This is a reinforcement learning agent in pysc2 environment. It's based on GA3C .

Note: This agent could reach 25 mean score on MoveToBeacon mini-game (which is good), but on DefeatRoaches it could only get around 60 mean score (No matter using Atari-net or FullyConv-net). This may caused by bad hyper-parameters or off-policy update during training. However, the throughput is better than single-machine A3C and batched A2C.

• python 3 or above
• pysc2 2.0.1
• tensorflow or tensorflow-gpu >= 1.8.0

1. Issue sh _clean.sh to clean the saved checkpoints of early experiments. (Make sure you change the directory name if you want to keep the checkpoints)
2. Run sh_train.sh command to start training.
3. You can change experiement parameters in Config.py.

• SC2_MAP_NAME: The map to train on
• IMAGE_SIZE: The image length of feature maps
• OPTIMIZER: The optimizer used in training
• LEARNING_RATE_START, LEARNING_RATE_END: Beginning learning rate and the learning rate in the end

Reinforcement Learning thorugh Asynchronous Advantage Actor-Critic on a GPU

StarCraft II: A New Challenge for Reinforcement Learning