Build instructions

CrazyAra can be build in different modes:

• MODE_CRAZYHOUSE: Build with crazyhouse only support (default)
• MODE_CHESS: Build with chess + chess960 only support
• MODE_LICHESS: Build with lichess variants support
• MODE_POMMERMAN: Build with Pommerman support
• MODE_OPEN_SPIEL: Build with OpenSpiel game interface support.

and different neural network inference back-ends:

• BACKEND_TENSORRT: Build with TensorRT back-end (default)
• BACKEND_MXNET: Build with MXNet backend (Blas/IntelMKL/CUDA/TensorRT) support
• BACKEND_TORCH: Build with Torch backend (CPU/GPU) support
• BACKEND_OPEN_VINO: Build with OpenVino backend CPU support

Only one of these mode can be active at a time with the exception that BACKEND_MXNET and BACKEND_TENSORRT can both be ON at a time. By default the native TensorRT back-end without MXNet is used.

The CMake option USE_RL allows building the executable with reinforcement learning support and requires additional dependencies for exporting the generated data set.

Custom MCTS-Build options:

Build flag	Description
MCTS_STORE_STATES	Allows storing the states within the nodes. This is highly memory consuming but can be beneficial for environments with a high do_action() runtime cost and either low memory consumption or low total number of nodes in the search tree.
MCTS_SINGLE_PLAYER	The MCTS is only computed for a single player. The value won't be negated during back-propagation.
MCTS_TB_SUPPORT	The MCTS uses tablebase support during search.

For all options, please refer to CMakeLists.txt.

---

This guide is structured into three parts and is available for Linux, Mac and Windows.

• 1. Download dependencies
• 2. Download / build neural network inference library
• 3. Build CrazyAra binary