Train an AI agent to play QWOP using reinforcement learning and imitation learning. Check out the ACER video summary, DDQN world record video or the write-up on Medium.
Install Python 3.7
conda create -n py37 python=3.7
conda activate py37
Install dependencies
pip install -r requirements.txt
Install Windows-specific dependencies if on Windows
conda install -c intel mpi4py
Install Chromedriver for interacting with Chrome through Selenium. Then, start HTTP server for the game
python host_game.py
Train agent
python main.py --train
Test agent
python main.py --test
To save or use new models, the constant MODEL_PATH in main.py. Note that the agent
plays the game in real-time, so differences in latency between training and testing
computer will affect performance.
Record games using human input
python main.py --record
Train agent from recorded games
python main.py --imitate
Track training performance live or view metrics previous sessions
tensorboard --logdir ./tensorboard/
- The original QWOP game at http://www.foddy.net/Athletics.html
- Lachlan for his help in getting QWOP to run offline. https://github.com/etopiei/QWOP-Bot
- Keyboard input inspiration from https://github.com/juanto121/qwop-ai