Description

---

Reimplementation of Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor.

Contributions are welcome. If you find any mistake (very likely) or know how to make it more stable, don't hesitate to send a pull request.

Requirements

---

• mujoco-py
• Plotly
• PyTorch

Run

---

Use the default hyperparameters.

For SAC :

python main.py --env-name Humanoid-v2 --scale_R 20 

For SAC (Hard Update):

python main.py --env-name Humanoid-v2 --scale_R 20 --tau 1 --value_update 1000

For SAC (Deterministic, Hard Update):

python main.py --env-name Humanoid-v2 --scale_R 20 --deterministic True --tau 1 --value_update 1000

TODO

---

• Gaussian Policy
• Reparameterization
• Gaussian Mixture Model
• Use 2 Q-functions
• Evaluate the trained Policy
• Deterministic Policy (hard target update)