PySC2 agents

This is a simple implementation of DeepMind's PySC2 RL agents. In this project, the agents are defined according to the original paper, which use all feature maps and structured information to predict both actions and arguments via an A3C algorithm.

Requirements

PySC2 is a learning environment of StarCraft II provided by DeepMind. It provides an interface for RL agents to interact with StarCraft II, getting observations and sending actions. You can follow the tutorial in PySC2 repo to install it.

pip install s2clientprotocol==1.1
pip install pysc2==1.1

Python packages might miss: tensorflow and absl-py. If pip is set up on your system, it can be easily installed by running

pip install absl-py
pip install tensorflow-gpu

Getting Started

Clone this repo:

git clone https://github.com/xhujoy/pysc2-agents
cd pysc2-agents

Testing

Download the pretrained model from here and extract them to ./snapshot/.
Test the pretrained model:

python -m main --map=MoveToBeacon --training=False

You will get the following results for different maps.

	MoveToBeacon	CollectMineralShards	DefeatRoaches
Mean Score	~25	~62	~87
Max Score	31	97	371

Training

Train a model by yourself:

python -m main --map=MoveToBeacon

Notations

Different from the original A3C algorithm, we replace the policy penalty term with epsilon greedy exploration.
When train a model by yourself, you'd better to run several times and choose the best one. If you get better results than ours, it's grateful to share with us.

Licensed under The MIT License.