Skip to content

Berserk 8
Compare
Choose a tag to compare
@jhonnold jhonnold released this 05 Dec 18:19
· 346 commits to main since this release
8
4268296

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

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

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.
Removed the 512 binaries due to reported issues
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

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
  • Noob for his hardware on Open Bench. Tests run quickly and without issue thanks to Noob.
  • Kim Kåhre (Koivisto's Author) for rubber duck debugging and theorizing
  • Ipmanchess for validating my AVX512 binary works
Assets 7
Loading