xaxetrov/oscar

5IF project. Work on Starcraft2 + deepmind API.

OSCAR

OSCAR (Omniscient Starcraft Awesome Retaliator) is an autonomous team project conducted during a semester at INSA Lyon, France. It aims at experimenting artificial intelligence for Starcraft II, using the PySc2 API from DeepMind.
The project consists in:

• a hierarchical framework to unify some small specialized agent (read more)
• a scripted multi-agent bot playing the full game (read more, video)
• some tools and agents to experiment reinforcement learning (read more)

Installation (Unix)

The project requires Python 3.5.

git clone https://github.com/Xaxetrov/OSCAR.git
pip3 install pysc2
pip3 install gym
pip3 install baselines

Run (Unix)

cd OSCAR
Add current directory to Python path: export PYTHONPATH=$(pwd)

Scripted bot

The General agent should be launched with PySc2. For example,
python3 -m pysc2.bin.agent --agent_race terran --agent2_race terran --difficulty very_easy --map Flat64 --agent oscar.agent.commander.general.General

Deep Q-Network

training: python learning_tools/baseline/dqn/custom_learn.py
test: python learning_tools/baseline/dqn/custom_test.py

Asynchronous Actor-Critic Agents (A3C)

training: python learning_tools/A3C_learner/main.py
test: python learning_tools/A3C_test/a3c_tester.py

TODO

• Fix neural network save in DQL algorithm learning_tools/baseline/dqn/custom_learn.py
• Allow to generate the "general-learning-v0" environment with the config file as a parameter (and the map ?) ((this will not work when using the gym.make command but nothing forbids us to use our own constructor))

Authors

Bruno Godefroy

Edern Haumont

Jérome Liermann

Ruben Pericas-Moya

François Robion

Nicolas Six

oscar's Issues

Update

@anliec You are the last person to have worked on this project. What is the status of TODOs in the README.md and the issue you opened ?

@frobion @gegeAi @br-g @Sn0wFox tagging you fyi.

Gym 0.10

OSCAR is not compatible with gym 0.10 as they changed methods of the class gym.Env (_reset -> reset ...)
However gym 0.9.4 still work fine