CS7642_Week4

Question 1

What three ways can we modify reward functions?

Answer 1

1. Scale Multiplying by positive scalar constant ‘c’
2. Shift Adding a positive (or negative) scalar constant c
3. Potential-based (change-in-state-based bonuses using psi function that leave things unchanged)

Question 2

What is the danger of reward shaping?

Answer 2

Positive feedback loops. If

Question 3

Initializing Q-values randomly is okay? (True/False)

Answer 3

Dr. Isbell says False. (zero is a better choice in his opinion). This is the result of the fact that where we start (i.e. initialize) the Q-value does matter, because the agent is initially going to want to be in those states, which may slow down learning.