Skip to content

Latest commit

 

History

History
70 lines (48 loc) · 3.08 KB

README.md

File metadata and controls

70 lines (48 loc) · 3.08 KB
Raw

Equisetum

A UCI chess engine written in C++11. Equisetum uses Fail-Hard AB-search and NNUE evaluation. This approach resulted in an engine of moderate strength. One of the few strong engines using a fail-hard approach

NNUE

Equisetum uses an NNUE (efficiently updatable neural network) as an evaluation function.

Current implementation is a very simple (768-1024)x2 -> 1 network. It is trained on Equisetum self-play games using a mix of a play from a randomized startposition and a randomized positions derived from human high-bias positions.

Network is training using CudAD

Origins

Equisetum is basically a continuation of the Drofa chess engine, which in turn is started as fork of the Shallow Blue chess engine. My initial intention was to take weak, but stable and working chess engine and try to improve it, learning c++ along the way.

During my Drofa/Equisetum experiments huge chunk of knowlenge were received from:

  • VICE chess engine and tutorials.
  • Weiss chess engine, with clean and understandable implementations of complex features. Drofa/Equisetum use Weiss 1.0 LMP base reduction formulas. As well as HCE-tuning for Drofa versions
  • Several open source engines, mostly Ethereal and Stockfish

Special thanks to:

  • Terje Kirstihagen (Weiss author)
  • GediminasMasaitis (ChessDotCPP author) for explaining bulk of the NNUE concepts to me
  • Finn Eggers for creating CudAD which i use for NNUE training
  • Andrew Grant. AdaGrad paper and Ethereal chess engine are great sources of knowledge; Ethereal tuning dataset was a great help in tuning. As well as allowing me on main OpenBench instance
  • Kim Kahre, Finn Eggers and Eugenio Bruno (Koivisto team) for allowing Drofa on Koi OpenBench instance and motivating me to work on the engine
  • Jay Honnold (Berserk author) for helping me with NN stuff
  • OpenBench community for helping me with motivation, in finding bugs, teaching me (even if unknowingly) good programming practices and interesting discussions

Strength

Current Equisetum is estimated to be somewhere before Stockfish 11 and Stockfish 12 strength.

UCI commands

Equisetum supports following UCI commands:

  • BookPath
  • OwnBook
  • Threads (1 to 172),
  • Hash (16 to 65536)

These options can be set from your chess GUI or the UCI interface as follows:

setoption name OwnBook value true
setoption name BookPath value /path/to/book.bin

License

During developement, many concepts found in GPL-3 engines were used. Thus, as of 4.0.0, Equisetum also will be licensed as GPL-3.

2017 - 2019 © Rhys Rustad-Elliott (original Shallow Blue creator)

2020 - 2024 © Litov Alexander