Any demo for soft actor critic on discrete action space?

[KK666-AI]

Cool Work.

It seems that you have implemented sac to support discrete action space.
I wonder whether this project contains tiny demo from running soft actor critic on scenario of discrete action space.

Best,
Ken

[kengz]

Hi @lihuiknight, yes, you can run any spec file using the lab command.
For this particular case, u can run SAC on LunarLander, which is a discrete environment. For the command:

• spec file, which you can find at: slm_lab/spec/benchmark/sac/sac_lunar.json
• spec name: "sac_lunar" in the file
• lab mode: "train"

All in all, you can run it with the command

python run_lab.py slm_lab/spec/benchmark/sac/sac_lunar.json sac_lunar train

This should train within a few hours on a laptop.

Benchmark results and graphs can be found here