- Python 3.5
- Tensorforce (scripts without "ppo"; use latest version from Github)
- OpenAI baselines (scripts with "ppo" in name)
- OpenAI gym
- pygame
- imageio (only to create gif movies)
- matplotlib (to display the learning curve)
- tensorflow
Each script starting with "train" corresponds to a separate model. By default, each script loads pre-trained parameters and displays the agent in action:
python train_XXX.py
To retrain a model from scratch (this might overwrite the saved parameters in the corresponding subfolder), use:
python train_XXX.py --train
To load pre-trained weights and save the first 3 episodes as a GIF-movie, use
python train_XXX.py --savegif
For a detailed description and evaluation of each model see the corresponding blog post. To display the learning curve of model XXX use
python analyze_learning_curve.py -d XXX/