Exploiting Symmetries in Reinforcement Learning