Skip to content

Aurora, a chess engine with an MCTS-Minimax hybrid search and an NNUE evaluation

License

Notifications You must be signed in to change notification settings

kjljixx/Aurora-Chess-Engine

Repository files navigation

Aurora

Logo generated by Bing Image Creator

General

Aurora is an actively developed UCI (Universal Chess Interface) compatible chess engine. This means that Aurora is a command-line-based engine and you will need a UCI-Compatible GUI in order to intuitively use Aurora. See the "Usage" section below for usage instructions.

See Playing Strength Testing for Aurora's playing strength.

Usage

To use Aurora, simply download a binary from the latest release, for example on Windows you would download aurora-x-windows.exe, where x is the version of the latest release. Go to a chess GUI of your choice, and find instructions for adding engines, and add this binary.

If you want to build Aurora from the source code, download the source code, navigate to the directory, and run

make

You should now have a file named aurora.exe, which is the engine. Now go to your GUI and find instructions for adding engines. When the GUI prompts you to choose a file/engine, choose aurora.exe.

See an explanation of all the options on the wiki

Technical Details

Aurora uses a UCT (Upper Confidence Bounds applied to Trees) search with an NNUE (Efficiently Updatable Neural Network)

Search

Here are some details about the search:

  • Builds a search tree with nodes
  • Nodes to expand are selected with UCT
  • Values backpropagated up the tree are done so with minimax

Evaluation

A NNUE trained with bullet and options (the options are for andromeda-3.nnue):

  • Arch: (768->256)x2->1
  • Scale: 400
  • Batch Size: 16384
  • Batches / Superbatch: 6104
  • Positions / Superbatch: 100007936
  • Superbatches: 88
  • WDL scheduler: constant 0.0
  • LR scheduler: start 0.001 gamma 0.1 drop every 44 superbatches
  • Device: NVIDIA GeForce MX550
  • Threads: 4
  • Positions: 154143704

About

Aurora, a chess engine with an MCTS-Minimax hybrid search and an NNUE evaluation

Resources

License

Stars

Watchers

Forks

Packages

No packages published