Skip to content

darthbatman/football-imitation-learning

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

9 Commits
 
 
 
 
 
 

Repository files navigation

football-imitation-learning

Application of imitation learning to football (soccer) video game.

Usage

Setup

First, navigate to the football directory with cd football.

To install the required dependencies and build the environment, run

pip3 install -r requirements.txt
pip3 install .

Episode Data Collection

To execute the AI in the environment and collect episode data, run python3 execute_ai.py SCENARIO REPRESENTATION ACTION_DATA_FILE_NAME >> ACTION_DATA_FILE_NAME'. where SCENARIO is the scenario the AI is executing in, REPRESENTATION is the observation type (raw, pixels, etc.), and where ACTION_DATA_FILE_NAME is the file to which action data will be written to by redirecting STDOUT.

The episodes (states and corresponding actions) for this execution will be stored at episodes/episodes_SCENARIO_TS.pkl, respectively, where TS is the ending timestamp of the execution in seconds from epoch time. The file to which action data was written to, denoted by ACTION_DATA_FILE_NAME, can be deleted.

The following is an example of how to execute the AI in the environment and collect episode data.

python3 execute_ai.py 11_vs_11_stochastic raw ad.csv >> ad.csv

The episodes for this execution will be stored at episodes/episodes_11_vs_11_stochastic_TS.pkl, where TS is the ending timestamp of the execution in seconds from epoch time.

Training and Validation

To run training and validation, run

python3 train.py --env_name ENV_NAME

where ENV_NAME is the name of the environment/scenario (e.g. academy_empty_goal_close).

The number of validation episodes, max length of each test episode, directory with the expert data, number of epochs, and batch size can be set using the flags num_episodes_val, episode_len, data_dir, num_epochs, and batch_size, respectively.

Notes

At the moment, the assumption is made that the agent will control the left team.

If pixels or pixels_gray is selected as the observation type, the execution will be rendered and displayed.

References

We are using the Google Football Research environment.