31
Arthur CHARPENTIER - Multi-attribute Utility & Copulas Multi-Attribute Utility & Copulas (based on Ali E. Abbas contributions) A. Charpentier (Université de Rennes 1 & UQàM) Université de Rennes 1 Workshop, April 2016. http://freakonometrics.hypotheses.org @freakonometrics 1

Multiattribute utility copula

Embed Size (px)

Citation preview

Page 1: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Multi-Attribute Utility & Copulas

(based on Ali E. Abbas contributions)

A. Charpentier (Université de Rennes 1 & UQàM)

Université de Rennes 1 Workshop, April 2016.

http://freakonometrics.hypotheses.org

@freakonometrics 1

Page 2: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Olivier’s Talk, part 2, on Independence & Additivity

@freakonometrics 2

Page 3: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Olivier’s Talk, part 2, on Utility Independence

see also Keeney & Raiffa (1976)

@freakonometrics 3

Page 4: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Olivier’s Talk, part 2, on Mutual Utility Independence

@freakonometrics 4

Page 5: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Olivier’s Talk, part 2, on Additive Utility Independence

@freakonometrics 5

Page 6: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Olivier’s Talk, part 2, on Additive Utility Independence

@freakonometrics 6

Page 7: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Olivier’s Talk, part 2, on Mutual Utility Independence

@freakonometrics 7

Page 8: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Olivier’s Talk, part 2, on Mutual Utility Independence

@freakonometrics 8

Page 9: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

What are we looking for?

See Sklar (1959) for cumulative distribution function for random vector X ∈ Rn,

F (x1, · · · , xn) = C[F1(x), · · · , Fn(xn)]

where F (x) = P[X ≤ x] and Fi(xi) = P[Xi ≤ xi].

@freakonometrics 9

Page 10: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

What are we looking for?

@freakonometrics 10

Page 11: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Historical Perspective

When everything else remains constant whichdo you prefer

(x1, y1) or (x2, y2)

X can be consumptionY can be health(remaining life time expectancy)

Matheson & Howard (1968) : use a deterministic real-valued function V : Rd → Rand then use a utility function over the value function,

U(x) = U(x1, · · · , xd) = u(V (x1, · · · , xd)),

e.g. U(x) = u(x1 + · · ·+ xd) or u(min{x1, · · · , xd}).

@freakonometrics 11

Page 12: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Historical Perspective

See Matheson & Abbas (2005), e.g. V (x, y) = xyη,

see also Sheldon’s acoustic sweet spot or peanut butter/jelly sandwich preferencefunction

@freakonometrics 12

Page 13: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Historical Perspective

Alternative approach: assesss utilities over individual attributes, and combinetime into a functional form

Keeney & Raiffa (1976) : use some utility independence assumption

Mutual utility independence : U(x, y) = kxux(x) + kyuy(y) + kxyux(x)uy(y)where kxy = 1− kx − kyAdditive and Product forms

U(x, y) = kxux(x) + kyuy(y) with kx − ky = 1

U(x, y) = kxyux(x)uy(y)

Utility Independence is an intersting property, but it might be a simplifying one.

@freakonometrics 13

Page 14: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

How to Construct Multi-Attribute Utility Functions

From Abbas & Howard (2005), in dimension d = 2,

U(x, y) ∈ [0, 1] (normalization )

U(x, y) = U(x, y) = 0 (attribute dominance condition)

@freakonometrics 14

Page 15: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

How to Construct Multi-Attribute Utility Functions

Non-decreasing with arguments:

• given y, x1 < x2 implies (x1, y) � (x2, y)

• given x, y1 < y2 implies (x, y1) � (x, y2)

U(x, y) = ux(x) and U(x, y) = uy(y)

@freakonometrics 15

Page 16: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Conditional Utility

We can define conditional utility

Uy|x(y|x) = U(x, y)ux(x)

@freakonometrics 16

Page 17: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Conditional Utility

Bayes’ Rule for Attribute Dominance Utility

U(x, y) = ux(x) · Uy|x(y|x) = uy(y) · Ux|y(x|y).

@freakonometrics 17

Page 18: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Copula Structures for Attribute Dominance Utility

With two attributes, consider U(x, y) = C(ux(x), uy(y))

Since copulas are related to probability measures, function C are 2-increasing.

C is the cumulative didstribution function of some U , and

P(U ∈ [a, b]) ≥ 0

implies positive mixed partial derivatives, ∂2C(u, v)∂u∂v

≥ 0 (weaker condition exist).

Not a necessary condition for attribute dominance utility theory...

@freakonometrics 18

Page 19: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Understanding the Two Attribute Framework

C might be on a normalized domain, with a normalized range C : [0, 1]2 → [0, 1],with C(0, 0) = 0 and C(1, 1) = 1.

From Keeney & Raiffa (1976)

X independent of Y (preferences for lotteries over x do not depend on y)

U(x, y) = k2(y)U(x, y0) + d2(y)

Y independent of X (preferences for lotteries over y do not depend on x)

U(x, y) = k1(x)U(x0, y) + d1(x)

C should satisfy some marginal property: there are u0 and v0 such that

C(u0, v) = αu0v + βu0 and C(u, v0) = αv0u+ βv0 .

Margins are non decreasing, ∂C(u, v)∂u

> 0 and ∂C(u, v)∂v

> 0.

@freakonometrics 19

Page 20: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Understanding the Two Attribute Framework

Abbas & Howard (2005) defined some Class 1 Multiattribute Utility Copulas suchthat

C(1, v) = αu0v + βu0 and C(u, 1) = αv0u+ βv0 .

Proposition Any multi-attribute utility function U(x1, · · · , xn) that iscontinuous, bounded and strictly increasing in each argument can be expressed interms of its marginal utility functions u1(x1), · · · , un(xn) and some class 1multiattribute utility copula

U(x1, · · · , xn) = C[u1(x1), · · · , un(xn)].

@freakonometrics 20

Page 21: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Archimedean Copulas

On probability cumulative distribution functions

C(u1, · · · , ud) = φ−1(φ(u1) + · · ·+ φ(ud)) = φ−1

n∑j=1

φ(uj)

with φ : [0, 1]→ R+ an additive generator, or with ψ = φ−1 completely monotone

C(u1, · · · , ud) = ψ(ψ−1(u1) + · · ·+ ψ−1(ud)) = ψ

n∑j=1

ψ−1(uj)

One can define some mutiplicative generator, λ(t) = e−φ(t)

C(u1, · · · , ud) = λ−1(λ(u1)× · · · × λ(ud)) = λ−1

n∏j=1

λ−1(uj)

E.g. φ(t) = − log(t) or λ(t) = t, independent copula, C = Π = C⊥

@freakonometrics 21

Page 22: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Archimedean Utility Copulas

In the context of utility functions,

C(v1, · · · , vd) = αψ−1

(d∏i=1

ψ(γi + [1− γi]vi))

+ [1− α]

with γi ∈ [0, 1], and such that a =[ψ−1

(d∏i=1

ψ(γi))]−1

.

ψ continuous strictly increasing, ψ(0) = 0 and ψ(1) = 1.

E.g. ψ(t) = t, then

C(v1, v2) = α[γ1 + (1− γ1)v1][γ2 + (1− γ2)v2] + (1− α)

i.e. multiplicative form of mutual independence.

@freakonometrics 22

Page 23: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Alternative to this Two Attribute Framework

By relaxing the condition of ‘attribute dominance’, Abbas & Howard (2005)defined some Class 2 Multiattribute Utility Copulas such that

C(0, v) = αu0v + βu0 and C(u, 0) = αv0u+ βv0 .

Define a multiattribute utility copula C as a multivariate function of d variablessatisfying C : [0, 1]d → [0, 1], with C(0) = 0, C(1) = 1, the following marginalproperty

C(0, · · · , 0, vi, 0, · · · , 0) = αivi + βi, with αi > 0

and with ∂C(v)/∂vi > 0

@freakonometrics 23

Page 24: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Alternative to this Two Attribute Framework

To define some Class 2 Archimedean utility copulas, let h be continuous on [0, d],strictly increasing, with h(0) = 1 and h(1)d ≤ h(d). Then set

C(v1, · · · , vd) =h−1

(∏dj=1 h(ωjvj)

)h−1

(∏dj=1 h(ωj)

) , with 0 ≤ ωj ≤ 1.

E.g. h(t) = et, then C(U1(x1), · · · , Ud(xd)) = ω̃1U1(x1) + · · ·+ ω̃dUd(xd), whereω̃j = ωj/[ω1 + · · ·+ ωd], i.e. additive form of utility independence.

@freakonometrics 24

Page 25: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

One-Switch Utility Independence

Introduced in Abbas & Bell (2011)

Consider two attributes x and y, utility function U(x, y).

x is one-switch independent of y if and only if the ordering of any two lotteriesover x switches at most once as y increases

Proposition x is one-switch independent of y if and only if

U(x, y) = g0(y) + g1(y)[f1(x) + f2(x) · ϕ(y)]

where g1 has a constant sign, and ϕ is monotone.

@freakonometrics 25

Page 26: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

One-Switch Utility Independence

U(x, y) = g0(y) + g1(y)[f1(x) + f2(x)ϕ(y)]

It is possible to express those function in terms of utility

- g0(y) = U(x, y)

- g1(y) = [U(x, y)− U(x, y)]

- f1(x) = U(x|y)

- f2(x) = [U(x|y)− U(x|y)]

ϕ(y) =U(x|y)− U(x|y)U(x|y)− U(x|y)

@freakonometrics 26

Page 27: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Utility Trees and Bidirectional Utility Diagrams

From Abbas (2011), let x = (xi,x(i))

Condister the normalized conditional utility for xi at x,

U(xi|x(i)) =U(xi,x(i))− U(xi,x(i))U(xi,x(i))− U(xi,x(i))

Note that

U(xi,x(i)) = U(xi,x(i)) · U(xi|x(i)) + U(xi,x(i)) · [1− U(xi|x(i))]

Thus, for two attributes

U(x, y) = U(x, y) · U(x|y) + U(x, y) · [1− U(x|y)]

@freakonometrics 27

Page 28: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Utility Trees and Bidirectional Utility Diagrams

U(x, y) = U(x, y) · U(x|y) + U(x, y) · [1− U(x|y)]

But it is also possible to expand it

U(x, y) = U(x, y)︸ ︷︷ ︸=U(y|x)·U(x,y)

+[1−U(y|x)]·U(x,y)

U(x|y) + U(x, y)︸ ︷︷ ︸=U(y|x)·U(x,y)

+[1−U(y|x)]·U(x,y)

[1− U(x|y)]

which give four terms.

Simplified version can be obtained with additional assumptions:

Utility independence, U(x|y) = U(x|y) = U(x|y) ∀y

Boundary independence, U(x|y) = U(x|y)

@freakonometrics 28

Page 29: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Utility Trees and Bidirectional Utility Diagrams

U(x, y) = U(x, y)︸ ︷︷ ︸=U(y|x)·U(x,y)

+[1−U(y|x)]·U(x,y)

U(x|y) + U(x, y)︸ ︷︷ ︸=U(y|x)·U(x,y)

+[1−U(y|x)]·U(x,y)

[1− U(x|y)]

@freakonometrics 29

Page 30: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Utility Trees and Bidirectional Utility Diagrams

... and one can define directional utility diagrams

x ↔ y : mutual utility independence

x → y : Directional utility independence, x independent of y

x ← y : Directional utility independence, y independent of x

x ↔ y : no independence

In higher dimension, it is more complex...

@freakonometrics 30

Page 31: Multiattribute utility copula

Arthur CHARPENTIER - Multi-attribute Utility & Copulas

Abbas, A. E, R. A. Howard. 2005. Attribute Dominance Utility. Decisions Analysis, 2 (4)

Abbas, A. E and D. E. Bell. 2011. One-Switch Independence for Multiattribute UtilityFunctions, Operations Research, 59(3) 764-771.

Abbas, A. E. 2009. Multiattribute Utility Copulas. Operations Research, 57 (6), 1367-1383.

Abbas, A. E. 2013. Utility Copula Functions Matching all Boundary Assessments. OperationsResearch, 61(2), 359-371.

Abbas, A. E. 2011. General Decompositions of Multiattribute Utility Functions. J.Multicriteria Decision Analysis, 17 (1, 2), 37–59.

Abbas, A.E and D.E. Bell. 2011. One-Switch Independence for Multiattribute UtilityFunctions. Operations Research, 59 (3) 764-771.

Abbas, A.E. 2011. The Multiattribute Utility Tree. Decision Analysis, 8 (3), 165-169 .

Abbas, A.E. 2011. Decomposing the Cross-Derivatives of a Multiattribute Utility Function intoRisk Attitude and Value. Decision Analysis, 8 (2) 103-116.

Clemen, R.T. and T. Reilly. 1999. Correlations and Copulas for Decision and Risk Analysis.Management Science, Vol 45, No. 2.

Keeney, R.L., H. Raiffa. 1976. Decisions with Multiple Objectives. Wiley

Matheson, J.E., R.A. Howard. 1968. An Introduction to Decision Analysis in The Principlesand Applications of Decision Analysis.

@freakonometrics 31