It described about MDP, Monte-Carlo, Time-Difference, sarsa, and q-learning method, and used for Reinforcement Learning study group's lecture, where is belonged to Korea Artificial Intelligence Laboratory.

1. MDP, MC, TD
sarsa, q-learning
Uijin Jung

2. 𝑉𝜋 𝑠 = 𝐸 𝑟 𝑡+1
+ 𝛾 ∙ 𝑉𝜋 𝑆𝑡+1 𝑆𝑡 = 𝑠]
Bellman equation
𝑉𝜋 𝑠 : value function
𝑟 𝑡+1 ∶ reward
𝛾 : discount factor

3. S
A1 A2
Vπ(s) ↤ s
𝑞 𝜋(𝑠, 𝑎) ↤ 𝑎
Vπ s =
𝑎∈𝐴
𝜋 𝑎 𝑠 𝑞 𝜋(𝑠, 𝑎)
Vπ(s′) ↤ s’
𝑞 𝜋 𝑠, 𝑎 = 𝑅 𝑠
𝑎
+ 𝛾
𝑠′∈𝑆
Ρ𝑠𝑠′
𝑎
𝑉𝜋(𝑠′
)
𝑆1
′
𝑆2
′
𝑉𝜋 𝑠 = 𝐸 𝑟 𝑡+1 + 𝛾 ∙ 𝑉𝜋 𝑆𝑡+1 𝑆𝑡 = 𝑠]

4. Vπ s =
𝑎∈𝐴
𝜋 𝑎 𝑠 𝑞 𝜋(𝑠, 𝑎)
𝑞 𝜋 𝑠, 𝑎 = 𝑅 𝑠
𝑎
+ 𝛾
𝑠′∈𝑆
Ρ𝑠𝑠′
𝑎
𝑉𝜋(𝑠′
)
Vπ s =
𝑎∈𝐴
𝜋 𝑎 𝑠 (𝑅 𝑠
𝑎
+ 𝛾
𝑠′∈𝑆
Ρ𝑠𝑠′
𝑎
𝑉𝜋(𝑠′
))

5. Limitation of Bellman equation
• We must know MDP model perfectly.

6. Monte-Carlo
• Learn from the episodes
• Action selected by policy-> end of episode-> calculate value function
based on the episode
• If you iterate an episode several time, calculate value function by
averaging each rewards took from the state 𝑠𝑡.

7. =
1
𝑡
𝐺𝑡 +
𝑗=1
𝑡−1
𝐺𝑗
𝑗=1
𝑡−1
𝐺𝑗 = (𝑡 − 1)𝑉𝑡−1
=
1
𝑡
𝐺𝑡 + (𝑡 − 1)𝑉𝑡−1
𝑉𝑡 =
1
𝑡
𝑗=1
𝑡
𝐺𝑗 𝑉𝑡−1 =
1
𝑡 − 1
𝑗=1
𝑡−1
𝐺𝑗
(𝑡 − 1)𝑉𝑡−1=
𝑗=1
𝑡−1
𝐺𝑗

8. 𝑉𝑡 =
1
𝑡
𝐺𝑡 + (𝑡 − 1)𝑉𝑡−1
=
1
𝑡
𝐺𝑡 + 𝑉𝑡−1 −
1
𝑡
∙ 𝑉𝑡−1
= 𝑉𝑡−1 +
1
𝑡
𝐺𝑡 − 𝑉𝑡−1
𝑉𝑡−1 +
1
𝑡
𝐺𝑡 −
1
𝑡
∙ 𝑉𝑡−1
= 𝑉𝑡−1 + 𝑎 𝐺𝑡 − 𝑉𝑡−1

9. 𝑉𝑡 = 𝑉𝑡−1 + 𝑎 𝐺𝑡 − 𝑉𝑡−1
Reward predicted by
previous value function
Actual reward
Previous
value function
Value function
𝑉𝜋 𝑠𝑡 ← 𝑉𝜋(𝑠𝑡) + 𝑎 𝐺𝑡 − 𝑉(𝑠𝑡)
This is how we update value function
Using Monte-Carlo method

10. Limitations of Monte-Carlo
• Episode must end before we update the value function.
• It is difficult to train if the environment is endless or if the time to finish is too
long.

11. Time difference
• Let's change the Monte-Carlo method to a real time processing
method.
𝐺𝑡 = 𝑅𝑡+1 + 𝛾𝐺𝑡+1
𝑅𝑡+1 + 𝛾𝑉 (𝑆𝑡+1)
𝑉𝑡 = 𝑉𝑡 + 𝑎 𝑅𝑡+1 + 𝛾𝑉 (𝑆𝑡+1) − 𝑉𝑡
𝑉𝜋 𝑠𝑡 ← 𝑉𝜋(𝑠𝑡) + 𝑎 𝐺𝑡 − 𝑉(𝑠𝑡)

12. Time difference
• advantage
• You can update the value function in real time.
• disadvantage
• The reference 𝐺𝑡 is not true, it is derived
from the expected value : (we call this phenomenon, bootstrap)

13. Sarsa
• Algorithm to find optimal q value through TD method
𝑄 𝑆, 𝐴 ← 𝑄 𝑆, 𝐴 + 𝛼(𝑅 + 𝛾𝑄 𝑆′, 𝐴′ − 𝑄 𝑆, 𝐴 )
State, Action, Reward, next State, next Action

14. Sarsa pseudo code (https://stackoverflow.com/questions/32846262/q-learning-vs-sarsa-with-greedy-select)

15. Me
exploration
predict
Sarsa
-1 +1
• On-Policy
• There is a possibility
of being biased.

16. Q-learaning
• Algorithm to find optimal q value through TD method using off-policy.

17. Sarsa Q-learning
𝑄 𝑠, 𝑎 ← 𝑄 𝑠, 𝑎 + 𝛼(𝑅 + 𝛾𝑄 𝑠′, 𝑎′
− 𝑄 𝑠, 𝑎 ) 𝑄 𝑠, 𝑎 ← 𝑄(𝑠, 𝑎) + 𝛼(𝑅 + 𝛾 ∙ 𝑚𝑎𝑥 𝑎′ 𝑄 𝑠′
, 𝑎′
− 𝑄 𝑠, 𝑎 )

18. Q-learning pseudo code (https://stackoverflow.com/questions/32846262/q-learning-vs-sarsa-with-greedy-select)

19. Me
exploration
predict
Q-learning
-1 +1
v

20. The end