   1 LC0 239_0df31cb9 N714   :   306.4  117.2    34.0      40    85
   2 LC0 125_738edc3d N2k    :     0.0   ----     6.0      40    15

White advantage = 34.64 +/- 60.72
Draw rate (equal opponents) = 50.00 % +/- 0.00

N2k is "nodes=2000"

[majortom]

LCZero ID239 (714 nodes per move) vs. ID173 (2000 nodes per move):

Code: Select all

   # PLAYER                  :  RATING  ERROR  POINTS  PLAYED   (%)
   1 LC0 239_0df31cb9 N714   :    84.2   44.1    77.0     125    62
   2 LC0 173_101eeb41 N2k    :     0.0   ----    48.0     125    38

White advantage = 39.42 +/- 23.24
Draw rate (equal opponents) = 50.00 % +/- 0.00

N2k is "nodes=2000"

[majortom]

Code: Select all

   # PLAYER                  :  RATING  ERROR  POINTS  PLAYED   (%)
   1 LC0 239_0df31cb9 N714   :    91.5   35.4   125.5     200    63
   2 LC0 173_101eeb41 N2k    :     0.0   ----    74.5     200    37

White advantage = 9.37 +/- 18.39
Draw rate (equal opponents) = 50.00 % +/- 0.00

Games

Code: Select all

   # PLAYER                   :  RATING  ERROR  POINTS  PLAYED   (%)
   1 Stockfish 130418 D15     :   353.4   55.0   176.5     200    88
   2 LC0 239_0df31cb9 N714    :     0.0   ----    23.5     200    12

White advantage = -4.24 +/- 27.48
Draw rate (equal opponents) = 50.00 % +/- 0.00

Games

[majortom]

Code: Select all

   # PLAYER                   :  RATING  ERROR  POINTS  PLAYED   (%)
   1 Stockfish 130418 D15     :   663.5  133.8   556.0     600    93
   2 LC0 239_0df31cb9 N714    :   308.1  142.8    23.5     200    12
   3 LC0 173_101eeb41 N2k     :   233.2  147.7    16.0     200     8
   4 LC0 125_738edc3d N2k     :     0.0   ----     4.5     200     2

White advantage = 30.72 +/- 17.07
Draw rate (equal opponents) = 50.00 % +/- 0.00

[majortom]

Code: Select all

   # PLAYER                  :  RATING  ERROR  POINTS  PLAYED   (%)
   1 LC0 239_0df31cb9 N714   :   235.2   40.9   158.0     200    79
   2 LC0 125_738edc3d N2k    :     0.0   ----    42.0     200    21

White advantage = 42.70 +/- 21.89
Draw rate (equal opponents) = 50.00 % +/- 0.00

Games

[majortom]

Code: Select all

   # PLAYER                   :  RATING  ERROR  POINTS  PLAYED   (%)
   1 LC0 239_0df31cb9 N714    :   236.6   33.0   283.5     400    71
   2 LC0 173_101eeb41 N2k     :   147.8   31.9   215.0     400    54
   3 LC0 125_738edc3d N2k     :     0.0   ----   101.5     400    25

White advantage = 17.29 +/- 11.07
Draw rate (equal opponents) = 50.00 % +/- 0.00