instead of a tightly coupled DDPG / param noise actor interface, create a separate parameter-space noise module that owns its own copy of the actor network (and then has a tf_copy_weights method that creates the copy op or something).
Then the relevant flags for param noise adaptation should be only in this file.
instead of a tightly coupled DDPG / param noise actor interface, create a separate parameter-space noise module that owns its own copy of the actor network (and then has a tf_copy_weights method that creates the copy op or something).
Then the relevant flags for param noise adaptation should be only in this file.