32930-112x9-se and ender-112x9-se available

[dkappe]

At the usual place: https://github.com/dkappe/leela-chess-w ... d-Networks

[Eduard]

Thank You!

In my opinion, the current NN 50775 is tactical stronger than the distilled 32930-112x9-se. I tested in Analysis mode on 2x2,4 GHz CPU the bad tactical Lc0 positions know to me. Here is an example:

Rybka 4.1 x64 - Lc0 v0.20.1 1-0, ASUS-PC, Schnellschach 20m+5s 2019
[d]8/6p1/4Qpk1/p4n1p/2P2P2/1r1R2PP/2q1N1K1/8 w - - 0 1

Analysis by Lc0 v0.21.0 (Network 50775):

54.Txb3 Dxb3 55.De8+ Kh7 56.Dxh5+ Sh6 57.Dxa5 Dxc4 58.Kf2 De4 59.Dc3 Dh1 60.g4 Dh2+
+- (1.74) Tiefe: 10/17 00:00:49 5kN
54.Txb3 Dxb3 55.Dxf5+ Kxf5 56.Sd4+ Ke4 57.Sxb3 a4 58.Sc5+ Kd4 59.Sxa4 Kxc4 60.Sb6+ Kc5 61.Sc8 g6 62.Se7
+- (3.02) Tiefe: 10/17 00:00:54 7kN

Analysis by Lc0 v0.21.0se (Network 32930-112x9-se):

54.De8+ Kh7 55.Txb3 Dxb3 56.Dxh5+ Sh6 57.Dxa5 Dxc4 58.Kf2 De6 59.g4 De4 60.Dd2 Sg8 61.De3 Dd5
= (0.14) Tiefe: 11/18 00:02:03 17kN
54.Txb3 Dxb3 55.Dxf5+ Kxf5 56.Sd4+ Ke4 57.Sxb3 a4 58.Sc5+ Kd4 59.Sxa4 Kxc4 60.Sb2+ Kc3 61.Sd1+ Kd2 62.Sf2
+- (8.02) Tiefe: 11/18 00:02:06 18kN

[dkappe]

All of my testing at 2+2 on cpu indicates that 11258-112x9-se is better than 32930-112x9-se, t35 and t50, at least at low-ish nodes. Go figure. Maybe 50782 has a shot.

[Graham Banks]

At the usual place: https://github.com/dkappe/leela-chess-w ... d-Networks

Thanks.

[Guenther]

Thank You!

In my opinion, the current NN 50775 is tactical stronger than the distilled 32930-112x9-se. I tested in Analysis mode on 2x2,4 GHz CPU the bad tactical Lc0 positions know to me. Here is an example:

...

Still very early days, but nevertheless, here my current tests of small and distilled networks for being best on slow gpu:

Code: Select all

#       PLAYER                  RATING  ERROR   POINTS  PLAYED  (%)     CCRL 40/4   Diff      W       D       L
1       LC0_0211-ID50782        3300.31 142.06  12.5    20      62.50   *           *         8       9       3
2       Laser_17-64             3255.26 72.43   45.0    70      64.29   3266        -10.74   32      26      12
3       Xiphos_04-64            3234.31 105.39  21.0    32      65.63   3226          8.31   16      10       6
4       Booot_631-64            3201.02 89.37   23.5    40      58.75   3251        -49.98   15      17       8
5       Nemorino_500-64         3160.77 81.14   23.5    50      47.00   3119         41.77   14      19      17
6       Fizbo_20-64             3160.54 82.58   25.5    48      53.13   3246        -85.46   16      19      13
7       LC0_0202-ID11258#112x9  3154.20 28.01   251.0   460     54.57   *           *       174     154     132
8       ArasanX_212-64          3146.97 91.36   20.5    40      51.25   3102         44.97   14      13      13
9       Hannibal_17-64          3140.87 75.56   32.5    60      54.17   3110         30.87   22      21      17
10      Senpai_20-64            3138.07 89.08   20.0    40      50.00   3083         55.07   13      14      13
11      Chess22k_112-64         3131.79 73.45   26.0    52      50.00   3082         49.79   17      18      17
12      LC0_0210-ID11258#120x9  3121.94 63.03   54.0    110     49.09   *           *        35      38      37
13      Rofchade_20-64          3118.55 86.48   18.0    40      45.00   3127        -8.45    11      14      15
14      Ethereal_1000-64        3083.54 98.23   14.5    32      45.31   3143        -59.46    8      13      11
15      Pedone_18-64            3056.58 89.41   15.5    40      38.75   3105        -48.42    9      13      18
16      LC0_0210-ID50498        3056.05 66.21   31.5    89      35.39   *           *        16      31      42
17      SmarThink_198-64        3045.43 84.21   18.0    50      36.00   3038          7.43   10      16      24
18      Wasp_350-64             3036.06 88.04   16.0    45      35.56   3065        -28.94   10      12      23
19      Demolito_181029-64      2970.87 97.36   10.5    40      26.25   3021        -50.13    7       7      26
Gauntlet Opp Rating             3132.27                 679             3132.27      -6.89   draw rate 34.17%

Infos and lots of stats here: (will add some benchmarks for this slow gpu with the selected networks soon)
https://docs.google.com/spreadsheets/d/ ... XCH9joVodI

I am already surprised that I can reach those ratings on such a slow gpu.

[Eduard]

Thanks! Question: what is the goal of the 51xxx networks?

[dkappe]

Thanks! Question: what is the goal of the 51xxx networks?

To quote from the #dev-log channel on discord:

**Test51**: Compared to T50... fpu settings have changed (fpu reduction 0.5, 0 at root, instead of -1 absolute). No longer using gamma regularization - in favor of forcing some gammas to 1.0, although maybe I'll still need to add some in later. Also using renorm from the start - currently in 'enabled but set to 1,0 which means it is basically batch norm' mode. Will relax that restriction over the next couple of days.

- fpu is a technique for goosing the mcts/puct when you have unexplored nodes by giving them a certain outcome by default (win = 1.0, loss = -1.0, etc.). Lots of fiddling to help with tactics, maybe.
- put very simplistically, regularization and normalization are ways of keeping things like inputs, outputs, weights and so on, small and in a certain range. They are different concepts used differently and for different purposes during training. I do use NN’s professionally, but I haven’t really followed the details of the t40 and t50 issues. You’ll have to ask someone on discord about that.

[dkappe]

Well, the 50782 net started out great guns on CPU, but at 2+2 it’s a little behind. Will keep running.

Score of 50782 vs 11258-112x9-se: 26 - 36 - 84 [0.466] 146

[Guenther]

Well, the 50782 net started out great guns on CPU, but at 2+2 it’s a little behind. Will keep running.

Score of 50782 vs 11258-112x9-se: 26 - 36 - 84 [0.466] 146

Did you play more or other test games with ID50782 later?

My current results look like this:

Code: Select all

   # PLAYER                    :   RATING  ERROR  POINTS  PLAYED    (%)    W    D    L
   1 Xiphos_04-64              :  3268.45  80.35    36.0      52  69.23   28   16    8
   2 Laser_17-64               :  3251.06  70.78    45.0      70  64.29   32   26   12
   3 LC0_0211-ID50782          :  3215.48  51.77    80.5     134  60.07   55   51   28
   4 Booot_631-64              :  3197.83  88.10    23.5      40  58.75   15   17    8
   5 Fizbo_20-64               :  3181.93  66.57    35.5      68  52.21   22   27   19
   6 Hannibal_17-64            :  3179.27  62.35    44.0      80  55.00   30   28   22
   7 LC0_0202-ID11258#112x9    :  3157.06  26.82   251.0     460  54.57  174  154  132
   8 Nemorino_500-64           :  3147.30  77.22    23.5      50  47.00   14   19   17
   9 ArasanX_212-64            :  3137.72  70.74    28.0      60  46.67   17   22   21
  10 Chess22k_112-64           :  3133.87  71.94    30.0      60  50.00   19   22   19
  11 Ethereal_1000-64          :  3121.68  79.24    23.0      52  44.23   14   18   20
  12 Rofchade_20-64            :  3121.48  89.94    18.0      40  45.00   11   14   15
  13 LC0_0210-ID11258#120x9    :  3112.74  54.04    59.0     130  45.38   37   44   49
  14 LC0_0210-ID50498          :  3087.31  54.67    45.0     112  40.18   25   40   47
  15 Senpai_20-64              :  3082.60  74.29    23.5      60  39.17   13   21   26
  16 Pedone_18-64              :  3067.59  76.21    20.5      54  37.96   11   19   24
  17 SmarThink_198-64          :  3045.98  80.44    18.0      50  36.00   10   16   24
  18 Wasp_350-64               :  3029.73  73.69    21.5      60  35.83   13   17   30
  19 Demolito_181029-64        :  2974.07  97.58    10.5      40  26.25    7    7   26

White advantage = 38.01 +/- 9.94
Draw rate (equal opponents) = 36.61 % +/- 1.72

https://docs.google.com/spreadsheets/d/ ... XCH9joVodI

[dkappe]

Code: Select all

   # PLAYER            :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 11258-112x9-se    :       0     15   132.5     250  53.0      92   62  141   47  56.4
   2 50782             :     -21     15   117.5     250  47.0     ---   47  141   62  56.4

White advantage = 40.99 +/- 14.96
Draw rate (equal opponents) = 57.55 % +/- 2.99

2+2

At low nodes, there’s an inflection point that 112x9 seems to reach with reasonable smarts sooner than t35 and t50. I suspect at longer tc the 128x10-se nets might do better on cpu or slow gpu.