Bisimulation-based Approximate Lifted Inference