Stable-baselines: DDPG and SAC for discrete action space.
[question] Is there any reason why DDPG and SAC don't have the implementation for discrete action space? And will appreciate it there are any suggestions for applying the DDPG with continuous action space on the discrete one. Thanks!
soloist96
All 4 comments
Hello,
For DDPG, you can already find an answer here: https://github.com/hill-a/stable-baselines/issues/37
For SAC, the implementation with discrete actions is not trivial and it was developed to be used on robots, so with continuous actions. Those are the main reason. Meanwhile, if you want to work with discrete actions, you have plenty of other algorithms that can do that (ACER, PPO, DQN, A2C, ACKTR, ...).
araffin
on 26 Jul 2019
It was more than a year ago. Any news on this topic recently @araffin ?
It would be nice if SAC can take discrete action space input.
cosmir17
on 5 Dec 2020
We have an issue about that in Stable-Baselines3 repo: https://github.com/DLR-RM/stable-baselines3/issues/157
But I would favor QR-DQN first in the contrib repo.
araffin
on 5 Dec 2020
Thank you for letting me know @araffin
cosmir17
on 5 Dec 2020
Related issues
smorad
路
3Comments
saeid93
路
3Comments
Antalagor
路
3Comments
HareshKarnan
路
3Comments
shwang
路
3Comments
Most helpful comment
Hello,
For DDPG, you can already find an answer here: https://github.com/hill-a/stable-baselines/issues/37
For SAC, the implementation with discrete actions is not trivial and it was developed to be used on robots, so with continuous actions. Those are the main reason. Meanwhile, if you want to work with discrete actions, you have plenty of other algorithms that can do that (ACER, PPO, DQN, A2C, ACKTR, ...).