BlackCore v6.0

The new release improves many aspects of the search and also increases the number of hidden layer neurons from 512->768. Now BlackCore supports MultiPV search, keep in mind that enabling it drastically decreases playing strength.

Selfplay elo gain with Pohl opening book and 8+0.08s time control:

Result of BlackCore-v6-0 vs BlackCore-v5-1: 510 - 146 - 464 (0.66)
Ptnml:        WW     WD  DD/WL     LD     LL
              79    251    187     41      2
Elo difference: 117.16 +/- 15.76

This version also introduces an aesthetically more pleasing command line interface.
Sending the uci command automatically disables it.

Feel free to contact me with any questions or problems regarding BlackCore. For choosing a binary please see the installation section in the readme.

Changelog

• Pretty command line interface
• New network
• MultiPV
• Better mate reporting
• Improved time management
• Faster search
• Threshold based SEE
• Fixed memory overflow
• Razoring tweak
• New NMP scale (by Zuppa)
• Fixed search stack (by Disservin)
• Tuned search parameters
• More aggressive aspiration window

BlackCore v5.1

This new minor release contains only a few, but more significant patches compared to the previous version, which resulted a big selfplay elo gain. The biggest change is to the new 4 king buckets system in the NNUE network.

ELO   | 126.18 +- 17.78 (95%)
CONF  | 8.0+0.08s Threads=1 Hash=16MB
GAMES | N: 1000 W: 489 L: 141 D: 370

FRC/DFRC support is being planned for the next major release.

Feel free to contact me with any questions or problems regarding BlackCore. For choosing a binary please see the installation section in the readme.

Changelog

• Cleaner code base
• King buckets
• Killer moves (BlackCore used one of may own inventions, history difference heuristic a replacement to killer moves, but it's removal unfortunately gained elo)
• HH decay
• Fixed mate scores
• Fixed counter moves

BlackCore v5.0

During the development of v5.0 the previous NNUE trainer was rewritten in plain c++ to support multi-threading which should significantly cut down the training time (1 week->1-3 days) as I don't have any GPU resources. This allows me to experiment with different ideas and hopefully make a new net, that actually gains elo.

A critical bug in the repetition detection was fixed and many aspects of the search and the transposition table were improved. Huge portion of the codebase was documented and cleaned up.
The new version also supports the Syzygy tablebase using Fathom.

Estimated elo on 1 thread: 3150-3250
Estimated elo on 4 threads: 3200-3300

Feel free to contact me with any questions or problems regarding BlackCore. For choosing a binary please see the installation section in the readme.

Changelog

• Syzygy support
• More UCI support
• Time management changes
• Randomize draw values
• IID changes
• Razoring in PV nodes
• Tuned parameters
• RFP only in non PV nodes (by Disservin)
• LMR formula change
• Prevent TT cutoff after a null move
• TT always replace
• 16 byte TT entries
• Cleaned up code
• Added logo by Graham
• Capped HH score
• Fixed rare crash
• Fixed stack move
• Fixed repetition detection

BlackCore v4.0

This version brings general increase to search stability, which can be easily noticable in LTC playing strength and also multi-threading support with lazy SMP.

Estimated elo on 1 thread: ~3100
Estimated elo on 4 threads: ~3200

Feel free to contact me with any questions or problems regarding BlackCore. For choosing a binary please see the installation section in the readme.

HOTFIX 1:
Fixed crash with games longer than 250 ply.

Changelog

• Lazy SMP
• Counter move heuristic
• LMP now prunes only quiet moves
• Better history scores
• Root node move ordering by nodes searched
• Singular extension
• History difference
• Flipped improving in LMR
• LMR can extend depth

BlackCore v3.0

Stronger than ever! BlackCore v3.0 is smarter with a perspective aware NNUE architecture (trained on Berserk data using CoreTrainer) and faster than the last release. The overall search stability was also increased and evaluations became a lot more precise. Selfplay gain: ~140 elo.

Estimated elo: 3050-3100

Feel free to contact me with any questions or problems regarding BlackCore. For choosing a binary please see the installation section in the readme.

Changelog

• Better time management
• SEE en passant fix
• Forward futility pruning
• Improved capture move ordering
• New net architecture: 2 x (768 -> 256) -> 1
• Smarter late move reduction
• Templated search
• Better TT entries
• Fixed move generation "bug", that restricted search speed

BlackCore NNUE v2.0

After a lot of work, I am finally ready to release BlackCore v2.0. This version introduces NNUE which is the biggest reason of the giant elo improvement. v1.1 unable to beat v2.0 from balanced positions so estimating the real gain is hard, from selfplay it's 600 elo (but again, the only losses are from openings which are very unbalanced).

Elo on CCRL (as of 2022 nov. 9): ~2970

Feel free to contact me with any questions or problems regarding BlackCore. For choosing a binary please see the installation section in the readme.

Changelog

• Evaluation is now based on a neural network trained using CoreTrainer
• Vectorized accumulator updates
• Many changes to transposition table
• Tuned search parameters
• Late move pruning improvements

HOTFIX 1
Fix for time losses in some tc

BlackCore v1.1

v1.1 added many improvements which is resulted an ~170 elo gain against 1.0.

Changelog

• Small improvements
• Fixed memory handling issues
• TT in Qsearch
• Small TT speed-ups
• Better move ordering
• Using log formula for LMR
• Built-in texel tuning (values are still untuned)
• Piece-square tables
• Fixed a1a1 invalid move
• More advanced king safety
• Improved time management
• Aspiration window

BlackCore v1.0

BlackCore v1.0 Release!

I'm happy to announce the first major release of BlackCore!
You can download the windows and linux executables, there are 3 options for each, popcnt/modern/bmi2. For any further details see the readme.

Estimated elo: ~2200