1. Reinforcement Learning Introduction Passive Reinforcement Learning Temporal Difference Learning Active Reinforcement Learning Applications Summary

2. Introduction Supervised Learning: Example Class Reinforcement Learning: Situation Reward …

3. Examples Playing chess: Reward comes at end of game Ping-pong: Reward on each point scored Animals: Hunger and pain - negative reward food intake – positive reward

4. Framework: Agent in State Space 123 R=+5 6 R=  9 9 R=  6 10 8 R=+4 5 R=+3 4 7 e e s s s nw x/0.7 w n sw x/0.3 n s s Problem: What actions should an agent choose to maximize its rewards ? ne Example: XYZ-World Remark: no terminal states

5. 123 R=+5 6 R=  9 9 R=  6 10 8 R=+4 5 R=+3 4 7 e e s s s nw x/0.7 w n sw x/0.3 n s s ne XYZ-World: Discussion Problem 12 (3.3, 0.5) (3.2, -0.5) (0.6, -0.2) Bellman TD P P : 1-2-3-6-5-8-6-9-10-8-6-5-7-4-1-2-5-7-4-1. Explanation of discrepancies TD for P/Bellman: Most significant discrepancies in states 3 and 8; minor in state 10 P chooses worst successor of 8; should apply operator x instead P should apply w in state 6, but only does it only in 2/3 of the cases; which affects the utility of state 3 The low utility value of state 8 in TD seems to lower the utility value of state 10  only a minor discrepancy I tried hard but: any better explanations?

6. 1 0.145 2 0.72 3 0.58 R=+5 6 -8.27 R=  9 9 -5.98 R=  6 10 0.63 8 3.17 R=+4 5 3.63 R=+3 4 0.03 7 0.001 e e s s s nw x/0.7 w n sw x/0.3 n s s ne XYZ-World: Discussion Problem 12 Bellman Update  =0.2 Discussion on using Bellman Update for Problem 12: No convergence for  =1.0; utility values seem to run away! State 3 has utility 0.58 although it gives a reward of +5 due to the immediate penalty that follows; we were able to detect that. Did anybody run the algorithm for other  e.g. 0.4 or 0.6 values; if yes, did it converge to the same values? Speed of convergence seems to depend on the value of .

7. 123 R=+5 6 R=  9 9 R=  6 10 8 R=+4 5 R=+3 4 7 e e s s s nw x/0.7 w n sw x/0.3 n s s ne XYZ-World: Discussion Problem 12 (0.57, -0.65) (-0.50, 0.47) (-0.18, -0.12) TD TD inverse R P : 1-2-3-6-5-8-6-9-10-8-6-5-7-4-1-2-5-7-4-1. Other observations: The Bellman update did not converge for  =1 The Bellman update converged very fast for  =0.2 Did anybody try other values for  (e.g. 0.6)? The Bellman update suggest a utility value for 3.6 for state 5; what does this tell us about the optimal policy? E.g. is 1-2-5-7-4-1 optimal? TD reversed utility values quite neatly when reward were inversed; x become –x+  with  [-0.08,0.08]. (2.98, -2.99)

8. XYZ-World --- Other Considerations R(s) might be known in advance or has to be learnt. R(s) might be probabilistic or not R(s) might change over time --- agent has to adapt. Results of actions might be known in advance or have to be learnt; results of actions can be fixed, or may change over time.

9. Example: The ABC-World 123 R=+5 6 R=  9 9 R=  10 R=+9 8 R=  3 5 R=  4 4 R=  1 7 e e s s s nw x/0.9 w n ne x/0.1 n s n Problem: What actions should an agent choose to maximize its rewards ? ne Remark: no terminal states sw w To be used in Assignment3:

10. Basic Notations and Preview T(s,a,s’) denotes the probability of reaching s’ when using action a in state s; it describes the transition model A policy  specifies what action to take for every possible state s  S R(s) denotes the reward an agent receives in state s Utility-based agents learn an utility function of states uses it to select actions to maximize the expected outcome utility. Q-learning, on the other hand, learns the expected utility of taking a particular action a in a particular state s (Q-value of the pair (s,a) Finally, reflex agents learn a policy that maps directly from states to actions

11. Reinforcement Learning Introduction Passive Reinforcement Learning Temporal Difference Learning Active Reinforcement Learning Applications Summary

12. Passive Learning We assume the policy Π is fixed. In state s we always execute action Π(s) Rewards are given.

13. Figure 21.1a Terminal State 0.8 0.1 The Agent follows arrows with probability 0.8 and moves right or left of an arrow with probability 0.1; agents are reflected off walls and transferred back to the original state, if they move towards a wall. All non-terminal states have reward  0.04; The two terminal states have rewards +1 and  1.

14. Typical Trials (1,1) -0.04  (1,2) -0.04  (1,3) -0.04  (1,2) -0.04  (1,3) -0.04 …  (4,3) +1 Goal: Use rewards to learn the expected utility U Π (s)

15. Expected Utility U Π (s) = E [ Σ t=0 γ t R(s t ) | Π, S 0 = s ] Expected sum of rewards associated with s when the policy is followed.

16. Example (1,1) -0.04  (1,2) -0.04  (1,3) -0.04  (2,3) -0.04  (3,3) -0.04  (4,3) +1 Total reward for s(1,1) U π (s)=(-0.04 x 5) + 1 = 0.80

17. Direct Utility Estimation Convert the problem to a supervised learning problem: (1,1)  U = 0.72 (2,1)  U = 0.68 … Learn to map states to utilities. Problem: utilities are not independent of each other!

18. Bellman Equation Utility values obey the following equations: U  (s) = R(s) + γ  max a Σ s’ T ( s,a,s’ )  U  (s’) Can be solved using dynamic programming. Assumes knowledge of transition model T and reward R; the result is policy independent! Assume γ =1, for this lecture! Incorrect formula replaced on March 10, 2006

19. Example U(1,3) = 0.84 U(2,3) = 0.92 We hope to see that: U(1,3) = -0.04 + U(2,3) The value is 0.88. Current value is a bit low and we must increase it. (1,3)(2,3)

20. Bellman Update (Section 17.2 of textbook) If we apply the Bellman update indefinitely often, we obtain the utility values that are the solution for the Bellman equation!! U i+1 (s) = R(s) + γ max a ( Σ s’ (T(s, a,s’)  U i (s’))) Some Equations for the XYZ World: U i+1 (1) = 0+ γ*U i (2) U i+1 (5) = 3+ γ *max(U i (7),U i (8)) U i+1 (8) =  + γ *max(U i (6),0.3*U i (7) + 0.7*U i (9) ) Bellman Update:

21. Updating Estimations Based on Observations: New_Estimation = Old_Estimation*(1-  ) + Observed_Value*  New_Estimation= Old_Estimation + Observed_Difference*  Example: Measure the utility of a state s with current value being 2 and observed values are 3 and 3 and the learning rate is 0.2: Initial Utility Value:2 Utility Value after observing 3: 2x0.8 + 3x0.2=2.2 Utility Value after observing 3,3: 2.2x0.8 +3x0.2= 2.36

22. Reinforcement Learning Introduction Passive Reinforcement Learning Temporal Difference Learning Active Reinforcement Learning Applications Summary

23. Temporal Difference Learning Idea: Use observed transitions to adjust values in observed states so that the comply with the constraint equation, using the following update rule: U Π (s)  U Π (s) + α [ R(s) + γ U Π (s’) - U Π (s) ] α is the learning rate; γ discount rate Temporal difference equation. No model assumption --- T and R have not to be known.

24. Fig. 21.5a

25. Fig. 21.5b

26. TD-Q-Learning Goal: Measure the utility of using action a in state s, denoted by Q(a,s); the following update formula is used every time an agent reaches state s’ from s using actions a: Q(a,s)  Q(a,s) + α [ R(s) + γ *max a’ Q(a’,s’)  Q(a,s) ] α is the learning rate;  is the discount factor Variation of TD-Learning Not necessary to know transition model T!

27. Reinforcement Learning Introduction Passive Reinforcement Learning Temporal Difference Learning Active Reinforcement Learning Applications Summary

28. Active Reinforcement Learning Now we must decide what actions to take. Optimal policy: Choose action with highest utility value. Is that the right thing to do?

29. Active Reinforcement Learning No! Sometimes we may get stuck in suboptimal solutions. Exploration vs Exploitation Tradeoff Why is this important? The learned model is not the same as the true environment.

30. Explore vs Exploit Exploitation: Maximize its reward vs Exploration: Maximize long-term well being.

31. Simple Solution to the Exploitation/Exploration Problem Choose a random action once in k times Otherwise, choose the action with the highest expected utility (k-1 out of k times)

32. Another Solution --- Combining Exploration and Exploitation U + (s)  R(s) + γ  max a f(u,n) u=  s’ (T(s,a,s’)*U + (s’)); n=N(a,s) U + (s) : optimistic estimate of utility N(a,s): number of times action a has been tried. f(u,n): exploration function (idea: returns the value u, if n is large, and values larger than u as n decreases) Example: f(u,n):= if n>n avg then u else max(n/n avg *u, u avg ) n avg being the average number of operator applications. Idea f: Utility of states/actions that have not been explored much is increased artificially.

33. Reinforcement Learning Introduction Passive Reinforcement Learning Temporal Difference Learning Active Reinforcement Learning Applications Summary

34. Applications Game Playing Checker playing program by Arthur Samuel (IBM) Update rules: change weights by difference between current states and backed-up value generating full look-ahead tree

35. Reinforcement Learning Introduction Passive Reinforcement Learning Temporal Difference Learning Active Reinforcement Learning Applications Summary

36. Goal is to learn utility values of states and an optimal mapping from states to actions. Direct Utility Estimation ignores dependencies among states  we must follow Bellman Equations. Temporal difference updates values to match those of successor states. Active reinforcement learning learns the optimal mapping from states to actions.