Off-policy training (e.g. TD3, SAC) with a SharedModel

[jpmartin42]

I'm working on a project that requires training on time-series data with a custom architecture, and it may be too expensive to train multiple networks to learn the same time-series relationships; instead, actors and critics might work better as separate heads off of a single shared backbone. As the observation space I'm looking at is large and both sample efficiency and exploration are important, I'm looking at the possibility of off-policy agents like SAC or TD3.

Has anyone used actor and critic heads on the same network with skrl for off-policy algorithms specifically? I have not been able to find any examples of this online (all SharedModel examples I've found on the website use PPO, which uses Single Forward-Pass instead of Multiple Forward-Pass), and figuring out how to handle shared architectures for these training pipelines with the skrl library has been confusing.