LastOrder is a StarCraft AI bot developed by Bilibili AI Lab.
LastOrder uses a pre-trained model to do macro action selection during game. The model combines APE-X DQN with LSTM and is trained distributively on cluster.
The detailed info can be found in:
Macro action selection with deep reinforcement learning in StarCraft
LastOrder comprises of two parts:
- Actor. running the StarCraft client and processing messages with model.
- In training mode, actor periodically receive updated model from learner.
- In evaluation mode, actor load the local pre-trained model and run independently without learner.
- Learner. running the training procedure and updating model to actors.
- BWAPI 4.2.0.
- VisualStudio 2017 supplied with the Development Tools for C++.
- Python 3.6 with tensorflow and zmq
- Setting the env variables:
BWAPI_DIR
, pointing to BWAPI 4.2.0 installation. - Copy the libzmq_LastOrder.dll into C:\Windows
- Copy pre-trained model data to AI\saved_model
- Python 3.6 with tensorflow and zmq
- StarcraftAITournamentManager