Zahak 10.0

What's Changed

• Search tweaks
• Reintroduce val phase
• Faster make/unmake
• TT replacement changes
• Add move overhead UCI option
• New and slightly improved neural network (same architecture)
• Introduced a new Capture History
• Much better SMP scaling
• Introduce IIR again
• Reset Killer moves in between searches

This version is almost 50+ elo over Zahak 9.0

9.0

Changes since 8.0

Usability improvements

• MultiPV Support
• Supporting go searchmoves ..., go mate ... and go nodes ... UCI go commands, which makes the engine better at finding mate and also in analyzing (searchmoves).
• Support of 7 different skill levels (1 being the weakest, while 7 is the strongest). Anchored around Rustic Alpha 3, I found that, based on CCRL the ratings will probably translate to the following:
  • Skill Level 1: 1270
  • Skill Level 2: 1440
  • Skill Level 3: 1630
  • Skill Level 4: 1856
  • Skill Level 5: 2004
  • Skill Level 6: 2074

Strength improvements

• A much stronger network is included, trained on more than 500 million self-play positions.
• A much more improved search
• Better move ordering thanks to counter history and follow-up history
• Lots of bug fixes!
• Much better multi-threading support.

Which binary to use?

• Ubunut/Linux 64-bit (not Raspberry Pi, AVX only): zahak-linux-amd64-9.0-avx
• Apple Mac OSX (64-bit, Intel, AVX only): zahak-darwin-amd64-9.0-avx
• MS Windows 64-bit (AVX only): zahak-windows-amd64-9.0-avx.exe

Acknowledgment

It is worth mentioning that all the skills from 1 to 6 are based on networks that are trained by Alan Cooper (Scally on CCC). Many thanks to him.

8.0

Changes since 7.1

• New network architecture
• EGBT support (Syzygy)
• Improved search
• +300 elo hopefully

Which binary to use?

• Ubunut/Linux 64-bit (not Raspberry Pi, AVX only): zahak-linux-amd64-8.0-avx
• Apple Mac OSX (64-bit, Intel, AVX only): zahak-darwin-amd64-8.0-avx
• MS Windows 64-bit (AVX only): zahak-windows-amd64-8.0-avx.exe

Thanks to Basti and Scally for having interest in Zahak and generating either networks for lower skill levels (Scally), or training data to train the default network

7.1

This version is bit by bit equivallent to 7.0, except the UCI has been "fixed" to work with BanksiaGUI and DroidFish

7.0

Changes since 6.2

• Original and brand-new NNUE based evaluation function
• Fix reporting mate scores
• Removing futility pruning in quiescence search
• Various search parameter tuning to adapt it to NNUE
• Exposing EvalFile and BookPath UCI options
• NPS improvement

Which binary to use?

• Ubunut/Linux 64-bit (not Raspberry Pi, AVX only): zahak-linux-amd64
• Raspberry Pi 3 family/Android: zahak-linux-arm32
• Raspberry Pi 4: zahak-linux-arm64
• Apple Mac OSX (64-bit, Intel, AVX only): zahak-darwin-amd64
• MS Windows 32-bit: zahak-windows-386
• MS Windows 64-bit (AVX only): zahak-windows-amd64

6.2

Fixes mutl-threading, bit-by-bit equivalent for single-threaded searches with 6.0 and 6.1

6.1

Bit-by-bit equivalent to 6.0 on 64-bit architecture, but fixes a segfault issue in 32-bit architecture (RaspberryPi and Android builds)

6.0

Changes since 5.0

Search

• Lazy SMP (Multi Threading)
• ProbCut
• Singular Extension
• Multi-Cut
• Countermove heuristics for move ordering
• History Leaf Pruning

Evaluation

• Draw Pattern Recognition: RvN, RvB, RvRB, RvRN, KPK, OCB, Wrong Bishop
• King safety, and king virtual mobility
• Incremental Eval
• Retuned Eval based on LiChess dataset
• Passed pawn evaluation based on rank

Performance

• Move generation speedup
• NPS increase

Which binary to use?

• Ubunut/Linux 64-bit (not Raspberry Pi): zahak-linux-amd64
• Raspberry Pi 3 family/Android: zahak-linux-arm32
• Raspberry Pi 4: zahak-linux-arm64
• Apple Mac OSX (64-bit, Intel): zahak-darwin-amd64
• Apple Mac OSX (M1 CPU): zahak-darwin-m1-arm64
• MS Windows 32-bit: zahak-windows-386
• MS Windows 64-bit: zahak-windows-amd64

5.0

Changelog:

• SEE pruning in normal search
• LMR and pruning tweaks
• Re-doing move history heuristic
• Adding evaluation terms for Bishop pair
• Adding evaluation terms for Knight outposts
• New and improved time management
• Re-implemented mobility/aggressivity, so that different types of pieces are treated differently
• Pawnhash
• Fixing various bugs in search

Which binary to use?

• Ubunut/Linux 64-bit (not Raspberry Pi): zahak-linux-amd64
• Raspberry Pi 3 family/Android: zahak-linux-arm32
• Raspberry Pi 4: zahak-linux-arm64
• Apple Mac OSX (64-bit, Intel): zahak-darwin-amd64
• Apple Mac OSX (M1 CPU): zahak-darwin-m1-arm64
• MS Windows 32-bit: zahak-windows-386
• MS Windows 64-bit: zahak-windows-amd64

4.0

Changelog:

• King Safety, and Pawn Shield
• Pawn structure evaluation is redone, with various mis-evaluation fixes
• More aggressive Null Move Pruning
• NPS increase and performance optimizations (almost 40% more NPS)
• Add Internal Iterative Deepening
• Add Enhanced Futility Pruning
• Pseudo Legal Moves
• Fix SEE pruning
• Fix Hash-move and move ordering for pawn promotion
• Search Parameter tweaking

Which binary to use?

• Ubunut/Linux 64-bit (not Raspberry Pi): zahak-linux-amd64
• Raspberry Pi 3 family/Android: zahak-linux-arm32
• Raspberry Pi 4: zahak-linux-arm64
• Apple Mac OSX (64-bit, Intel): zahak-darwin-amd64
• Apple Mac OSX (M1 CPU): zahak-darwin-m1-arm64
• MS Windows 32-bit: zahak-windows-386
• MS Windows 64-bit: zahak-windows-amd64