Relies on the repo https://github.com/simonmeister/pysc2-rl-agents for implementations of action / observation space pre-processing, network architectures.
Entire backend is in Pytorch, with TF for tensorboard.
python run.py --envs 32 --map MoveToBeacon
MoveToBeacon | CollectMineralShards |
---|---|
![]() |
![]() |
- Train on other mini-games
- Use replay data
- Optimize the Runner to work with Torch tensor instead of numpy array
- Multi-GPU training