Towards Heisenberg limit without critical slowing down via quantum reinforcement learning