Berserk 8 Released

[jhonnold]

Downloads: https://github.com/jhonnold/berserk/releases/tag/8

Details
Another release with an improved NNUE evaluation and small search patches (same as last time).

The newest network was trained using a custom trainer on 1.2B FENs from Berserk 6 self play games. Most of the improvements for this release come from tweaks in the trainer and not a largely improve architecture or better data. I was hoping to improve the architecture for this release, but all attempts at this time have failed (miserably I may add).

Minor search patches are included in this release as well.

STC

Code: Select all

ELO   | 68.92 +- 4.47 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=8MB
GAMES | N: 10000 W: 3160 L: 1202 D: 5638

LTC

Code: Select all

ELO   | 47.90 +- 3.68 (95%)
CONF  | 40.0+0.40s Threads=1 Hash=64MB
GAMES | N: 10000 W: 2183 L: 813 D: 7004

Self play results vs Berserk 7, this Elo gain may not reflect exactly when playing against other engines

Choosing a binary
Binaries currently require `popcnt`.
Binaries with the label `avx512` require your processor to support `avx512`.
Binaries with the label `avx2` require your processor to support `avx2`.
Binaries with the label `pext` require your processor to support `bmi2`.

If you're unsure which to use or what your processor supports
- Download them all
- Open them and run `go depth 24`
- If it crashes, you can't use it.
- Pick the fastest one

Changes
- NN Trainer improvements
- Tuned hyper parameters
- Tuned Eval vs WDL weighting from training data
- L1 Penalty
- LMR Tweaks
- Cut Node reduction added to captures/promotions in LMR
- Simplified out useless reductions
- RFP Tweaks
- RFP Margin Adjusted
- NMP Tweaks
- Margin adjusted via `improving` flag
- Raw null search score returned

Bugs
- Resolved an issue with Berserk suffered from EG blindness due to NMP (Example)

Thanks
- Connor McMonigle (Seer's Author) for continued guidance and advice on NNs
- Andrew Grant (Ethereal's Author) for Open Bench. Development of Berserk would not be possible without this resource
- Kim Kåhre (Koivisto's Author) for rubber duck debugging and theorizing
- Ipmanchess for validating my AVX512 binary works

[Madeleine Birchfield]

Congrats on another release!

[jhonnold]

Thanks
- Connor McMonigle (Seer's Author) for continued guidance and advice on NNs
- Andrew Grant (Ethereal's Author) for Open Bench. Development of Berserk would not be possible without this resource
- Kim Kåhre (Koivisto's Author) for rubber duck debugging and theorizing
- Ipmanchess for validating my AVX512 binary works

I should also include NoobPwnFTW here as well who's resources are also critical in the development of Berserk!

[Damir]

Thanks a lot for this new release !

[matejst]

Impressive and fast progress! I hope you will successfully improve the NN architecture for the next version! (Anyway, I had just started testing the B7 version, when B8 is already there... ouf, ouf...)

[Rebel]

Berserk 8 for the GRL.

Games : 1600

Elo pool : 3471

http://rebel13.nl/a/grl.htm

[Wolfgang]

https://cegt.forumieren.com/t1628-testi ... 8-0nn#3130
No. of games: 1800
Av.Opp.ELO: 3435

[Jouni]

Great fight for number 3. position: SlowChess, Koivisto and Berserk! BTW Koivisto is the only one, that works 100% in old Fritz GUI .

[Rebel]

Estimated elo gain for Berserk 8 : +56

Code: Select all

Gambit Rating List
Running      : Gauntlet Berserk 8
Time Control : Time control : 40/120
Games        : 1600

Results from file gauntlet-berserk-8.pgn:

No. Name                 Win Draw Loss Unf.  Score Games       %
----------------------------------------------------------------
  1 Berserk 8           +431 =924 -245   *0  893.0  1600   55.8%
  2 Komodo-Dragon 2      +62 =119  -19   *0  121.5   200   60.8%
  3 Koivisto 7.9         +28 =136  -36   *0   96.0   200   48.0%
  4 SlowChess 2.8        +28 =132  -40   *0   94.0   200   47.0%
  5 Stockfish 10         +44  =93  -63   *0   90.5   200   45.2%
  6 Stockfish 9          +30 =101  -69   *0   80.5   200   40.2%
  7 Ethereal 13.25-NNUE  +13 =131  -56   *0   78.5   200   39.2%
  8 RubiChess 2.2        +16 =119  -65   *0   75.5   200   37.8%
  9 Komodo 14            +24  =93  -83   *0   70.5   200   35.2%

Total Games:    1600
White Wins:      337 (21.1%)
Black Wins:      339 (21.2%)
Draws:           924 (57.8%)
Unfinished:        0 (0.0%)

Estimated elo gain for Berserk_8
Elo pool : 3471
Berserk 7 : 3452.0
Berserk_8 : 3507.9
Difference : 55.9

Provisional Top-10

Code: Select all

   # PLAYER                 :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)     W     D     L  D(%)
   1 Stockfish 14.1         :  3709.2   13.0  2368.0    3000    79      57  1768  1200    32    40
   2 Komodo-Dragon 2.5      :  3646.6   11.8  2332.5    3400    69     100  1492  1681   227    49
   3 Lc0 v28                :  3581.0   12.7   591.5    1000    59     100   335   513   152    51
   4 Berserk 8              :  3512.2    7.9   893.0    1600    56      88   431   924   245    58
   5 Koivisto 7.9           :  3500.7   12.0  1252.0    2400    52      81   607  1290   503    54
   6 SlowChess 2.8          :  3485.1   10.3  1340.0    2600    52      67   627  1426   547    55
   7 Ethereal 13.25-NNUE    :  3453.0   10.3  2795.5    5800    48      55  1318  2955  1527    51
   8 RubiChess 2.2          :  3407.3   10.6  2547.5    6000    42      90  1073  2949  1978    49
   9 Seer 2.4.0             :  3400.4   13.7   804.5    1400    57      52   478   653   269    47
  10 Igel 3.0.5             :  3354.8    7.3  2289.5    6100    38      59   846  2887  2367    47

Berserk now third best original CPU engine.

[Madeleine Birchfield]

Estimated elo gain for Berserk 8 : +56

Code: Select all

Gambit Rating List
Running      : Gauntlet Berserk 8
Time Control : Time control : 40/120
Games        : 1600

Results from file gauntlet-berserk-8.pgn:

No. Name                 Win Draw Loss Unf.  Score Games       %
----------------------------------------------------------------
  1 Berserk 8           +431 =924 -245   *0  893.0  1600   55.8%
  2 Komodo-Dragon 2      +62 =119  -19   *0  121.5   200   60.8%
  3 Koivisto 7.9         +28 =136  -36   *0   96.0   200   48.0%
  4 SlowChess 2.8        +28 =132  -40   *0   94.0   200   47.0%
  5 Stockfish 10         +44  =93  -63   *0   90.5   200   45.2%
  6 Stockfish 9          +30 =101  -69   *0   80.5   200   40.2%
  7 Ethereal 13.25-NNUE  +13 =131  -56   *0   78.5   200   39.2%
  8 RubiChess 2.2        +16 =119  -65   *0   75.5   200   37.8%
  9 Komodo 14            +24  =93  -83   *0   70.5   200   35.2%

Total Games:    1600
White Wins:      337 (21.1%)
Black Wins:      339 (21.2%)
Draws:           924 (57.8%)
Unfinished:        0 (0.0%)

Estimated elo gain for Berserk_8
Elo pool : 3471
Berserk 7 : 3452.0
Berserk_8 : 3507.9
Difference : 55.9

Provisional Top-10

Code: Select all

   # PLAYER                 :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)     W     D     L  D(%)
   1 Stockfish 14.1         :  3709.2   13.0  2368.0    3000    79      57  1768  1200    32    40
   2 Komodo-Dragon 2.5      :  3646.6   11.8  2332.5    3400    69     100  1492  1681   227    49
   3 Lc0 v28                :  3581.0   12.7   591.5    1000    59     100   335   513   152    51
   4 Berserk 8              :  3512.2    7.9   893.0    1600    56      88   431   924   245    58
   5 Koivisto 7.9           :  3500.7   12.0  1252.0    2400    52      81   607  1290   503    54
   6 SlowChess 2.8          :  3485.1   10.3  1340.0    2600    52      67   627  1426   547    55
   7 Ethereal 13.25-NNUE    :  3453.0   10.3  2795.5    5800    48      55  1318  2955  1527    51
   8 RubiChess 2.2          :  3407.3   10.6  2547.5    6000    42      90  1073  2949  1978    49
   9 Seer 2.4.0             :  3400.4   13.7   804.5    1400    57      52   478   653   269    47
  10 Igel 3.0.5             :  3354.8    7.3  2289.5    6100    38      59   846  2887  2367    47

Berserk now third best original CPU engine.

Also stronger than Stockfish 11 on the GRL.