An implementation of AlphaZero for the board game Tak. See also https://github.com/ViliamVadocz/tak
The repository contains several libraries and binaries:
takzero
is the main library which implements MCTS and the neural networksselfplay
is used during training to generate replays and exploitation targetsreanalyze
computes fresh targets from old replayslearn
takes targets fromselfplay
andreanalyze
to train new modelsevaluation
pits models against each otherpuzzle
runs the puzzle benchmarkanalysis
includes interactive game analysisgraph
computes Bayesian Elo from match results (fromevaluation
) and creates a graphtei
a TEI implementation
I also have a python
directory for miscellaneous Python scripts.
You will need the C++ Pytorch library (LibTorch). See tch-rs for installation instructions.
It's possible you may not be able to find these versions anymore.
In that case try downloading the newest and update the tch-rs
version in Cargo.toml
.
You may also need to set LIBTORCH_BYPASS_VERSION_CHECK
to 1
.
If you find some version works, please let me know so I can add it here.
Worked:
- Stable (2.1.2), CUDA 11.8, Release
- Stable (2.1.2), CUDA 11.8, Debug
Did not work:
- TODO
Worked:
- Stable (2.1.2), CUDA 11.8, Pre-cxx11 ABI
Did not work:
- TODO