Vous êtes sur la page 1sur 30

DADSS

Multiattribute Utility Theory

Administrative Details
Homework Assignment 6 is due Monday. (slightly shorter)
Homework Assignment 7 posted tonight will be due Monday, March 24th
Project meetings

Complex Choices
Multi-Objective Decision Making (MODM)
Multiple, Competing Goals
Maximize Tax Revenue
Minimize Tax Rate
Maximize Compliance

Municipal Fiscal Policy

Multi-Attribute Decision/Utility Theory (MAUT)


Diverse Characteristics Aggregated to Single Value Measure
Price
Safety
Performance

Buying a Car

MAUT
MODM is typically dealt with using techniques such as Goal

Programming and the Analytic Hierarchy Process


We will not cover MODM

MAUT involves an extension of our existing techniques to incorporate

trade-offs
Trade-offs are expressions of preference

Attributes
Basic Party Problem
Everything is reduced to dollars
MAUT Party Problem
U(x) is the utility of x
U(Party) = U(Cost) + U(Fun) + U(Attendance)
Multiple factors (attributes) influence our preferences for various outcomes
U(Party) is essentially a utility measure with multiple factors
MAUT Key: Can the attributes be traded-off?
Could the party still be good if the Cost goes up, provided that Fun and Attendance also go up?
THINK: Additive vs Multiplicative Value

Choice Strategies
Non-Compensatory Strategies
Methods for choosing alternatives that do not allow for trade-offs between attributes
Compensatory Strategies
Decision maker can give up/get some of one attribute in exchange for another attribute or attributes
to increase total value

Non-Compensatory Strategies
Similar to simple heuristics
Easy to apply
Prone to biases and can be misleading

Lexicographic
Elimination-by-Aspects
Conjunctive
Disjunctive
Combinations

Lexicographic Rule
Rank the attributes in order of importance
Rank all options on the most important attribute
Break ties by using next most important attribute
Pick option with best value on most important attribute
Problem: Only considers a single attribute when other attributes may also

be important

Elimination-by-Aspects Rule
Rank the attributes in order of importance
Establish a minimum acceptable level on each attribute
Eliminate alternatives that are unacceptable with respect to the most

important attribute
Continue elimination with next most important attributes until only one
alternative remains
Problem: Difficult to determine attribute importance independently of

acceptability thresholds; acceptability can be arbitrary

Conjunctive Rule
Establish a minimum or maximum acceptable level on each attribute
Alternatives found to be unacceptable on any attribute are eliminated
If no alternatives remain, weaken the acceptability level; if two or more

remain, strengthen the acceptability level

Problem: Same issues as optimism/pessimism

Disjunctive Rule
Establish a minimum or maximum excellence level on each attribute
Alternatives found to be excellent on any attribute are accepted
If no alternatives remain, weaken the excellence level; if two or more

remain, strengthen the excellence level

Problem: Same issues as optimism/pessimism

Compensatory Strategies
Additive Value Functions
V Alternative 1v1 Attribute 1 2 v2 Attribute 2 3v3 Attribute 3
n

V Alternative i vi Attribute i
i 1

Two questions:
How are the weights (i) determined?
How are the individual attribute values (vi) determined?

Additive Value Functions


Now, trade-offs are allowed:

V Alternative 1v1 Attribute 1 2 v2 Attribute 2


=

Weights: How important is the cars price relative to its performance?


Values: How much more valuable is a 0-100km time of 5.5 seconds over 6.5 seconds?

Additive Value Functions in 5 Steps


Step 1: Check the validity of the additive value model
Step 2: Assess the single attribute value functions (vi)
Step 3: Assess the scaling constants (i)
Step 4: Compute the overall value of each alternative
Step 5: Perform sensitivity analysis
Example: Buying a Car
3 Choices
3 Attributes (price, performance, braking)

Step 1: Validity Check


With a linear AVF (additive value function) what types of preference are

ruled out?

I only like high performance cars if theyre black


I would never work in a large city unless it was for an investment bank

For choices using an AVF to be rational, they must not only satisfy

completeness and transitivity, but we will also require independence


Independence is an additional requirement for being able to use an additive value

function to represent preferences

Independence
Preferential Independence
Your preferences for more or less of one attribute are not influenced by the levels of

other attributes
Choosing among job offers: Salary levels in NYC vs Erie
Difference Independence

The degree of preference among one attribute cannot be affected by another

attribute
If you prefer NYC twice as much as Erie at a salary of $50K, then you must
maintain that same degree of preference (2x) at a salary of $80K
Trade-Off Independence
How you trade-off any two attributes cannot be affected by a third

Step 2: Value Functions


Determine the ranking of alternatives for a specific attribute (e.g., price)
Let the worst alternative be 0 and the best one be 100
Determine intermediate values based on their relative similarity
Fit (or interpolate) a value function
Challenges? Checking all possible combinations/differences for

inconsistencies and intransitivities

Assessing Single Attribute Value Functions


Considering new cars costing $20-$50K
Set end points
v($50,000) = 0
v($20,000) = 100
Where should $35K be?
Suppose v($42K) = 50
Ask: is v($42K) = 0.5 v($20K)?
Ask: is v($42K) v($50K) = v($20K)
v($42K)?
Elicit other points to complete curve

Step 3: Comparing Attributes


Which of the attributes matters most? Are they equally influential?
Each individual attribute has now been measured on a scale of 0 100,

but is v1 = 25 the same as v2 = 25?


We need an exchange rate to allow us to compare different attributes on
the same scale?
i should reflect the relative importance of the ranges of outcomes on the
different attributes

Method 1: Swing Weights


Consider an alternative having the worst level of each attribute
Suppose you could increase one attribute to its best level
Which one? Which would be second?
Assign a value of 100 to the most important attribute and values to the remaining

attributes to reflect their relative importance

Method 1: Swing Weights


Suppose A1 is most important, A2 is 1/2 as important, and A3 is 1/3 as

important
If A1 = 100, then A2 = 50 and A3 = 33.3
Normalize to sum to 1
100 + 50 + 33.3 = 183.3
Weights by swing weight method:
A1 = 100/183.3 = 0.545
A2 = 50/183.3 = 0.273
A3 = 33.3/183.3 = 0.182

Method 2: Direct Trade-Off


The direct trade-off method requires continuity (money yes; city no)
Swing weight method doesnt

Infer values from comparative judgments


Tends to work best when money is one of the attributes and can be used as the

medium of exchange

Suppose X and Y are two attributes


Let + and reflect the best and worst levels for each attribute

Method 2: Direct Trade-Off


Consider 2 alternatives:
A1 = {X+, Y-}
A2 = { ___, Y+}
Y+ is preferable to Y What value makes the decision maker indifferent?
The blank is typically monetary

Indifference implies that V(A1) = V(A2)


Since we know the attribute values (from Step 2), we can easily solve for

the weights

Method 2: Direct Trade-Off


We have 3 unknowns:
1, 2, and 3

Suppose A1 = {$16K, 9 s.} and A2 = {___, 6 s.}


What price produces indifference between A1 and A2?
Or phrased differently, how much would you pay to improve the acceleration from 9 sec to 6 sec?
Suppose its $21,000

Method 2: Direct Trade-Off


Now compare another 2 attributes
A3 = {$16K, 160 ft.}
A4 = {___, 150 ft.}
How much more would you pay to move from 160 ft to 150 ft?
Suppose youd pay $18K
Now we have 3 equations for our 3 unknowns
A1 = A2, A3 = A4, and i = 1

Method 2: Direct Trade-Off


Use values elicited in Step 2
Equation 1:

Equation 2:

1v1 16K 2 v2 9sec 1v1 21K 2 v2 6sec


1001 0 2 681 100 2
2 0.321

1v1 16K 3v3 160 ft 1v1 18K 3v3 150 ft


1001 0 3 901 100 3

3 0.101

Equation 3:

1 2 3 1
1 0.321 0.101 1

1 0.704, 2 0.225, 3 0.070

Step 4: Compute Overall Value


V Car 1v1 Price 2 v2 Acceleration 3v3 Braking

V Car 1 1v1 $16K 2 v2 9 sec. 3v3 160 ft.

Step 5: Sensitivity Analysis


Compare rankings of alternatives using swing weights with those

produced by direct trade-off

Procedural Invariance
Consistency/Biases in Elicitation?

Compare results with equal weighting


Do the weights matter? Robustness?
Is there a clear winner?
Suppose you are unsure about your preferences (at least expressed numerically)

Step 5: Sensitivity Analysis


Vary weights given to less important attributes
Suppose the decision maker is only confident of the importance ranking, but not

necessarily the values


Does it matter if A2 is 1/2 as important as A1?

3/4?

1/4?

3/8?

Examine more trade-offs using the direct trade-off method


Search thoroughly for any intransitivities or inconsistencies in preferences

Challenges to MAUT
GIGO (garbage in, garbage out)
These steps are all just meaningless calculations unless the elicitation is done properly
Also, if Step 1 (validity of the AVF) isnt satisfied, the methodology is unreliable
Cognitive biases?

Vous aimerez peut-être aussi