RL algorithms with JAX

Hands-on project after finishing Deepmind's RL course. Decided to learn and use JAX for the implementations.

🚧 The project is still in progress.

TODO

Utilities:

• Experience accumulator
  • ✔️ by episodes
  • 🔲 by transitions (🚧 in progress)
• ✔️ Training experiment

Environments:

• gym
  • ✔️ Black-jack
  • ✔️ Cartpole
  • 🔲 Atari (🚧 in progress)
• mujoco ?
• evogym (so cool, must try)

Algorithms:

• Value function approximator
  • ✔️ Tabular
  • ✔️ Linear
  • ✔️ Neural Nets
• Value approximation/heuristic
  • TD
    • ✔️ TD(0)
    • ✔️ n-step TD
    • ✔️ TD(λ)
  • Q-learning
    • ✔️ vanilla q-learning
    • 🔲 λ q-learning
• Simple agents
  • ✔️ Tabular + TD,Q (with ε-greedy)
  • ✔️ Linear + Q (with ε-greedy)
• DQN
  • ✔️ Barebones (NN + Q)
  • 🔲 Vanilla DQN (:construction: in progress)
  • 🔲 Rainbow?
• Policy Gradient
  • 🔲 Vanilla
  • 🔲 Trust Region/PPO?
• Model-based ?
• GVF ?
• Combining with Evolutionary ?