Denys88/rl_games

question about central value

Opened this issue · 1 comments

@Denys88 thank you for the nice repo.
I noticed you use a central value network when using asymmetric actor-critic. Could you please elaborate on what the central value net is exactly doing? Is this just the critic net?

yes it is a critic network but with additional inputs. For example policy can have only a few parameters as obs but critic can have the whole world state because we don't use critic during inference.