Decision Theory Game Theory

Decision Theory &
Game Theory

Decision Making


choose the optimum strategy from all the
alternatives

Decision Making Situations
Perfect Information  Maximize –
Minimize
 Partial or Imperfect Information:





Decisions under Risk
Decisions under Uncertainty


Certainty

Risk

Uncertainty

Decisions under Risk


Based on criteria:





Expected value (of profit or loss)
Combined expected value and variance
Known aspiration level
Most likely occurrence of a future state


Expected Value Criterion
Expected Value includes the probability to
gain profit + the probability to suffer loss
 Mathematical Expectation (E(x))


Expected Value and Variance
Expected Value + Variance determine the
Risk Aversion Factor (K)
 Risk Aversion Factor indicates the
“importance” of an alternative
 The higher value of K, the more important
alternative


Aspiration-Level Criterion
First alternative generally treats as the
“best” alternative
 Decision is made in a short span of time
 Role of intuition



Most Likely Future Criterion
Simplification of probabilistic problem to
deterministic
 Generalization of what happen in the
(quite) similar problem


Probabilities for Under Risk
Prior Probabilities: the known probability
 Posterior Probabilities: modification of
prior probabilities … Conducting
experiment …


Under Risk: Decision Tree


Nodes:



Square:

decision point



Circle:

chance event (alternative)

Decision Tree: The Example
A company considers alternatives of 10year plan (partitioned in 2-year and 8-year
plans)
 Stage 1: At the beginning of the 2-year
plan





Build large plant: – 5M
High Demand (prob. = 0.75) yields 1 M/year
 Low Demand (prob. = 0.25) yields 0.3 M/year




Build small plant: – 1M
High Demand (prob. = 0.75) yields 0.25 M/year
 Low Demand (prob. = 0.25) yields 0.2 M/year


Decision Tree


Stage 2: at the beginning of the 8-year
plan



Expand the small plant: – 4.2 M
High Demand (prob. = 0.75) yields 0.9 M/year
 Low Demand (prob. = 0.25) yields 0.2 M/year




Do not Expand the small plant:
High Demand (prob. = 0.75) yields 0.25 M/year
 Low Demand (prob. = 0.25) yields 0.2 M/year


H. demand

1 M/y

0.75
2
Bu
Pl ild

an la
t
rg
e

5M

L. demand

Bu
P il
– lan d s
1M t m
a

0.25

ll

3


d
an
em /y
d
H. 25M
0.
0.
75

1

0.3 M/y

L.

de
ma
0.
nd

25

an
p
Ex

4
Do


no

tE

d

5

M
2

.
4

xp

H. demand
0.9 M/y
0.75
L. demand
0.2 M/y
0.25
H. demand
0.75

an

d

0.25 M/y


6
L. demand

0.2 M/y

0.25
0.2 M/y

Stage 1: 2 years

Stage 2 : 8 years

(2)= (10  0.75 1) + (10  0.25  0.3) = 8.25
(1)  (2)= 8.25 – 5 = 3.25 (build the large plant now)
(5)= (8  0.75 0.9) + (8  0.25  0.2) = 5.8
(6)= (8 0.75  0.25) + (8  0.25  0.2) =1.9
(4)  (5),(6) = 5.8 + 1.9 – 4.2 = 3.5
(3)  (4) = 3.5 + (2 0.75 0.25) + (10  0.25 0.20)
= 4.375
(1)  (3) = 4.375 – 1 = 3.375

Decision under Uncertainty
The Laplace Criterion  optimistic
 The Minimax (Maximin) Criterion  less
optimistic
 The Savage Criterion  “less
conservative”
 The Hurwicz Criterion  ranging from
optimistic to pessimistic


Decision under Uncertainty
Rows : possible action (ai)
 Columns: future states or condition (j)


Laplace Criterion



Based on the principle of insufficient reason
Unknown probabilities of the occurrence of

j


j = 1, 2, … n

Laplace Criterion

1 n

  v(ai,  j )
max
n
ai
 j 1

1
Probability of
n
j = 1, 2, … n

j

Laplace Criterion: Example
Minimize total cost of actions
Customer Category

Supplies
Level

a1
a2
a3
a4

1

2

3

4

5

10

18

25

8

7

8

23

21

18

12

21

30

22

19

15

Probability P{ =j} = ¼
j = 1, 2, 3, 4

• E{a1} = ¼ (5 + 10 + 18 + 25) = 14.5
• E{a2} = ¼ (8 + 7 + 8 + 23)

= 11.5 

• E{a3} = ¼ (21 + 18 + 12 + 21) = 18.0
• E{a4} = ¼ (30 + 22+ 19 + 15) = 21.5

Laplace Criterion: Example
Minimize total cost of actions
Customer Category

Supplies
Level

a1
a2
a3
a4

1

2

3

4

5

10

18

25

8

7

8

23

21

18

12

21

30

22

19

15

Minimax (Maximin) Criterion






Making the best out of
the worst
Minimax:

n

 v(ai ,  j )
max
min
ai
j
 j 1


Maximin:

n

 v(ai ,  j )
max
min
j
ai
 j 1


Minimax Criterion: Example
Minimax Strategy

a1
a2
a3
a4

1

2

3

4

5

10

18

25

8

7

8

23

21

18

12

21

30

22

19

15

Minimax Criterion: Example
Minimax Strategy

1

2

3

4

max{v(ai, j)}

a1

5

10

18

25

25

v(ai, j) a2

8

7

8

23

23

j

Minimax value

a3 21

18

12

21

21

a4 30

22

19

15

30

Maximin Criterion: Example
Maximin Strategy

a1
a2
a3
a4

1

2

3

4

5

10

18

25

8

7

8

23

21

18

12

21

30

22

19

15

Maximin Criterion: Example
Maximin Strategy

1

2

3

4

min{v(ai, j)}

a1

5

10

18

25

5

v(ai, j) a2

8

7

8

23

7

a3 21

18

12

21

12

j

Maximin value

a4 30

22

19

15

15

Savage Minimax Regret Criterion
Construct new loss or profit matrix
 v(ai, j) is replaced by r(ai, j) which is


defined by





r(ai, j)

max{v(ak, j)} – v(ai, j) if v is profit
ak
v(ai, j) – min {v(ak, j)} if v is loss
ak

Only the Minimax criterion can be applied
to r(ai, j)

Savage Minimax Regret Criterion
v is loss

v(ai, j)

1

2

3

4

a1

5

10

18

25

a2

8

7

8

23

a3

21

18

12

21

a4

30

22

19

15

Savage Minimax Regret Criterion
1

2

3

4

a1

5

10

13

25

a2

8

7

8

23

a3

21

18

12

21

a4

30

22

19

15

5

7

8

15

v is loss

v(ai, j)

min {v(ak, j)}
ak

Savage Minimax Regret Criterion
v is loss

r(ai, j)

1

2

3

4

a1

0

3

10

10

a2

3

0

0

8

a3

16

11

4

6

a4

25

15

11

0

Savage Minimax Regret Criterion
v is loss

a1

1

2

3

4

max r(ai, j)

0

3

10

10

10

j
Minimax value

r(ai, j) a2

3

0

0

8

8

a3

16

11

4

6

16

a4

25

15

11

0

25

Hurwicz Criterion


Balancing between extreme pessimism
and extreme optimism

Hurwicz Criterion
v(ai, j) : profit or gain
 max { max v(ai, j) + (1 – )min v(ai, j)
ai
j
j


Hurwicz Criterion
v(ai, j) : profit or gain
 Most

optimistic: max max{v(ai, j)}

 Most

pesimistic: max min{v(ai, j)}

ai j
weigth = 

ai j
weigth = 1 – 

where 0    1
  = ½ (reasonable)


Hurwicz Criterion
v(ai, j) : cost
 Most

optimistic: min min{v(ai, j)}

 Most

pesimistic: min max{v(ai, j)}

ai j
weigth = 

ai j
weigth = 1 – 

where 0    1
  = ½ (reasonable)


Hurwicz Criterion
v(ai, j) : cost
 min { min v(ai, j) + (1 – )max v(ai, j)
ai
j
j


Minimize total cost of actions
min v(ai, j) max v(ai, j)
1 2 3 4
j
j

a1
a2
a3
a4

5

10

18

25

5

25

8

7

8

23

7

23

21

18

12

21

12

21

30

22

19

15

15

30

=½

=½
min v(ai, j)
j

a1
a2

max v(ai, j)  min v(ai,
j) j
j

+

1–
max v(ai, j)
j

5

25

0.5  5 + 0.5  25 =15

7

23

0.5  7 + 0.5  23 =15

a3

12

21

a4

15

30

0.5  12 + 0.5  21
=16.5
0.5  15 + 0.5  30
=22.5

min
ai

Reading Assignment #5


Game Theory
2

person zero-sum games
 Mixed Strategies

Reading Assignment #6


Introduction to Queueing Theory

The End


This is the end of Chapter 8A