You don’t remove any information. You just use the right model. The action distribution is a delta function for your own actions. You still train on how the world is affected by or reacts to your actions

→ View original post on X — @nandodf