This is the pytorch example of the google football research, more contents will be updated soon!:soccer:
- python-3.6.8
- openai-baselines
- pytorch-1.1.0
- gfootball
- add more tasks and examples - full game is in plan.
- remove openai-baseline's functions.
- add more algorithms: IMPALA and Ape-X DQN.
- add multi-agent reinforcement learning algorithms (MARL)
Please install the gfootball
according to the instructions here.
- Make sure your
pip
is less than19
, the lastest version ofpip
will disable--process-dependency-links
.
conda install pip==18.1
- Install tensorflow (well, we don't need it, but it's required for
gfootball
).
pip install tensorflow
- Install gfootball.
git clone https://github.com/google-research/football.git
cd football
pip install .[tf_cpu] --process-dependency-links (we don't need GPU for tensorflow)
Train the simple example - academy_empty_goal_close
python train_example.py --cuda (if you have a GPU)
Play the demo:
python demo.py
academy_empty_goal_close | academy_run_to_score |
---|---|
academy_3_vs_1_with_keeper | academy_counterattack_easy |
---|---|