Questions

1. Why can you not use DQN for the TORCS problem?
2. We used the Xavier weights initializer for the neural network weights. What other weight initializers are you aware of, and how well will the trained agent perform with them?
3. Why is the abs() function used in the reward function, and why is it used for the last two terms but not for the first term?
4. How can you ensure smoother driving than what was observed in the video?
5. Why is a replay buffer used in DDPG but not in PPO?