Decision Making Under Uncertainty

VARIOUS STEPS

UtilityUtility-Based Agent
sensors

?
environment agent actuators

NonNon-deterministic vs. Probabilistic Uncertainty
? ?

a {a,b,c}

b

c

a

b

c

{a(pa),b(pb),c(pc)} decision that maximizes expected utility value Probabilistic model

decision that is best for worst case Non-deterministic model ~ Adversarial search

xn and distribution (p1..pn) E. U is the utility of a state The expected utility of A is EU[A] = 7i=1.Expected Utility Random variable X with n values x1.g.: X is the state reached after doing an action A under uncertainty Function U of X E.g. . .n p(xi|A)U(xi) . .

2 + 50 x 0.7 + 70 x 0.1 = 20 + 35 + 7 = 62 A1 s1 0.One State/One Action Example s0 EU(A1) = 100 x 0.7 50 s3 0.1 70 .2 100 s2 0.

EU(A2)} = 74 A2 A1 s1 0.One State/Two Actions Example s0 ‡ EU(AI) = 62 ‡ EU(A2) = 74 ‡ EU(S0) = max{EU(A1).8 80 .2 100 s2 0.1 70 s4 0.2 50 s3 0.7 0.

1 70 s4 0.7 0.2 100 s2 0.EU(A2)} = 57 A2 A1 -5 -25 s1 0.Introducing Action Costs s0 ‡ EU(A1) = 62 ± 5 = 57 ‡ EU(A2) = 74 ± 25 = 49 ‡ EU(S0) = max{EU(A1).2 50 s3 0.8 80 .

MEU Principle rational agent should choose the action that maximizes agent s expected utility this is the basis of the field of decision theory normative criterion for rational choice of action .

and we are able to solve much more complex decision theoretic problems than ever before .Not quite Must have complete model of:    Actions Utilities States Even if you have a complete model. will be 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.

17) . 16) Sequential decision making (ch.We ll look at Decision Theoretic Planning   Simple decision making (ch.

Decision Networks Extend BNs to handle actions and utilities Also called Influence diagrams Make use of BN inference Can do Value of Information calculations .

. as in BNs Decision nodes: actions that decision maker can take Utility/value nodes: the utility of the outcome state. Chance nodes: random variables.Decision Networks cont.

R&N example .

Prenatal Testing Example .

0 rain happiness U(~umb.0 P(~umb|~take)=1.4 Take Umbrella umbrella P(umb|take) = 1. rain) = -100 U(umb. ~rain) = 100 U(~umb.~rain) = 0 U(umb.Umbrella Network take/don¶t take P(rain) = 0.rain) = -25 .

using BN inference Calculate the resulting utility for action return the action with the highest utility .Evaluating Decision Networks Set the evidence variables for current state For each possible value of the decision node:    Set decision node to that value Calculate the posterior probability of the parent nodes of the utility node.

Umbrella Network take/don¶t take P(rain) = 0.rain | take) happiness U(~umb.4 Take Umbrella umbrella P(umb|take) = 1. rain) = -100 U(umb.~rain) = 0 0 0 1 1 U(umb.0 rain #1 umb rain 0 1 0 1 P(umb. ~rain) = 100 U(~umb.0 P(~umb|~take)=1.rain) = -25 #2: EU(take) .

rain) = -25 #2: EU(~take) .0 rain #1 umb rain 0 1 0 1 P(umb.0 P(~umb|~take)=1.4 Take Umbrella umbrella P(umb|take) = 1.Umbrella Network take/don¶t take P(rain) = 0. ~rain) = 100 U(~umb. rain) = -100 U(umb.rain | ~take) happiness U(~umb.~rain) = 0 0 0 1 1 U(umb.

E d! ma | ) A § U(Re sult ( A))P(Re sult ( A) | E .Value of Information (VOI) suppose agent s current knowledge is E. E )  EU( E | E ) . E d o( A)) i i i the value of information for E is: VOI (E d! ) § P( e k k | E )EU( E ek | e k . The value of the current best action E is EU( E | E ) ! ma A § U(Re sult ( A))P(Re sult ( A) | E i i i o ( A)) the value of the new best action (after new evidence E is obtained): EU( E dE .

2 0.8 0.~rain) = 0 0 1 1 U(umb.4 Take Umbrella umbrella P(umb|take) = 1.3 0.0 rain forecast happiness R 0 F 0 1 0 1 P(F|R) 0.0 P(~umb|~take)=1. ~rain) = 100 U(~umb.Umbrella Network take/don¶t take P(rain) = 0. rain) = -100 U(umb.7 U(~umb.rain) = -25 .

VOI VOI(forecast)= P(rainy)EU(Erainy) + P(~rainy)EU(E~rainy) EU(E) .

~rain) = 100 U(~umb.2 0.8 0.~rain) = 0 0 1 1 U(umb. rain) = -100 U(umb.3 0.4 Umbrella Network take/don¶t take F 0 0 1 1 R 0 1 0 1 P(R|F) 0.8 0.7 P(rain) = 0.4 Take Umbrella umbrella P(umb|take) = 1.3 0.P(F=rainy) = 0.0 rain forecast happiness R 0 F 0 1 0 1 P(F|R) 0.rain) = -25 .7 U(~umb.0 P(~umb|~take)=1.2 0.

rain | take. ~rainy) #1: EU(take|rainy) #3: EU(take|~rainy) umb 0 0 1 1 rain 0 1 0 1 P(umb. rainy) umb 0 0 1 1 rain 0 1 0 1 P(umb.umb 0 0 1 1 rain 0 1 0 1 P(umb.rain | take.rain |~take.rain | ~take. ~rainy) #2: EU(~take|rainy) #4: EU(~take|~rainy) . rainy) umb 0 0 1 1 rain 0 1 0 1 P(umb.

2 0.8 0.rain) = -25 .4 Umbrella Network take/don¶t take F 0 0 1 1 R 0 1 0 1 P(R|F) 0.3 0.8 0. rain) = -100 U(umb.0 rain forecast happiness R 0 F 0 1 0 1 P(F|R) 0. ~rain) = 100 U(~umb.7 P(rain) = 0.P(F=rainy) = 0.3 0.~rain) = 0 0 1 1 U(umb.0 P(~umb|~take)=1.4 Take Umbrella umbrella P(umb|take) = 1.2 0.7 U(~umb.

VOI VOI(forecast)= P(rainy)EU(Erainy) + P(~rainy)EU(E~rainy) EU(E) .

Sequential Decision Making Finite Horizon Infinite Horizon .

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

Probabilistic Transition Model In each state. then it does not move) .8 the robot moves up one square (if the robot is already in the top row. and L The effect of U is as follows (transition model): With probability 0. R. D. the possible actions are U.

8 the robot moves up one square (if the robot is already in the top row.1 the robot moves right one square (if the robot is already in the rightmost row. and L The effect of U is as follows (transition model): With probability 0. the possible actions are U. D.Probabilistic Transition Model In each state. then it does not move) . R. then it does not move) With probability 0.

then it does not move) . D.Probabilistic Transition Model In each state. the possible actions are U.8 the robot moves up one square (if the robot is already in the top row. and L The effect of U is as follows (transition model): With probability 0.1 the robot moves left one square (if the robot is already in the leftmost 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. R. then it does not move) With probability 0.

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

R) .2] 3 2 1 1 2 3 4 Planned sequence of actions: (U.Sequence of Actions [3.

2] .3] [4.Sequence of Actions 3 2 1 1 2 3 4 Planned sequence of actions: (U.2] [3. R) U is executed [3.2] [3.

Histories 3 2 1 1 2 [3.2] [3.3] [4.2] [3.1] [4.2] [3.1] [3.2] [4.3] 3 4 Planned sequence of actions: (U.3] [4.2] [3. 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! .

3] | R.R).2] | U.3] | R.3]) x P([3.8 P([4.2]) = 0.2]) P([4.1 P([4.[3.3] | (U.3] | U.Probability of Reaching the Goal Note importance of Markov property 2 in this derivation 1 1 2 3 4 3 P([4.[4.[4.[3.2]) x P([4.8 P([3.[3.3] | (U.3] | R.2]) + P([4.[3.[3.1 P([4.2] | U.[3.65 .2]) = P([4.3] | U.2]) = 0.2]) = 0.3] | R.2]) = 0.3]) = 0.[3.R).[3.

Utility Function 3 2 1 1 2 3 4 +1 -1 [4.2] is a sand area from which the robot cannot escape .3] provides power supply [4.

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

2] is a sand area from which the robot cannot escape The robot needs to recharge its batteries [4.Utility Function 3 2 1 1 2 3 4 +1 -1 [4.3] or [4.2] are terminal states .3] provides power supply [4.

3] provides power supply [4.Utility of a History 3 2 1 1 2 3 4 +1 -1 [4.2] are terminal states The utility of a history is defined by the utility of the last state (+1 or 1) minus n/25.2] is a sand area from which the robot cannot escape The robot needs to recharge its batteries [4. where n is the number of moves .3] or [4.

2] .R) from [3.Utility of an Action Sequence 3 2 1 1 2 3 4 +1 -1 Consider the action sequence (U.

2] [3.R) from [3.3] [4.1] [3. each with some probability .1] [4.2] Consider the action sequence (U.2] [4.Utility of an Action Sequence 3 2 1 1 2 3 4 [3.3] [4.2] [3.2] [3.3] +1 -1 [3.2] A run produces one among 7 possible histories.

R) from [3.2] [3.2] [3.1] [3.3] +1 -1 [3.Utility of an Action Sequence 3 2 1 1 2 3 4 [3.2] A run produces one among 7 possible histories.2] [4.2] [3. each with some probability The utility of the sequence is the expected utility of the histories: U = 7hUh P(h) .2] Consider the action sequence (U.3] [4.3] [4.1] [4.

2] [3.1] [3.2] A run produces one among 7 possible histories.3] +1 -1 [3.2] [3.2] [3.3] [4.R) from [3.3] [4.2] [4.1] [4.Optimal Action Sequence 3 2 1 1 2 3 4 [3.2] Consider the action sequence (U. 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] A run produces onethe sequence is histories.R) from [3.Optimal Action Sequence 3 2 1 1 2 3 4 [3.3] [4.1] [3.2] [3.2] [4.2] [3.3] +1 -1 [3.2] Consider the action sequence (U.1] [4. each with some only if among 7 possible executed blindly! 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? .3] [4.2] [3.

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

Policy (Reactive/Closed-Loop Strategy) (Reactive/Closed3 2 1 1 2 3 4 +1 -1 A policy 4 is a complete mapping from states to actions .

Reactive Agent Algorithm Repeat:  s sensed state  If s is terminal then exit  a 4(s)  Perform a .

2] is a dangerous mapping from states to actions state that the optimal policy The optimal policy 4* is the one that always yields a tries to avoid history (ending at a terminal state) with maximal expected utility Makes sense because of Markov property .Optimal Policy 3 2 1 1 2 3 4 +1 -1 A policy 4 is a completeNote that [3.

Optimal Policy 3 2 1 1 2 3 4 +1 -1 This problem states to A policy 4 is a complete mapping from is called a actions Markov Decision Problem (MDP) The optimal policy 4* is the one that always yields a history with maximal expected utility How to compute 4*? .

Additive Utility
History H = (s0,s1, ,sn) The utility of H is additive iff: U(s ,s , ,s ) = R(0) + U(s , ,s ) = 7R(i)
0 1 n 1 n

Reward

Additive Utility
History H = (s0,s1, ,sn) The utility of H is additive iff: U(s ,s , ,s ) = R(0) + U(s , ,s )
0 1 n 1 n

=

7R(i)

Robot navigation example:  R(n) = +1 if sn = [4,3]  R(n) = -1 if sn = [4,2]  R(i) = -1/25 if i = 0, , n-1

Principle of Max Expected Utility
History H = (s0,s1, ,sn) Utility of H: U(s ,s ,
0 1

,sn) =

7R(i)
+1 -1

First-step analysis

U(i) = R(i) + maxa 7kP(k | a.i) U(k)
4*(i) = arg maxa 7kP(k | a.i) U(k)

do: Ut+1(i) R(i) + maxa 7kP(k | a.i) Ut(k) 3 2 1 1 2 3 4 +1 -1 .Value Iteration Initialize the utility of each non-terminal state si to U (i) = 0 0 For t = 0. . 2. 1.

660 +1 -1 0.Value Iteration Note the importance of terminal state and Initialize the utility of each non-terminalstates si to connectivity of the U0(i) = 0 state-transition graph For t = 0.5 0 30 t 2 1 0.388 1 2 3 4 0 10 20 .705 0. 2.i) Ut(k) Ut([3.611 0. do: Ut+1(i) R(i) + maxa 7kP(k | a.762 0.868 0.812 0.1]) 0.918 3 0. .655 0.611 0. 1.

Policy Iteration Pick a policy 4 at random .

Policy Iteration Pick a policy 4 at random Repeat:  Compute the utility of each state for 4 Ut+1(i) R(i) + 7kP(k | 4(i).i) Ut(k) .

i) U(k) k .i) Ut(k) Compute the policy 4 given these utilities 4 (i) = arg maxa  7 P(k | a.Policy Iteration Pick a policy 4 at random Repeat:  Compute the utility of each state for 4 Ut+1(i) R(i) + 7kP(k | 4(i).

i) Ut(k) Compute the policy 4 given these utilities 4 (i) = arg maxa Or 7 P(k solve the set of linear equations: | a.i) U(k) k   If 4 = 4 then return 4a sparse system) (often U(i) = R(i) + 7kP(k | 4(i).i) U(k) .Policy Iteration Pick a policy 4 at random Repeat:  Compute the utility of each state for 4 Ut+1(i) R(i) + 7kP(k | 4(i).

Example: Tracking a Target The robot must keep the target in view The target s trajectory is not known in advance robot target .

Example: Tracking a Target .

Example: Tracking a Target .

histories are What if the robot lives forever? potentially unbounded and the same One trick: state can be reached many times 3 2 1 1 2 3 4 +1 -1 Use discounting to make infinite Horizon problem mathematically tractible .Infinite Horizon In many problems.g.. e. the robot navigation example.

and actually does not make sense (is not rational) .POMDP (Partially Observable Markov Decision Problem) A sensing operation returns multiple states. with a probability distribution Choosing the action that maximizes the expected utility of this state distribution assuming state utilities computed as above is not good enough.

Example: Target Tracking There is uncertainty in the robot s and target s positions. and this uncertainty grows with further motion There is a risk that the target escape behind the corner requiring the robot to move appropriately But there is a positioning landmark nearby. Should the robot tries to reduce position uncertainty? .

Summary Decision making under uncertainty Utility function Optimal policy Maximal expected utility Value iteration Policy iteration .

Sign up to vote on this title
UsefulNot useful