ECE 307 – Techniques for Engineering
Decisions
Value of Information
George Gross
Department of Electrical and Computer Engineering
University of Illinois at Urbana-Champaign
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
1
VALUE OF INFORMATION
While we cannot do away with uncertainty, there
is always a desire to attempt to reduce the
uncertainty about future outcomes
The reduction in uncertainty about future
outcomes may give us choices that improve
chances for a good outcome
We focus on the principles behind information
valuation
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
2
SIMPLE INVESTMENT EXAMPLE
market up (0.5)
k
c
to
s
k
ri s
gh
i
h
low-risk stock
1,700 – 200 = 1,500
flat
(0.3)
down
(0.2)
– 800 – 200
= – 1,000
up
(0.5)
1,200 – 200
= 1,000
flat
(0.3)
400 – 200 =
down
(0.2)
100 – 200 = – 100
300 – 200 =
savings account
100
200
500
stock investment entails a brokerage fee of $ 200
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
3
NOTION OF PERFECT INFORMATION
We say that an expert’s information is perfect if it
is always correct; we think of an expert as
essentially a clairvoyant
We can place a value on information in a decision
problem by measuring the expected value of info
( EVI )
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
4
NOTION OF PERFECT INFORMATION
We consider the role of perfect information in the
simple investment example
In this decision problem, the optimal policy is to
invest in high – risk stock since it has the highest
returns
Suppose an expert predicts that the market goes
up: this implies the investor still chooses the
high – risk stock investment and consequently
the perfect information of the expert appears to
have no value
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
5
NOTION OF PERFECT INFORMATION
On the other hand, suppose the expert predicts a
market decrease or a flat market: under this
information, the investor’s choice is the savings
account and the perfect information has value
because it leads to a changed outcome with improved results then would be the case otherwise
In worst case conditions: regardless of the
information, we take the same decision as
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
6
NOTION OF PERFECT INFORMATION
without the information and consequently
EVI = 0; the interpretation is that we are equally
well off without an expert
Cases in which we have information and in which
we change the optimal decision: these lead to
EVI > 0 since we make a decision with an improved outcome using the available information
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
7
EVI ASSESSMENT
It follows that the value of information is always
nonnegative, EVI ≥ 0
In fact, with perfect information, there is no
uncertainty and the expected value of perfect
information EVPI provides an upper bound for EVI
EVPI ≥ EVI
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
8
INVESTMENT EXAMPLE:
COMPUTATION OF EVPI
Absent any expert information, a value –
maximizing investor selects the high – risk stock
investment
The introduction of an expert or clairvoyant
brings in perfect information since there is perfect
knowledge of what the market will do before the
investor makes his decision and the investor’s
decision is based on this information
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
9
COMPUTATION OF EVPI
We use a decision tree approach to compute EVPI
by reversing the decision and uncertainty order:
we view the value of information in an a priori
sense and define
EVPI = E {decision with perfect information} –
E {decision without information}
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
10
COMPUTATION OF EVPI
For the investment problem,
EVPI = 1,000 – 580 = 420
We may view EVPI to represent the maximum
amount that the investor should be willing to pay
the expert for the perfect information resulting in
the improved outcome
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
11
COMPUTATION OF EVPI
ck
to
s
0
k
ris = 58
gh V
hi EM
low-risk stock
EMV = 540
up
(0.5)
flat
(0.3)
down
(0.2)
up
(0.5)
flat
(0.3)
down
(0.2)
savings account
1500
100
−1000
1000
200
−100
500
up
t
ke
ar .5)
m
(0
consult clairvoyant
EMV = 1000
market flat
(0.3)
m
ar
ke
(0 t d
.2 ) o w
n
high-risk stock
1500
low-risk stock
1000
savings account
high-risk stock
500
100
low-risk stock
200
savings account
500
high-risk stock
− 1000
low-risk stock
− 100
savings account
500
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
12
EXPECTED VALUE OF IMPERFECT
INFORMATION
In practice, we cannot obtain perfect information;
rather, the information is imperfect since there are
no clairvoyants
We evaluate the expected value of imperfect
information, EVII
For example we engage an economist to fore–
cast the future stock market trends; his forecasts
constitute imperfect information
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
13
EXPECTED VALUE OF IMPERFECT
INFORMATION
conditioning event
up
flat
down
“up”
0.8
0.15
0.2
“flat”
0.1
0.7
0.2
“down”
0.1
0.15
0.6
P{ “flat”| market is flat }
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
conditional probabilities
economist’s
prediction
true market state
14
EVII ASSESSMENT
We use the decision tree approach to compute
EVII
For the decision tree, we evaluate probabilities
using Bayes’ theorem
For the imperfect information, we define
with probability 0.5
⎧ up
market
⎪
M=
= ⎨ flat
with probability 0.3
performance ⎪
⎩down with probability 0.2
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
15
EVII ASSESSMENT
and the forecast r.v.
F =
⎧
⎪
⎪
⎪
⎪⎪
⎨
⎪
⎪
⎪
⎪
⎪⎩
"up"
"flat"
"down"
without the knowledge of the corresponding
probabilities of the two r.v.s
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
16
EVII COMPUTATION: INCOMPLETE
DECISION TREE
market activity
up
(0.5)
ck
to
s
k 80
ris = 5
gh V
hi EM low-risk
stock
EMV = 580
flat
(0.3)
down (0.2)
up
flat
(0.5)
(0.3)
down (0.2)
savings account
market activity
1500
100
high-risk stock
− 1000
1000 economist says
“market up”
200
− 100
500
low-risk stock
(?)
up
(?)
flat
(?)
down
(?)
up
(?)
flat
(?)
200
down (?)
− 100
savings account
consult the
economist
high-risk stock
economist says
“market flat”
low-risk stock
(?)
flat
(?)
up
(?)
flat
(?)
down (?)
(?)
savings account
high-risk stock
up
(?)
flat
(?)
down (?)
economist says
“market down”
low-risk stock
up
(?)
flat
(?)
down (?)
(?)
savings account
100
− 1000
1000
500
up
down (?)
economist’s
forecast
1500
1500
100
− 1000
1000
200
− 100
500
1500
100
− 1000
1000
200
− 100
500
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
17
COMPUTATION OF REVERSE
CONDITIONAL PROBABILITIES
P { M = down F = "up"} =
P {F = "up" M = down} P { M = down}
[ P {F = "up" M = down} P { M = down} +
P {F = "up"
P {F = "up"
M = down} P { M = up}
M = flat} P { M = flat}
+
]
0.2 ( 0.2 )
P {F = "up"} =
0.2 ( 0.2 ) + 0.5 ( 0.3 ) + 0.8 ( 0.5 )
we flip the probabilities in this way
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
18
EVII COMPUTATION: FLIPPING THE
“market down” (0.10)
“market up”
(0.15)
market flat
“market flat”
(0.70)
(0.3)
“market down” (0.15)
m
n
(0
.2)
“market flat”
(0.20)
what we have
(?)
market flat
(?)
market down (?)
market up
(?)
market flat
(?)
market up
(?)
market flat
(?)
?)
“market down” (0.60)
(
n”
ow
td
w
do
(0.20)
market up
market down (?)
ke
ar
“m
t
ke
ar
“market up”
“market flat”
(?)
actual market
performance
market down (?)
what we need
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
conditional probabilities with the conditioning on
the economists’ forecast
(0.10)
(?
)
“market flat”
p”
(0.80)
ar
ke
tu
“market up”
economist’s
forecast
“m
economist’s
forecast
m
ar
ke
t
up
(0
.5)
actual market
performance
conditional probabilities with the conditioning on
the actual market performance
PROBABILITY TREE
19
POSTERIOR PROBABILITIES
economist’s
prediction
market up
market flat
market down
“up”
0.8247
0.0928
0.0825
“flat”
0.1667
0.7000
0.1333
“down”
0.2325
0.2093
0.5581
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
conditional probabilities on
economists forecast
posterior probability for:
20
EVII COMPUTATION
We use conditional probabilities in the table to
build the posterior probabilities
For example
P {market up economist predicts "up"} = 0.8247
We then compute
P {F = "up"} = 0.485
P {F = "flat"} = 0.300
P {F = "down"} = 0.215
University of Illinois at Urbana-Champaign, All Rights Reserved.
© 2006 - 2009 George Gross,
21
EXPECTED VALUE OF IMPERFECT
INFORMATION
market activity
up (0.5)
1500
flat (0.3)
100
down (0.2)
−1000
up (0.5)
1000 economist says
flat (0.3)
“market up” (0.485)
200
down (0.2)
−100
ck
to
s
0
k
ris = 58
gh V
h i EM
low-risk
stock
EMV = 540
savings account
500
consult economist
EMV=822
economist says
“market flat” (0.300)
market activity
high-risk stock
up
(0.8247)
flat
(0.0928)
1500
low-risk stock
100
down (0.0825) − 1000
up
(0.8247)
1000
flat (0.0928)
200
EMV=835
down (0.0825)
EMV=1164
− 100
500
savings account
up
(0.1667)
high-risk stock
flat
(0.7000)
EMV=187
down (0.1333)
up
(0.1667)
low-risk stock
flat
(0.7000)
EMV=293
down (0.1333)
1500
100
− 1000
1000
200
− 100
savings account
economist says
“market down”(0.215)
up
(0.2325)
high-risk stock
flat
(0.2093)
EMV=188
down (0.5581)
up
(0.2325)
low-risk stock
flat
(0.2093)
EMV=219
down (0.5581)
500
1500
100
− 1000
savings account
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
1000
200
− 100
500
22
EVII COMPUTATION
The expected mean value for the decision made
with the economist information is
EMV |economist = 1,164(0.485) + 500(0.515) = 822
The expected mean value without information is
580
Consequently,
EVII = 822 – 580 = 242
This value represents the upper limit on the worth
of the economist’s forecast
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
23
EXAMPLE OF VALUE OF
INFORMATION
We consider the following decision tree
0.1
0.2
E
A
0.6
0.1
0.7
B
F
0.3
20
10
0
– 10
5
–1
with the events at E and F as independent
We perform a number of valuations of EVPI for
this simple decision problem
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
24
EVPI FOR F ONLY
EMV (A)
= 3.0
A
E
EMV (B)
= 3.2
B
F
0.1
20
0.2
10
0.6
0
0.1
−10
0.7
0.3
perfect
information EMV (info. about F) =
about F
4.4
EVPI (info. about F) =
EMV (info. about F) – EMV (B) =
4.4 – 3.2 =
1.2
0.1
5
0.2
E
−1
B=5
0.7
0.6
0.1
B
0.2
E
B = −1
0.3
10
0
− 10
5
0.1
F
20
20
10
0.6
0
0.1
−10
B
© 2006 - 2009 George Gross, University of Illinois at Urbana-Champaign, All Rights Reserved.
−1
25