1. Decision Making Under Uncertainty CMSC 471 – Fall 2011 Class #23-24 – Thursday, November 17 / Tuesday, November 22 R&N, Chapters 15.1-15.2, 16.1-16.3, 17.1-17.3 material from Lise Getoor, Jean-Claude Latombe, and Daphne Koller 1

2. MODELING UNCERTAINTY OVER TIME 2

3. Temporal Probabilistic Agent environment agent ? sensors actuators t 1, t 2, t 3, … 3

4. Time and Uncertainty The world changes; we need to track and predict it Examples: diabetes management, traffic monitoring Basic idea: Copy state and evidence variables for each time step Model uncertainty in change over time, incorporating new observations as they arrive Reminiscent of situation calculus, but for stochastic domains X t – set of unobservable state variables at time t e.g., BloodSugar t, StomachContents t E t – set of evidence variables at time t e.g., MeasuredBloodSugar t, PulseRate t, FoodEaten t Assumes discrete time steps 4

5. States and Observations Process of change is viewed as series of snapshots, each describing the state of the world at a particular time Each time slice is represented by a set of random variables indexed by t: 1. the set of unobservable state variables X t 2. the set of observable evidence variables E t The observation at time t is E t = e t for some set of values e t The notation X a:b denotes the set of variables from X a to X b 5

6. Stationary Process/Markov Assumption Markov Assumption: X t depends on some previous X i s First-order Markov process: P(X t |X 0:t-1 ) = P(X t |X t-1 ) kth order: depends on previous k time steps Sensor Markov assumption: P(E t |X 0:t, E 0:t-1 ) = P(E t |X t ) That is: The agent’s observations depend only on the actual current state of the world Assume stationary process: Transition model P(X t |X t-1 ) and sensor model P(E t |X t ) are time-invariant (i.e., they are the same for all t) That is, the changes in the world state are governed by laws that do not themselves change over time 6

7. Complete Joint Distribution Given: Transition model: P(X t |X t-1 ) Sensor model: P(E t |X t ) Prior probability: P(X 0 ) Then we can specify the complete joint distribution of a sequence of states: 7

8. Example Rain t-1 Umbrella t-1 Rain t Umbrella t Rain t+1 Umbrella t+1 R t-1 P(R t |R t-1 ) TFTF 0.7 0.3 RtRt P(U t |R t ) TFTF 0.9 0.2 This should look a lot like a finite state automaton (since it is one...) 8

9. Inference Tasks Filtering or monitoring: P(X t |e 1,…,e t ) Compute the current belief state, given all evidence to date Prediction: P(X t+k |e 1,…,e t ) Compute the probability of a future state Smoothing: P(X k |e 1,…,e t ) Compute the probability of a past state (hindsight) Most likely explanation: arg max x1,..xt P(x 1,…,x t |e 1,…,e t ) Given a sequence of observations, find the sequence of states that is most likely to have generated those observations 9

10. Examples Filtering: What is the probability that it is raining today, given all of the umbrella observations up through today? Prediction: What is the probability that it will rain the day after tomorrow, given all of the umbrella observations up through today? Smoothing: What is the probability that it rained yesterday, given all of the umbrella observations through today? Most likely explanation: If the umbrella appeared the first three days but not on the fourth, what is the most likely weather sequence to produce these umbrella sightings? 10

11. Filtering We use recursive estimation to compute P(X t+1 | e 1:t+1 ) as a function of e t+1 and P(X t | e 1:t ) We can write this as follows: This leads to a recursive definition: f 1:t+1 =  FORWARD (f 1:t, e t+1 ) QUIZLET: What is α? 11

12. Filtering Example Rain t-1 Umbrella t-1 Rain t Umbrella t Rain t+1 Umbrella t+1 R t-1 P(R t |R t-1 ) TFTF 0.7 0.3 RtRt P(U t |R t ) TFTF 0.9 0.2 What is the probability of rain on Day 2, given a uniform prior of rain on Day 0, U 1 = true, and U 2 = true? 12

13. DECISION MAKING UNDER UNCERTAINTY

14. Decision Making Under Uncertainty Many environments have multiple possible outcomes Some of these outcomes may be good; others may be bad Some may be very likely; others unlikely What’s a poor agent to do?? 14

15. Non-Deterministic vs. Probabilistic Uncertainty ? bac {a,b,c}  decision that is best for worst case ? bac {a(p a ), b(p b ), c(p c )}  decision that maximizes expected utility value Non-deterministic modelProbabilistic model ~ Adversarial search 15

16. Expected Utility Random variable X with n values x 1,…,x n and distribution (p 1,…,p n ) E.g.: X is the state reached after doing an action A under uncertainty Function U of X E.g., U is the utility of a state The expected utility of A is EU[A] =  i=1,…,n p(x i |A)U(x i ) 16

17. s0s0 s3s3 s2s2 s1s1 A1 0.20.70.1 1005070 U(A1, S0) = 100 x 0.2 + 50 x 0.7 + 70 x 0.1 = 20 + 35 + 7 = 62 One State/One Action Example 17

18. s0s0 s3s3 s2s2 s1s1 A1 0.20.70.1 1005070 A2 s4s4 0.20.8 80 U (A1, S0) = 62 U (A2, S0) = 74 U (S0) = max a {U(a,S0)} = 74 One State/Two Actions Example 18

19. s0s0 s3s3 s2s2 s1s1 A1 0.20.70.1 1005070 A2 s4s4 0.20.8 80 U (A1, S0) = 62 – 5 = 57 U (A2, S0) = 74 – 25 = 49 U (S0) = max a {U(a, S0)} = 57 -5-25 Introducing Action Costs 19

20. MEU Principle A rational agent should choose the action that maximizes agent’s expected utility This is the basis of the field of decision theory The MEU principle provides a normative criterion for rational choice of action 20

21. Not quite… Must have a complete model of: Actions Utilities States Even if you have a complete model, decision making is computationally intractable In fact, a truly rational agent takes into account the utility of reasoning as well (bounded rationality) Nevertheless, great progress has been made in this area recently, and we are able to solve much more complex decision-theoretic problems than ever before 21

22. Axioms of Utility Theory Orderability (A>B)  (A<B)  (A~B) Transitivity (A>B)  (B>C)  (A>C) Continuity A>B>C   p [p,A; 1-p,C] ~ B Substitutability A~B  [p,A; 1-p,C]~[p,B; 1-p,C] Monotonicity A>B  (p≥q  [p,A; 1-p,B] >~ [q,A; 1-q,B]) Decomposability [p,A; 1-p, [q,B; 1-q, C]] ~ [p,A; (1-p)q, B; (1-p)(1-q), C] 22

23. Money Versus Utility Money <> Utility More money is better, but not always in a linear relationship to the amount of money Expected Monetary Value Risk-averse: U(L) < U(S EMV(L) ) Risk-seeking: U(L) > U(S EMV(L) ) Risk-neutral: U(L) = U(S EMV(L) ) 23

24. Value Function Provides a ranking of alternatives, but not a meaningful metric scale Also known as an “ordinal utility function” Sometimes, only relative judgments (value functions) are necessary At other times, absolute judgments (utility functions) are required 24

25. Multiattribute Utility Theory A given state may have multiple utilities...because of multiple evaluation criteria...because of multiple agents (interested parties) with different utility functions We will talk about this more later in the semester, when we discuss multi-agent systems and game theory 25

26. SEQUENTIAL DECISION MAKING UNDER UNCERTAINTY

27. Sequential Decision Making Finite Horizon Infinite Horizon 27

28. Simple Robot Navigation Problem In each state, the possible actions are U, D, R, and L 28

29. Probabilistic Transition Model In each state, the possible actions are U, D, R, and L The effect of U is as follows (transition model): With probability 0.8, the robot moves up one square (if the robot is already in the top row, then it does not move) 29

30. Probabilistic Transition Model In each state, the possible actions are U, D, R, and L The effect of U is as follows (transition model): With probability 0.8 the robot moves up one square (if the robot is already in the top row, then it does not move) With probability 0.1 the robot moves right one square (if the robot is already in the rightmost row, then it does not move) 30

31. Probabilistic Transition Model In each state, the possible actions are U, D, R, and L The effect of U is as follows (transition model): With probability 0.8 the robot moves up one square (if the robot is already in the top row, then it does not move) With probability 0.1 the robot moves right one square (if the robot is already in the rightmost row, then it does not move) With probability 0.1 the robot moves left one square (if the robot is already in the leftmost row, then it does not move) 31

32. Markov Property The transition properties depend only on the current state, not on the previous history (how that state was reached) 32

33. Sequence of Actions Planned sequence of actions: (U, R) 2 3 1 4321 [3,2] 33

34. Sequence of Actions Planned sequence of actions: (U, R) U is executed 2 3 1 4321 [3,2] [4,2][3,3][3,2] 34

35. Histories Planned sequence of actions: (U, R) U has been executed R is executed There are 9 possible sequences of states – called histories – and 6 possible final states for the robot! 4321 2 3 1 [3,2] [4,2][3,3][3,2] [3,3][3,2][4,1][4,2][4,3][3,1] 35

36. Probability of Reaching the Goal P([4,3] | (U,R).[3,2]) = P([4,3] | R.[3,3]) x P([3,3] | U.[3,2]) + P([4,3] | R.[4,2]) x P([4,2] | U.[3,2]) 2 3 1 4321 Note importance of Markov property in this derivation P([3,3] | U.[3,2]) = 0.8 P([4,2] | U.[3,2]) = 0.1 P([4,3] | R.[3,3]) = 0.8 P([4,3] | R.[4,2]) = 0.1 P([4,3] | (U,R).[3,2]) = 0.65 36

37. Utility Function [4,3] provides power supply [4,2] is a sand area from which the robot cannot escape +1 2 3 1 4321 37

38. Utility Function [4,3] provides power supply [4,2] is a sand area from which the robot cannot escape The robot needs to recharge its batteries +1 2 3 1 4321 38

39. Utility Function [4,3] provides power supply [4,2] is a sand area from which the robot cannot escape The robot needs to recharge its batteries [4,3] or [4,2] are terminal states +1 2 3 1 4321 39

40. Utility of a History [4,3] provides power supply [4,2] is a sand area from which the robot cannot escape The robot needs to recharge its batteries [4,3] or [4,2] are terminal states The utility of a history is defined by the utility of the last state (+1 or –1) minus n/25, where n is the number of moves +1 2 3 1 4321 40

41. Utility of an Action Sequence +1 Consider the action sequence (U,R) from [3,2] 2 3 1 4321 41

42. Utility of an Action Sequence +1 Consider the action sequence (U,R) from [3,2] A run produces one of 7 possible histories, each with some probability 2 3 1 4321 [3,2] [4,2][3,3][3,2] [3,3][3,2][4,1][4,2][4,3][3,1] 42

43. Utility of an Action Sequence +1 Consider the action sequence (U,R) from [3,2] A run produces one among 7 possible histories, each with some probability The utility of the sequence is the expected utility of the histories: U =  h U h P(h) 2 3 1 4321 [3,2] [4,2][3,3][3,2] [3,3][3,2][4,1][4,2][4,3][3,1] 43

44. Optimal Action Sequence +1 Consider the action sequence (U,R) from [3,2] A run produces one among 7 possible histories, each with some probability The utility of the sequence is the expected utility of the histories The optimal sequence is the one with maximal utility 2 3 1 4321 [3,2] [4,2][3,3][3,2] [3,3][3,2][4,1][4,2][4,3][3,1] 44

45. Optimal Action Sequence +1 Consider the action sequence (U,R) from [3,2] A run produces one among 7 possible histories, each with some probability The utility of the sequence is the expected utility of the histories The optimal sequence is the one with maximal utility But is the optimal action sequence what we want to compute? 2 3 1 4321 [3,2] [4,2][3,3][3,2] [3,3][3,2][4,1][4,2][4,3][3,1] only if the sequence is executed blindly! 45

46. Accessible or observable state Repeat:  s  sensed state  If s is terminal then exit  a  choose action (given s)  Perform a Reactive Agent Algorithm 46

47. Policy (Reactive/Closed-Loop Strategy) A policy  is a complete mapping from states to actions +1 2 3 1 4321 47

48. Repeat:  s  sensed state  If s is terminal then exit  a   (s)  Perform a Reactive Agent Algorithm 48

49. Optimal Policy +1 A policy  is a complete mapping from states to actions The optimal policy  * is the one that always yields a history (ending at a terminal state) with maximal expected utility 2 3 1 4321 Makes sense because of Markov property Note that [3,2] is a “dangerous” state that the optimal policy tries to avoid 49

50. Optimal Policy +1 A policy  is a complete mapping from states to actions The optimal policy  * is the one that always yields a history with maximal expected utility 2 3 1 4321 This problem is called a Markov Decision Problem (MDP) How to compute  *? 50

51. Additive Utility History H = (s 0,s 1,…,s n ) The utility of H is additive iff: U (s 0,s 1,…,s n ) = R (0) + U (s 1,…,s n ) =  R (i) Reward 51

52. Additive Utility History H = (s 0,s 1,…,s n ) The utility of H is additive iff: U (s 0,s 1,…,s n ) = R (0) + U (s 1,…,s n ) =  R (i) Robot navigation example: R (n) = +1 if s n = [4,3] R (n) = -1 if s n = [4,2] R (i) = -1/25 if i = 0, …, n-1 52

53. Principle of Max Expected Utility History H = (s 0,s 1,…,s n ) Utility of H: U (s 0,s 1,…,s n ) =  R (i)  First-step analysis  U (i) = R (i) + max a  k P (k | a.i) U (k)  *(i) = arg max a  k P (k | a.i) U (k) +1 53

54. Defining State Utility Problem: When making a decision, we only know the reward so far, and the possible actions We’ve defined utility retroactively (i.e., the utility of a history is obvious once we finish it) What is the utility of a particular state in the middle of decision making? Need to compute expected utility of possible future histories 54

55. Value Iteration Initialize the utility of each non-terminal state s i to U 0 (i) = 0 For t = 0, 1, 2, …, do: U t+1 (i)  R (i) + max a  k P (k | a.i) U t (k) +1 2 3 1 4321 55

56. Value Iteration Initialize the utility of each non-terminal state s i to U 0 (i) = 0 For t = 0, 1, 2, …, do: U t+1 (i)  R (i) + max a  k P (k | a.i) U t (k) U t ([3,1]) t 0302010 0.611 0.5 0 +1 2 3 1 4321 0.7050.6550.3880.611 0.762 0.8120.8680.918 0.660 Note the importance of terminal states and connectivity of the state-transition graph 56 EXERCISE: Verify that U([3,3]) = 0.918 in the limit

57. Policy Iteration Pick a policy  at random 57

58. Policy Iteration Pick a policy  at random Repeat: Compute the utility of each state for  U t+1 (i)  R (i) +  k P (k |  (i).i) U t (k) 58

59. Policy Iteration Pick a policy  at random Repeat: Compute the utility of each state for  U t+1 (i)  R (i) +  k P (k |  (i).i) U t (k) Compute the policy  ’ given these utilities  ’(i) = arg max a  k P (k | a.i) U (k) 59

60. Policy Iteration Pick a policy  at random Repeat: Compute the utility of each state for  U t+1 (i)  R (i) +  k P (k |  (i).i) U t (k) Compute the policy  ’ given these utilities  ’(i) = arg max a  k P(k | a.i) U (k) If  ’ =  then return  Or solve the set of linear equations: U (i) = R (i) +  k P (k |  (i).i) U (k) (often a sparse system) 60

61. Infinite Horizon +1 2 3 1 4321 In many problems, e.g., the robot navigation example, histories are potentially unbounded and the same state can be reached many times One trick: Use discounting to make an infinite horizon problem mathematically tractable What if the robot lives forever? 61