A nice feature would be to able to manually specify which variables are known to be decoupled or if there is rotational / translational symmetry in the problem and then learn a relative model.
@stanleybak nice. let me draw up a formulation of how
1) how this information is encoded and how it's used to transform the model learning problem
2) how this affect possibly things like observables choices
A nice feature would be to able to manually specify which variables are known to be decoupled or if there is rotational / translational symmetry in the problem and then learn a relative model.