phraust1612 / MinervaSc2

machine learning project using DeepMind's PySc2

MinervaSc2

A machine learning project using DeepMind's PySC2 and Tensorflow.

I refered to Sunghun Kim's repository for DQN class, etc.

Usage

At first, please download my git and some prerequisites.

Here's my example.

git clone https://github.com/phraust1612/MinervaSc2.git
sudo pip3 install pysc2
sudo pip3 install numpy
sudo pip3 install tensorflow

Specify your saving directory. For default, DQN structure will be saved at 'saved/',

so be sure that 'saved/' directory exists or prepare for your own.

Reinforcement Learning

python3 trainingRL.py --start_episode (##) --num_episodes (##)

For default, starting episode and total number would be 0 and 100

Supervised Learning (with replays)

python3 trainingSL.py --replay (your replay directory) --repeat (no of repetition)

For default option, this will refer ~/StarCraftII/Replays/ for replay files

Composition

• trainingRL.py : run reinforcement learning loops via DQN.
• trainingSL.py : run supervised learning loops with your replay files.
• minerva_agent.py : contains an agent class which decides actions for every step.
• dqn.py : DQN network class, in order to devide target and learning networks.