please go APPENDIX – B Defensive Strategies Generation by Imitation Learning for this project.
- Clone this repo:
$ git clone https://github.com/chychen/RL_strategies- Environment Setting docker/ nvidia-docker:
$ docker pull jaycase/bballgail:init
$ nvidia-docker run --name {name} -it -p 127.0.0.1:6006:6006 -v {repo_path}RL_strategies/:/RL_strategies -w /RL_strategies jaycase/bballgail bash- Download the dataset. (You should put dataset right under the folder "{repo_path}/RL_strategies/data/") (QQ can't find the files, maybe you could try this one)
root@7bdb7335cde0:/RL_strategies# cd bball_strategies/data/
root@7bdb7335cde0:/RL_strategies/bball_strategies/data# wget http://140.113.210.14:6006/NBA/data/FPS5.npy
root@7bdb7335cde0:/RL_strategies/bball_strategies/data# wget http://140.113.210.14:6006/NBA/data/FPS5Length.npy- Filter out bad data.
root@7bdb7335cde0:/RL_strategies/bball_strategies/data# python3 preprocess.py- Create training dataset for GAIL. (transform data into state-action pair.)
root@7bdb7335cde0:/RL_strategies/bball_strategies/data# python3 create_gail_data.py- Reorder the data offense and defense positions by rule-based method.
- And, duplicate OrderedGAILTransitionData_52.hdf5 for multi process settings.
root@7bdb7335cde0:/RL_strategies/bball_strategies/data# python3 postprocess_data_order.py
root@7bdb7335cde0:/RL_strategies/bball_strategies/data# cp OrderedGAILTransitionData_52.hdf5 OrderedGAILTransitionData_522.hdf5All configurations are in "{repo_path}/RL_strategies/bball_strategies/scripts/gail/config.py"
root@7bdb7335cde0:/RL_strategies# python3 -m bball_strategies.scripts.gail.train --config=double_curiculumroot@7bdb7335cde0:/RL_strategies# tensorboard --logdir='logdir/gail_defense/{time stamp}-double_curiculum' --port=6006###Custimized Animation Players
- Buid upon kivy framework
- You can drag and drop the '.npz' files (recorded while tranining) into the player.
root@7bdb7335cde0:/RL_strategies# cd gui_tool/
root@7bdb7335cde0:/RL_strategies/gui_tool# python3 player.py- Visualize results by plotly framework
root@7bdb7335cde0:/RL_strategies# cd analysis/
root@7bdb7335cde0:/RL_strategies/analysis# python3 evaluation.pyBasically, the code is built upon the TensorFlow Agents Framework.
@article{hafner2017agents,
title={TensorFlow Agents: Efficient Batched Reinforcement Learning in TensorFlow},
author={Hafner, Danijar and Davidson, James and Vanhoucke, Vincent},
journal={arXiv preprint arXiv:1709.02878},
year={2017}
}