#reward-design

Question 1: Select all that apply

You design a Markov decision process for the Atari game Pong, in which the agent moves a paddle to return a ball past the opponent. Which choices are sound? Select all that apply.

Reward: +1 when the ball passes the opponent's paddle, −1 when it passes yours, 0 otherwise1
Actions: move the paddle up, move it down, or leave it where it is2
State: the last few screen frames stacked, so the ball's direction can be seen3
State: only the current screen frame, since it shows where everything is4
The transition probabilities must be known to the agent, or Pong is not an MDP5

Timed test