Reinforcement Learning - SAC with hybrid action spaces

Current implementation of Soft Actor Critic algorithm (SAC) in Matlab only applies to problems with continuous action spaces.

I must tackle a problem with hybrid action spaces (some actions are continuous, some are discrete).

I know that SAC implementations with hybrid action spaces have been proposed in literature. Is it possible to adjust standard Matlab implementations in order to account for hybrid actions?

If this is not possible, is there any other Reinforcement Learning Algorithm in Matlab that can (and should) be used in order to deal with hybrid action spaces?

Answer 1

[Emmanouil Tzorakoleftherakis]

Hi,

Unfortunately, hybrid action spaces are not currently supported out of the box. One potential workaround is to use multi-agent training and train one agent for the discrete portion of the action space, and another for the continous part.

It would be helpful to know some more details about the application as well as some links to the papers you mentioned. I could use this information to discuss with the development team.

Thanks