Comment by igorkraw

Could you highlight the difference between this and training a permutation invariant or equivariant policy network using standard supervised or RL methods? Assuming I also have a way of having an invariant/equivariant loss function