In reinforcement learning, which component is responsible for updating the policy?