Reinforcement learning algorithm that stabilizes updates with clipped objectives, used in RLHF.