Soft Actor Critic deploy mean path only

Hi, I'm wondering if there's a way to only deploy the mean path of the SAC agent after it's been trained? This is useful to create more stable actions after the network has been trained.

Should I extract the network weights manually, create a network, then extract an output path for the mean network?

Answer 1

[Emmanouil Tzorakoleftherakis]

Hello,

Please take a look at this option here which was added in R2021a to allow exactly the behavior you mentioned.

Hope this helps

[Tech Logg Ding]

Thank you for the reply. That setting works. I've also tried the roundabout way of extracting the actor neural network and modifying it to only have the mean path. Then I deploy the actor neural network into the simulation to act as a controller. Both method works!