Sunteți pe pagina 1din 26

Judgment and Decision Making in Information Systems

Decision Making, Sensitivity


Analysis, and the Value of
Information
Yuval Shahar, M.D., Ph.D.

Personal Decision Making:


The Party Problem
Joseph K. invites his friends to a party, but needs
to decide on the location:
Outdoors (O) on the grass (completely open)
On the Porch (P) (covered above, open on sides)
Inside (I) the living room

If the weather is sunny (S), outdoors is best,


followed by Porch and Indoors; if it rains (R), the
living room is best, followed by Porch and
Outdoors

The Party Problem:


An Ordered Preferences List
O-S (best)
P-S
I-R
I-S
P-R
O-R (worst)

Evaluating The Prospects


O-S (best)

P-S

I-R
I-S
P-R
O-R (worst)

1
0
0.95
0.05
0.67

B
W

B
W

0.33
0.57

0.43
0.32

0.68
0

B
B
B

The Party Problem: Adding Probabilities


S 0.4

R 0.6
S 0.4

P
R 0.6
I

S 0.4

R 0.6

The Party Problem: Substituting Prospects


S 0.4

R 0.6
S 0.4

P
R 0.6
I

S 0.4

R 0.6

1
0
0
1
0.95

B
W
B
W
B

0.05

0.32

0.68
0.57

0.43

0.67

0.33

The Party Problem: Simplifying The


Relationship to Preference Probabilities
0.4

0.6
0.57

B
W

P
0.43
I

0.63

0.37

The Party Problem: Using the Choice Rule


0.4

0.6
0.57

B
W

P
0.43
I

0.63

0.37

<=
W

Expected Preference-Probability Values


E-value =0.40
O
E-value= 0.57
P
I

E-value= 0.63

S 0.4

Preference Probability

R 0.6

S 0.4

0.95

R 0.6

0.32

S 0.4

0.57

R 0.6

0.67

Computation With Decision Trees:


Making the Decision
Decision trees are folded back to the top
most (leftmost, or initial) decision
Computation is performed by averaging
expected utility (in this case, preference
probability) recursively over tree branches
from right to left (bottom up), maximizing
utility for every decision made and assuming
that this is the expected utility for the subtree
that follows the computed decision

The Value of Information (VI)


We often need to decide what would be the next best
piece of information to gather (e.g., within a diagnostic
process); that is, what is the best next question to ask
(e.g., what would be the result of a urine culture?);
Alternatively, we might need to decide how much, if at
all, is a paticular additional piece of information worth
The Value of Information (VI) of feature f is the
marginal expected utility of an optimal decision made
knowing f, compared to making it without knowing f
The net value of information (NVI) of f = VI(f)-cost(f)
NVI is highly useful when deciding what would be the
next information item, if any, to investigate

Computing the Value of


Information: Requirements
Decision Makers are often faced with the option of
getting additional information for a certain cost
To assess the value of additional information to
the decision maker, we need a continuous, real
valued utility measure, such as money
Thus, we need to map preference probabilities to
a monetary scale or an equivalent one (such as
time)

Introducing Monetary Values:


Bringing in The Wizard
We ask the decision maker for her
willingness to pay for a change from any
state to any other state, assuming we have a
wizard who can perform the change
instantaneously
We can thus create a 1:1 correspondence
between preference probabilities and $
values

The Party Problem:


Adding Monetary Values
Prospect
Preference Probability
O-S (best) 1
100
P-S 0.95
90
I-R 0.67
50
I-S 0.57
40
P-R 0.32
20
O-R (worst)
0
0

$ Value

Utility

Joseph K.s Utility Curve


4
1
U( x) (1
3
2

1
0.5

x
50

0
$0 $34 $50K
$100K
Money
Note: Once we know JKs utilility curve, we can compute his
certain equivalent for ANY deal, e.g., a (<$100K, 50%; $0, 50%>
deal, which happens to be $34, using graphical or other methods;
or the $ certain equivalent of any outcome (e.g., Outdoor location)

The Value of Clairvoyance


To compute the value of information, we
assume a clairvoyant who knows with
certainty all outcomes and always tells the truth
However, clairvoyance has a price!
Thus, the question is, How Much should we
pay the clairvoyant for her forecast (e.g., for
telling Joseph K. if it will rain or not)

Computing the Value of Clairvoyance


We build a decision tree comparing the optimal decision
without clairvoyance to the optimal decision given that
the clairvoyant prophesies any of the possible outcomes
(in this case, with 100% accuracy)
We need to deduct the cost of clairvoyance from the $
value of all outcomes and role the decision tree back as
we did before, to determine the expected value of the
decision given clairvoyance (and paying for it!)
We compute e-values using u-values (which represent
the decision makers utility for each $ value)
At the end (root node) we convert u-values to $ values
We then know whether the clairvoyance is worth its cost

Computing The $ Value of Clairvoyance

No clairvoyance
U=0.63 ( $46)

U= 0.40 $26
O
U= 0.57 $40
P
U= 0.63 $46
I
U= 0.63 $46
$85
U= 0.92 $85

S
Buy clairvoyance for $15
U =0.67 ( $51)

0.4

O
$75
P
I $25

0.6

U= 0.50 $35

$-15
$5
$35

$value U value
100 1
0.6 R
0
0
0.4 S
90
0.95
0.6 R
91
0.32
0.4 S
40
0.57
0.6 R 50
0.67
$value $value-$cost
1 S
100
85
0 R
0
-15
1 S
90
75
0 R
91
5
1 S
40
25
0 R
50
35
100
85
0 S
0
-15
1 R
90
75
0 S
91
5
0 S
40
25
1 R
50
35
0.4 S

The Value of Partial Clairvoyance


Assume we have a rain detector which, when the
weather is going to be sunny or rainy, predicts Rain
or Sun 80% of the time, correspondingly; is it
worth $10 to use it?
In order to compute the value of an uncertain piece
of information, we first need to calculate the
probability for rain or sun, given each
corresponding prediction of the detector
This involves reversing the tree in which the
detector information is usually given, using Bayes
theorem

Computing The Value


of Partial Clairvoyance (I):
Representing the Detectors Profile
S 0.8
S 0.4

0.32

R 0.2

0.08

S 0.2

0.12

R 0.8

0.48

R 0.6

Computing The Value of


Partial Clairvoyance (II):
Calculating Accuracy by Reversing the Tree
S 0.727
0.44 S

R 0.273
S 0.143

0.32
0.08
0.12

R 0.56
R 0.857

0.48

Computing The Value of


Partial Clairvoyance (III):
Calculating the Optimal Decision
With and Without Clairvoyance
To actually calculate the value of uncertain information, we compare
the expected utility of the best decision without the information to the
expected utility of the decision with the information, as we did before
We use the distribution of real world states given the semiclairvoyants prediction (using our accuracy calculation)
In this particular case (80% accuracy), the value without detector is
U = 0.627, or $45.83; the value with partial information is U = 0.615,
or $44.65
Thus, for Joseph K., this particular detector is not worth the money

Sensitivity Analysis
The main insights into a decision are often given
by an analysis of the influence of the given
probabilities and utilities on the final decision and
its value
We thus get a sense as to how sensitive the
expected utility and optimal decisions are to each
parameter, and can focus our attention (and
perhaps, further elicitation efforts) on the most
important aspects of the problem

The Party Problem: Sensitivity Analysis

O
U=p

$value

U value

100

1-p

90

0.950

1-p

20

0.323

40

0.568

1-p

50

0.667

P
U = 0.323+0.627p

I
U = 0.667-0.099p

Sensitivity Analysis of the Party Problem


The value of information

100
90

Free clairvoyance:

0.667+0.333p
0.8
0.6

I: 0.667+0.0.099p
40

0.4
0.2

20

P: 0.323+0.628p
O:p
0.2

0.4

0.47

0.6

0.8

Tornado Diagrams
We can calculate the lower and upper bounds on the expected
utility E(U) of the optimal decision for the full range of changes in
each parameter Xi while leaving all other parameters at their
nominal (default) value
We get a quick sense of the potential influence of each parameter
on the expected utility
But:
X4
It ignores the change distribution
It ignores the dependencies between parameters

X12
X3
X5
X7
X8
0

S-ar putea să vă placă și