Lecture 18: Group Decision Making and Aggregation of ... · ME 597: Fall 2019 Lecture 18 5 / 32....
Transcript of Lecture 18: Group Decision Making and Aggregation of ... · ME 597: Fall 2019 Lecture 18 5 / 32....
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
Lecture 18: Group Decision Making andAggregation of Individual Preferences
Jitesh H. Panchal
ME 597: Decision Making for Engineering Systems Design
Design Engineering Lab @ Purdue (DELP)School of Mechanical Engineering
Purdue University, West Lafayette, INhttp://engineering.purdue.edu/delp
October 24, 2019
ME 597: Fall 2019 Lecture 18 1 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
Lecture Outline
1 Group Decision MakingArrow’s Impossibility Theorem
2 Aggregating Individuals’ Preferences1. Aggregation under Certainty2. Aggregation Under Uncertainty
3 Other ConsiderationsConsiderations of Equity
Keeney, R. L. and H. Raiffa (1993). Decisions with Multiple Objectives: Preferences and ValueTradeoffs. Cambridge, UK, Cambridge University Press. Chapter 10.
ME 597: Fall 2019 Lecture 18 2 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsArrow’s Impossibility Theorem
Focus of this Lecture: The Group Decision Problem
Decision Maker (in this lecture): Person who wishes to incorporate thefeelings, values, preferences, utilities of others into her own valueassessments. The Decision Maker’s preferences depend on the preferencesof the others.
How can a (supra) Decision Maker systematically incorporate the views ofothers into her own decision making framework?
ME 597: Fall 2019 Lecture 18 3 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsArrow’s Impossibility Theorem
Other Decision Problems (NOT covered in this Lecture)
1. Participatory Group Problem
A group of individuals collectively shares the responsibility of making aparticular decision.
In this problem, individuals i = 1, 2, . . . ,N are the members of thedecision-making group. The group as a whole must address the tradeoff andequity considerations to its members.
2. Interactive Decisions
When the decisions made by one decision maker depend on the decisionsmade by other decision makers.
ME 597: Fall 2019 Lecture 18 4 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsArrow’s Impossibility Theorem
Arrow’s Problem
Arrow’s Problem
Given the rankings of a set of alternatives by each individual in a decisionmaking group, what should the group ranking of these alternatives be?
Arrow’s outlook is individualistic and ordinalist. It is not permitted to ask anyfurther information to determine strength of preference.
ME 597: Fall 2019 Lecture 18 5 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsArrow’s Impossibility Theorem
Arrow’s Assumptions (1)
AssumptionsA: Complete Domain: There are at least two individuals in the group, at
least three alternatives, and a group ordering is specified for all possibleindividual members’ orderings (i.e., each individual is free to order thealternatives in any way).
B: Positive Association of Social and Individual Orderings:IF
the group ordering indicates A � B, andindividual’s paired comparison between alternatives other than A are notchanged, andeach individual’s paired comparison between A and any other alternativeeither remains unchanged or is modified in A’s favor
THEN the group ordering must imply A is still preferred to B. (An ordinalversion of monotonicity).
C: Independence of Irrelevant AlternativesD: Individual’s SovereigntyE: Non-dictatorship
ME 597: Fall 2019 Lecture 18 6 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsArrow’s Impossibility Theorem
Arrow’s Assumptions (2)
AssumptionsA: Complete DomainB: Positive Association of Social and Individual OrderingsC: Independence of Irrelevant Alternatives: If an alternative is eliminated
from consideration and the preference relations for the remainingalternatives remain invariant for all the group members, then the newgroup ordering for the remaining alternatives should be identical to theoriginal group ordering for these same alternatives. (Imposes transitivityin social order).
D: Individual’s Sovereignty : For each pair of alternatives A and B, there issome set of individual orderings such that the group prefers A to B.(Fairness: everyone’s preferences matter).
E: Non-dictatorship: There is no individual with the property that whenever(s)he prefers alternative A to B, the group also prefers A to B regardlessof the other individuals’ preferences.
ME 597: Fall 2019 Lecture 18 7 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsArrow’s Impossibility Theorem
Arrow’s Impossibility Theorem
Theorem (Arrow’s Impossibility Theorem)
Assumptions A, B, C, D, and E are inconsistent.
Interpretation: There is no procedure for combining individual rankings into agroup ranking that does not explicitly address the question of interpersonalcomparison of preferences!
ME 597: Fall 2019 Lecture 18 8 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsArrow’s Impossibility Theorem
Arrow’s Impossibility Theorem: Implications for Engineering Design
Implications for Engineering Design:
Aggregation of customers’ preferences
Group decision making
Multi-criteria decision making (MCDM) methods
ME 597: Fall 2019 Lecture 18 9 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Aggregating Individuals’ Preferences:From Ordinal to Cardinal Preferences
Assume that the consequences x of her decisions can be described in termsof attributes X1,X2, . . . ,XM .
Overall objective: Maximize the well-being of N specified individuals.
N lower-level objectives: Maximize the well-being of individual i , defined by vi
or ui .
ME 597: Fall 2019 Lecture 18 10 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
1. The Certainty Model
(Supra) Decision Maker’s value function: v over consequences x
Individuals’ value functions: v1, v2, . . . , vN
Goal: to find an appropriate model vD such that
v(x) = vD[v1(x), v2(x), . . . , vN(x)]
ME 597: Fall 2019 Lecture 18 11 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
1. The Certainty ModelAssumptions
Assumptions in this form:1 Decision maker’s preferences are entirely captured by vi ’s2 Individual i ’s preference structure is completely captured by vi , ∀i3 Decision Maker knows all the vi ’s
How realistic are these assumptions?
ME 597: Fall 2019 Lecture 18 12 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Complications
Situations where these assumptions may not be valid:
The (supra) Decision Maker may have an interest in the outcomes,beyond the implications on individuals that she represents.
The individuals themselves may not be capable of articulating theirpreferences to the Decision Maker.
Individuals may not be honest.
N may be large.
ME 597: Fall 2019 Lecture 18 13 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Restricting the Form of Group Value Functions
Assume that vD exists, such that
v(x) = vD[v1(x), v2(x), . . . , vN(x)]
Assumption 1: Preferential independence
The attributes {Vi ,Vj} are preferentially independent of their complement Vij ,for all i 6= j ,N ≥ 3.
Assumption 2: Ordinal positive association
Let certain alternatives A and B be equally preferred by the group. If A ismodified to A′ in such a manner that some individual i prefers A′ to A but allother individuals remain indifferent, then A′ is preferred to B by the group.
ME 597: Fall 2019 Lecture 18 14 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Assumption 1: Preferential Independence
Consider:
a three-person group N = 3
two consequences x′ and x′′
Suppose the third individual is indifferent between x′ and x′′.
The group preference of x′ vs. x′′ should NOT depend on whether the thirdindividual thinks x′ and x′′ are awful or delightful. The supra Decision Maker’sranking should only depend on the feelings of individuals 1 and 2.
ME 597: Fall 2019 Lecture 18 15 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Assumption 2: Ordinal positive association
v(x) = vD[v1(x), v2(x), . . . , vN(x)]
Ordinal positive association: Let certain alternatives A and B be equallypreferred by the group. If A is modified to A′ in such a manner that someindividual i prefers A′ to A but all other individuals remain indifferent, then A′
is preferred to B by the group.
Implication: vD is a positive monotonic function of each of its arguments.
ME 597: Fall 2019 Lecture 18 16 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Additive Group Value Functions – Theorem
Theorem
Given N ≥ 3, Assumption 1 (preferential independence) and Assumption 2(ordinal positive association) hold if and only if
v(x) =N∑
i=1
v∗i [vi (x)] =N∑
i=1
v+i (x)
where, for all i,1 vi is a value function for individual i scaled from 0 to 1.2 v∗i , a positive monotonic transformation of its argument vi , is the
Decision Maker’s value function over Vi , reflecting her interpersonalcomparison of the individuals preferences.
3 v+i defined as v∗i (vi ) is another value function for individual i consistently
scaled to reflect the Decision Maker’s interpersonal comparison ofpreference.
Does this result contradict Arrow’s impossibility theorem?
ME 597: Fall 2019 Lecture 18 17 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
2. The Uncertainty Model
Decision maker’s utility function: u over consequences x
Individuals’ utility functions: u1(), u2(), . . . , uN()
Goal: to find an appropriate model such that
u(x) = uD[u1(x), u2(x), . . . , uN(x)]
Assumptions in this form:1 Decision maker’s preferences are entirely captured by ui ’s2 Individual i ’s preference structure is completely captured by ui , ∀i3 Decision maker knows all the ui ’s
ME 597: Fall 2019 Lecture 18 18 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Example
Assume that the (supra) Decision Maker has a choice between:
Certainty consequence C
Lottery L: 〈A, 0.5,B〉
The Decision Maker ascertains that:
For individual i = 1 : uA1 = 1, uB
1 = 0, uC1 = 0.4
For individual i = 2 : uA2 = 0, uB
2 = 1, uC2 = 0.4
So, the Decision Maker has to choose between:
C, which will result in utilities (0.4, 0.4) for the two individuals
L, which will result in 〈(1, 0), 0.5, (0, 1)〉The decision maker must consider the tradeoffs among the impacts onindividuals i = 1 and i = 2. The decision maker must make theinterpersonal comparisons.
ME 597: Fall 2019 Lecture 18 19 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Additive Group Utility Functions
Assume the existence of uD such that
u(x) = uD[u1(x), u2(x), . . . , uN(x)]
In addition to this, the critical condition (Harsanyi, 1955) for
u(x) =N∑
i=1
λiui (x)
is:
Assumption H (Harsanyi)
If two alternatives, defined by probability distributions over the consequencesx, are indifferent to each individual, then they are indifferent for the group as awhole.
ME 597: Fall 2019 Lecture 18 20 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Additive Independence and Strategic Equivalence
Assumption H is equivalent to the following pair of assumptions:
Assumption 3: Additive Independence
The set of attributes U1,U2, . . . ,UN is additive independent.
Assumption 4: Strategic Equivalence
The Decision Maker’s conditional utility function u∗i over the attribute Ui
designating individual i ’s utility is strategically equivalent to individual i ’s utilityfunction ui . [Honesty Assumption]
ME 597: Fall 2019 Lecture 18 21 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
Additive Independence and Strategic Equivalence
Theorem
For N ≥ 2, Assumptions 3 (additive independence) and 4 (strategicequivalence) hold if and only if
u(x) =N∑
i=1
λiui (x)
where ui , i = 1, 2, . . . ,N, is a utility function for individual i scaled from 0 to 1,the λi ’s are positive scaling constants, and x is a consequence.
The Decision Maker’s interpersonal comparison of the individual’spreferences is required to assess the λi scaling factors.
ME 597: Fall 2019 Lecture 18 22 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
More General Group Utility Functions (1)
Assumption 5: Utility Independence
Attribute Ui , i = 1, 2, . . . ,N, is utility independent of the other attributes Ui .This implies
uD(u1, . . . , ui , . . . , uN) = gi (ui ) + fi (ui )u∗i (ui ) ∀i
Theorem
For N ≥ 2, Assumptions 4 (strategic equivalence) and 5 (utilityindependence) imply
u(x) = uD(u1, u2, . . . , uN)
u(x) =N∑
i=1
λiui (x) +N∑
i=1,j>i
λijui (x)uj (x) + · · ·+ λ12...Nu1(x)u2(x) . . . uN(x)
ME 597: Fall 2019 Lecture 18 23 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other Considerations
1. Aggregation under Certainty2. Aggregation Under Uncertainty
More General Group Utility Functions (2)
Assumption 1A: Preferential Independence
The attributes {Ui ,Uj} are preferentially independent of their complement Uij
for all i 6= j ,N ≥ 3
Theorem
For N ≥ 3, Assumptions 1 or 1A (preferential independence), 4 (strategicequivalence) and 5 (utility independence) imply
u(x) = uD(u1, u2, . . . , uN)
u(x) =N∑
i=1
λiui (x) + λ
N∑i=1,j>i
λiλjui (x)uj (x) + . . .
+ λN−1λ1λ2 . . . λNu1(x)u2(x) . . . uN(x)
where u and the ui ’s are scaled from 0 to 1, the λ’s are scaling constants,0 < λi < 1 for all i , and λ > −1.
ME 597: Fall 2019 Lecture 18 24 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsConsiderations of Equity
Prior vs. Posterior Equity
A Decision Maker is interested in the preferences of two individuals and herutility function is:
u(x) = uD(u1(x), u2(x)) = 0.5u1(x) + 0.5u2(x)
Let (0.4, 0.6) designate the alternative where u1 = 0.4 and u2 = 0.6
Consider the following alternatives:
Alternative A: (1, 0)
Alternative B: 〈(1, 0), (0, 1)〉Alternative C: 〈(1, 1), (0, 0)〉
Which of these alternatives should the Decision Maker choose?
ME 597: Fall 2019 Lecture 18 25 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsConsiderations of Equity
Prior vs. Posterior Equity
The additive utility function cannot differentiate among Alternatives A,B,C.
Alternatives B and C are a priori equitable, whereas alternative A is not.
Alternative C is more equitable in terms of posterior equity.
ME 597: Fall 2019 Lecture 18 26 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsConsiderations of Equity
Prior vs. Posterior Equity
Another utility function:
u(x) = 0.4u1(x) + 0.4u2(x) + 0.2u1(x)u2(x)
Expected utilities for alternatives A,B, and C are 0.4, 0.4, and 0.5respectively.
Provides a formal consideration of posterior equity.
ME 597: Fall 2019 Lecture 18 27 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsConsiderations of Equity
Pareto Optimality
Consider the following alternatives:
Alternative A: (1, 0)
Alternative B: 〈(1, 0), (0, 1)〉Alternative C: 〈(1, 1), (0, 0)〉Alternative D: (0.48, 0.48)
If u(x) = 0.5u1(x) + 0.5u2(x), then B � DIf u(x) = 0.4u1(x) + 0.4u2(x) + 0.2u1(x)u2(x), then D � B
The decision maker is faced with a tradeoff of amount of equity versusdegree to which the narrow interpretation of Pareto optimality is violated.
ME 597: Fall 2019 Lecture 18 28 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsConsiderations of Equity
Universal Agreement
Assumption: Universal Agreement
If all members of a group have the same utility function, then the group utilityfunction should be the common utility function.
In some cases, this assumption may not be reasonable! For example, thegroup may be less risk averse than the individuals.
Can you think of a reason?
ME 597: Fall 2019 Lecture 18 29 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsConsiderations of Equity
Assumption 5: Utility Independence
Utility independence may not be appropriate in certain conditions.
Consider two situations where all but one individual (i) is indifferent betweenthe consequences:
1 All individuals that are indifferent between alternatives are impacted in avery desirable manner.
2 All individuals that are indifferent between alternatives are impacted in avery undesirable manner.
If everyone’s utility is low, then because of the pressures among individuals,the Decision Maker may prefer consequences where the individual (i) alsoreceives an undesirable consequence rather than a desirable one.
ME 597: Fall 2019 Lecture 18 30 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsConsiderations of Equity
Summary
1 Group Decision MakingArrow’s Impossibility Theorem
2 Aggregating Individuals’ Preferences1. Aggregation under Certainty2. Aggregation Under Uncertainty
3 Other ConsiderationsConsiderations of Equity
ME 597: Fall 2019 Lecture 18 31 / 32
Group Decision MakingAggregating Individuals’ Preferences
Other ConsiderationsConsiderations of Equity
References
1 Keeney, R. L. and H. Raiffa (1993). Decisions with Multiple Objectives:Preferences and Value Tradeoffs. Cambridge, UK, Cambridge UniversityPress. Chapter 10.
2 Arrow, K. J. (1951). Social Choice and Individual Values, 1st edition,Wiley, New York, NY. Available: http://cowles.yale.edu/sites/default/files/files/pub/mon/m12-2-all.pdf
3 Hazelrigg, G. A. (1996). “The implication of Arrow’s ImpossibilityTheorem on Approaches to Optimal Engineering Design”, ASME Journalof Mechanical Design, Vol. 118, No. 2, pp. 161-164.
ME 597: Fall 2019 Lecture 18 32 / 32