#cliff-walking

Question 1: Single choice

In the Cliff World grid (start and goal at the two ends of the bottom row, the cells between them a cliff with reward −100, every other move −1), Q-learning and SARSA are both trained with ε-greedy exploration. Q-learning's greedy policy walks right along the cliff edge, while SARSA's goes one row higher. Why?

SARSA values its own exploring policy, which sometimes slips off the edge; Q-learning values greedy play1
SARSA uses a larger discount factor, so it places more value on the safer path2
Q-learning explores more, so it was the only one of the two to discover the shorter path3
SARSA cannot represent the path along the cliff edge in its tabular Q-function4

Timed test