Berserk 4.0.0

Changes

• SMP functionality added
• Various speed improvements
• King safety reverted to pre-3.3.0 implementation
• Endgame specific functionality ended
  • KPK Bitbase (credit to Cheng)
  • KXK knowledge
  • Scale strong side scoring based on board state
• Tactical moves included in LMR
• Improvement on time management
• IID

Special Thanks

• Koivisto authors for guidance/assistance
• Lars for preliminary testing

Testing

Preliminary results by Lars show that this release is most likely +120 Elo over 3.2.1

Berserk 3.3.0

Changes

• Singular extension
• LMR based around historical score
• Improvements around passed pawn evaluation
• Finer tuning of values
• Outpost improvements
  • Bishop outposts and "seeing" outposts
• King safety shelter built into quadratic calculation
  • Previously was a separate linear calculation

Note: Bolded changes reported largest Elo gains in self-play testing

Special Thanks

• Lars for consistent encouragement
• Martin Sedlak for guidance
• Aleksey Glebov for hosting a wonderful tournament with Berserk in it, and finding the picture I now use

Testing

Games played at 120+1 TC using cutechess-cli

Score of Berserk 3.3.0-rc vs Cheng 4.40: 28 - 24 - 44  [0.521] 96
...      Berserk 3.3.0-rc playing White: 16 - 11 - 21  [0.552] 48
...      Berserk 3.3.0-rc playing Black: 12 - 13 - 23  [0.490] 48
...      White vs Black: 29 - 23 - 44  [0.531] 96
Elo difference: 14.5 +/- 51.5, LOS: 71.0 %, DrawRatio: 45.8 %

Score of Berserk 3.3.0-rc vs Mr Bob v1.0.0: 28 - 21 - 47  [0.536] 96
...      Berserk 3.3.0-rc playing White: 13 - 8 - 27  [0.552] 48
...      Berserk 3.3.0-rc playing Black: 15 - 13 - 20  [0.521] 48
...      White vs Black: 26 - 23 - 47  [0.516] 96
Elo difference: 25.4 +/- 49.9, LOS: 84.1 %, DrawRatio: 49.0 %

Score of Berserk 3.3.0-rc vs Monolith 2: 28 - 28 - 40  [0.500] 96
...      Berserk 3.3.0-rc playing White: 19 - 11 - 18  [0.583] 48
...      Berserk 3.3.0-rc playing Black: 9 - 17 - 22  [0.417] 48
...      White vs Black: 36 - 20 - 40  [0.583] 96
Elo difference: 0.0 +/- 53.4, LOS: 50.0 %, DrawRatio: 41.7 %

Berserk 3.2.1

Changes

This is a small set of changes, very little Elo gain (all from additional tuning)

• Retuned evalution parameters from 0 using a larger dataset
• Increase input buffer to support games 200+ moves
• Attempted to resolve zombie's when using OpenBench
  • Don't think it's working

Berserk 3.2.0

Changes

• Values tuned using Texel's Tuning Method
  • Used quiet-labeled.epd as the training set

Berserk 3.1.0

Changes

This release is mostly for NPS and bug fixing

NPS

• Improvement piece tracking in Board
• Heavier usage on arrays instead of operations
• Evaluation stored in TT
• Fifty move rule combined with isRepetition

Bugs

• Zobrist collisions due to bad PRNG resolved
  • Special thanks to Martin (author of Cheng) for assisting me with this
• MVV_LVA sorting of EP moves
• Reset killers at start of new ply

Bench: 10593829
Short test vs Berserk 3.0.0

ELO   | 36.32 +- 35.94 (95%)
SPRT  | 60+1s Threads=1 Hash=32MB
LLR   | 1.71 (-2.94, 2.94) [-10.00, 10.00]
Games | N: 192 W: 60 L: 40 D: 92

Berserk 3.0.0

Changes

• Improved LMR
• Added LMP
• King Safety evaluation improvement
• Threat evalution
• SEE pruning in qsearch
• Aspiration window adjustments

There are probably other's that I'm missing.

Largest improvements for this release seem to come from LMP, SEE pruning in qsearch, and threat evaluation.

Tests against other engines:

Score of Berserk 3.0.0 vs Delfi 5.4: 48 - 17 - 31  [0.661] 96
...      Berserk 3.0.0 playing White: 24 - 9 - 15  [0.656] 48
...      Berserk 3.0.0 playing Black: 24 - 8 - 16  [0.667] 48
...      White vs Black: 32 - 33 - 31  [0.495] 96
Elo difference: 116.4 +/- 59.4, LOS: 100.0 %, DrawRatio: 32.3 %

Score of Berserk 3.0.0 vs Donna 4.1: 38 - 16 - 42  [0.615] 96
...      Berserk 3.0.0 playing White: 22 - 5 - 21  [0.677] 48
...      Berserk 3.0.0 playing Black: 16 - 11 - 21  [0.552] 48
...      White vs Black: 33 - 21 - 42  [0.563] 96
Elo difference: 81.1 +/- 52.8, LOS: 99.9 %, DrawRatio: 43.8 %

Score of Berserk 3.0.0 vs Berserk 2.0.0+popcnt: 65 - 3 - 28  [0.823] 96
...      Berserk 3.0.0 playing White: 30 - 1 - 17  [0.802] 48
...      Berserk 3.0.0 playing Black: 35 - 2 - 11  [0.844] 48
...      White vs Black: 32 - 36 - 28  [0.479] 96
Elo difference: 266.9 +/- 66.0, LOS: 100.0 %, DrawRatio: 29.2 %

Score of Berserk 3.0.0 vs Arminius 2018-12-23: 38 - 18 - 40  [0.604] 96
...      Berserk 3.0.0 playing White: 18 - 9 - 21  [0.594] 48
...      Berserk 3.0.0 playing Black: 20 - 9 - 19  [0.615] 48
...      White vs Black: 27 - 29 - 40  [0.490] 96
Elo difference: 73.5 +/- 53.8, LOS: 99.6 %, DrawRatio: 41.7 %

Score of Berserk 3.0.0 vs Velvet v1.2.0: 40 - 24 - 32  [0.583] 96
...      Berserk 3.0.0 playing White: 25 - 5 - 18  [0.708] 48
...      Berserk 3.0.0 playing Black: 15 - 19 - 14  [0.458] 48
...      White vs Black: 44 - 20 - 32  [0.625] 96
Elo difference: 58.5 +/- 57.6, LOS: 97.7 %, DrawRatio: 33.3 %

Score of Berserk 3.0.0 vs Devel 3.8.4: 32 - 26 - 38  [0.531] 96
...      Berserk 3.0.0 playing White: 14 - 17 - 17  [0.469] 48
...      Berserk 3.0.0 playing Black: 18 - 9 - 21  [0.594] 48
...      White vs Black: 23 - 35 - 38  [0.438] 96
Elo difference: 21.7 +/- 54.4, LOS: 78.5 %, DrawRatio: 39.6 %

Score of Berserk 3.0.0 vs Cheng 4.40: 8 - 63 - 25  [0.214] 96
...      Berserk 3.0.0 playing White: 4 - 27 - 17  [0.260] 48
...      Berserk 3.0.0 playing Black: 4 - 36 - 8  [0.167] 48
...      White vs Black: 40 - 31 - 25  [0.547] 96
Elo difference: -226.5 +/- 67.9, LOS: 0.0 %, DrawRatio: 26.0 %

Berserk 2.0.0

Changes

The largest changes that come from this release are evaluation changes. PSQT's dominated in 1.2.2, but that lead to poor pawn structure and poor king safety. This has been addressed in 2.0.0 with all PSQT's being flattened (more understandable human values) and specific evaluation parameters around pawn structure and king safety. Along with this, piece mobility has been added.

Estimated Elo ~2450

Berserk 1.2.2

Changes

• Attempt to support newer versions of gcc better
• Berserk supports both popcount and non-popcount versions
  • make popcount
  • make

Berserk 1.2.1

Changes

• Fixes issue with LMR dropping into quiesce search

Berserk 1.2.0

Changes

• Pruning
  • Null Move
  • Futility
  • SEE
• Extension
  • Check
• Reductions
  • LMR
• Move Ordering
  • MVV/LVA
  • SEE
  • Killers
  • Counters

Still no evaluation changes at this time

• Still tapered material + psqt