mirror of
https://github.com/wassname/pytorch-soft-actor-critic.git
synced 2026-08-05 13:20:53 +08:00
41 lines
966 B
Markdown
41 lines
966 B
Markdown
### Description
|
|
------------
|
|
Reimplementation of [Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement
|
|
Learning with a Stochastic Actor](https://arxiv.org/pdf/1801.01290.pdf).
|
|
|
|
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](https://github.com/openai/mujoco-py)
|
|
- [Plotly](https://plot.ly/)
|
|
- [PyTorch](http://pytorch.org/)
|
|
|
|
### Run
|
|
------------
|
|
Use the default hyperparameters.
|
|
|
|
#### For SAC (Gaussian Policy):
|
|
|
|
```
|
|
python main.py --algo SAC --env-name HalfCheetah-v2
|
|
```
|
|
#### For SAC (Gaussian Mixture Policy):
|
|
|
|
```
|
|
python main.py --algo SAC(GMM) --env-name HalfCheetah-v2 --k 4
|
|
```
|
|
|
|
### TODO
|
|
------------
|
|
|
|
- [x] Gaussian Policy
|
|
- [x] Reparameterization
|
|
- [x] Gaussian Mixture Model
|
|
- [x] Use 2 Q-functions
|
|
- [ ] Deterministic Policy
|
|
- [ ] Soft Actor-Critic (hard target update)
|
|
- [ ] Evaluate the trained Policy
|
|
|