TODO-IDEA.txt


- Compare SGD/Adam (SGD should be better for testing)

- Statistics:
    - Keep track of memory usage
    - Save videos of train and test
    - Save curves

- Save/load
    - Save model during training. -> Done (Adrien)

- Run Q-learning OR Q-learning+ by just changing one boolean parameter

- Implements Deep Q learning+ (Step 3+)
    - Fixed Q-target. -> Done. (Lostindark); Question : How to find 'optimal' tau ?
    - double DQN. -> Done. (Lost)
    - dueling DQN. -> Done. (Lost)
    - PER. -> Done. (Lost); Question : Optimal a ?

- Lauch code with parameters (Scripts)

- Try new architectures
- Try other games


References :

Tutorial link:
https://simoninithomas.github.io/Deep_reinforcement_learning_Course/#syllabus

Reference paper:
chrome-extension://oemmndcbldboiebfnladdacbdfmadadm/https://nihit.github.io/resources/spaceinvaders.pdf

Another DQN implementation:
https://github.com/tambetm/simple_dqn

More advanced tuto on RL:
https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A/videos

About Bellman equations :
https://joshgreaves.com/reinforcement-learning/understanding-rl-the-bellman-equations/