PPO
IMPALA
In the parameters.json configuration file, you can replace the algorithm by changing the algo entry.
This repository targets the SMAC2 environment.
And, this repository has been validated for training on 5vs5 sc2 marine battles.
You must ensure the complete installation of the SMAC2 environment on the machine that will be used as a Worker.
This repository is developed assuming the use of tmux-interactive-session and a Python Miniconda environment on an Ubuntu machine.
In the run.py script, you need to set the virtual environment name vir_env_name.
In the machines.json configuration file, you need to set the account, IP, and port for the machines to be used as Learner, Manager, and Worker.
Additionally, in the Worker machine settings, you need to decide how many independent worker processes to use.
After all settings are properly configured, you need to run python run.py
Build the library for your system or use pre-built binaries for convenience.
- Follow the guide to build the library from source.
- Alternatively, pre-built binaries are available:
- Windows:
*.dll - Linux:
*.so
- Windows:
WSL2 and NVIDIA-GPU According to the JAX official GitHub repository, using NVIDIA GPUs in WSL2 is still considered "experimental". It is recommended to use a native Linux environment instead for better stability.
Compatibility of NVIDIA Driver, CUDA Toolkit, and cuDNN Ensure strict compatibility between the NVIDIA driver, CUDA Toolkit, and cuDNN versions. Any incompatibility here could result in numerous errors throughout the JAX project and significantly impact the overall code functionality.
num_worker: 30
