How can a policy neural network control a continuous agent? What's the KL divergence? What's the main idea behind TRPO? How is the KL divergence used in TRPO? What's the main benefit of PPO? How does PPO achieve good sample efficiency?