This repository contains a series of tutorials on Deep Reinforcement Learning (DRL). This includes slides as well as experiments. Going forward I plan on adding exercises as well as complementary blog posts. So stay tuned!
- Slides: Includes DQN, Double DQN, Prioritized Experience Replay & Dueling DQNs
- Experiments: Provides code to implement all of the above.
- Blog Post I/II: Covering all algorithms from Fitted Q-Learning to Categorical DQNs.
- Replicating the experiments:
-
Create & activate a virtual env. Install the requirements.
-
Afterwards you can run all experiments by executing:
time bash run_experiments_dqn.sh dqn <cuda_device_id> time bash run_experiments_dqn.sh double-dqn <cuda_device_id> time bash run_experiments_dqn.sh per-dqn <cuda_device_id> time bash run_experiments_dqn.sh dueling-dqn <cuda_device_id> -
The visualizations for the different experiments as well as the mini double DQN illustration can be replicated by executing the notebook:
jupyter notebook viz_results.ipynb -
Finally, in order to visualize an episode rollout of a DQN agent at different stages do the following:
python train_dqn.py --SAVE_AGENT python enjoy_dense.py --AGENT 5000_MLP-DQN --TITLE 5000 python enjoy_dense.py --AGENT 40000_MLP-DQN --TITLE 40000 python enjoy_dense.py --AGENT 500000_MLP-DQN --TITLE 500000
-
