Making Choquet Expected Utility Tractable for the Study of Ambiguity Peter P. Wakker, December 20, 2005 Tel-Aviv University Aim:Make Choquet-expected utility,

Slides:



Advertisements
Similar presentations
From risk to opportunity Lecture 11 John Hey and Carmen Pasca.
Advertisements

From risk to opportunity Lecture 10 John Hey and Carmen Pasca.
6.896: Topics in Algorithmic Game Theory Lecture 20 Yang Cai.
Paradoxes in Decision Making With a Solution. Lottery 1 $3000 S1 $4000 $0 80% 20% R1 80%20%
Lecture XXIII.  In general there are two kinds of hypotheses: one concerns the form of the probability distribution (i.e. is the random variable normally.
Decision Theory Lecture 8. 1/3 1 1/4 3/8 1/4 3/8 A A B C A B C 1/2 A B A C Reduction of compound lotteries 1/2 1/4 A B C.
Lecture 4 Environmental Cost - Benefit - Analysis under risk and uncertainty.
1 12. Principles of Parameter Estimation The purpose of this lecture is to illustrate the usefulness of the various concepts introduced and studied in.
1 Lower Distribution Independence Michael H. Birnbaum California State University, Fullerton.
Evaluating Non-EU Models Michael H. Birnbaum Fullerton, California, USA.
Rational choice: An introduction Political game theory reading group 3/ Carl Henrik Knutsen.
EC941 - Game Theory Prof. Francesco Squintani Lecture 8 1.
Chapter 13 Risk Attitudes ..
1 st lecture Probabilities and Prospect Theory. Probabilities In a text over 10 standard novel-pages, how many 7-letter words are of the form: 1._ _ _.
Judgment and Decision Making in Information Systems Utility Functions, Utility Elicitation, and Risk Attitudes Yuval Shahar, M.D., Ph.D.
Lecture 4 on Individual Optimization Risk Aversion
1 A Brief History of Descriptive Theories of Decision Making Kiel, June 9, 2005 Michael H. Birnbaum California State University, Fullerton.
Omission or Paternalism Peter P. Wakker (& Bleichrodt & Pinto & Abdellaoui); Seminar at University of Chicago, School of Business, June 23, Hypothetical.
Using Prospect Theory to Study Unknown Probabilities ("Ambiguity") by Peter P. Wakker, Econ. Dept., Erasmus Univ. Rotterdam (joint with Mohammed Abdellaoui.
Making Rank-Dependent Utility Tractable for the Study of Ambiguity Peter P. Wakker, June 16, 2005 MSE, Université de Paris I Aim:Make rank-dependent utility.
Tutorial on Risk and Uncertainty Peter P. Wakker Part 1: Introduction into Prospect Theory. Part 2: Using Prospect Theory to Better Describe and Prescribe.
Decision-Principles to Justify Carnap's Updating Method and to Suggest Corrections of Probability Judgments Peter P. Wakker Economics Dept. Maastricht.
Maximum likelihood (ML) and likelihood ratio (LR) test
Utilitarianism for Agents Who Like Equity, but Dislike Decreases in Income Peter P. Wakker (& Veronika Köbberling)
Ambiguity will concern Dow Jones & Nikkei indexes today: Topic: Quantitative measurement of capacities under ambiguity. Question: How do people perceive.
½U(1000) + ½U(x 1 ) = ½U(8000) + ½U(x 0 ) _ ( U(8000)  U(1000) ) Tradeoff (TO) method for EU x2x x 1 ~ ½ ½ ½ ½ x4x x 3 ½ ½ ~ ½ ½.
Tractable Quantifications of Ambiguity Attitudes by Peter P. Wakker, Econ. Dept., Erasmus Univ. Rotterdam (joint with Mohammed Abdellaoui & Aurélien Baillon)
1 A Brief History of Descriptive Theories of Decision Making: Lecture 2: SWU and PT Kiel, June 10, 2005 Michael H. Birnbaum California State University,
Lecture 03: Risk Preferences and Expected Utility Theory
Example, on Dow Jones & Nikkei indexes today: Peter P. Wakker & Enrico Diecidue & Marcel Zeelenberg Domain: Individual Decisions under Ambiguity (events.
On Dow Jones & Nikkei indexes today: Peter P. Wakker & Enrico Diecidue & Marcel Zeelenberg This file will be on my homepage on coming Monday. Domain: Decisions.
Combining Bayesian Beliefs and Willingness to Bet to Analyze Attitudes towards Uncertainty by Peter P. Wakker, Econ. Dept., Erasmus Univ. Rotterdam (joint.
Uncertainty RUD, June 28, 2006, Peter P. Wakker Paper: Chapter for New Palgrave. Its didactical beginning is not optimal for this specialized audience.
Adapting de Finetti's Proper Scoring Rules for Measuring Subjective Beliefs to Modern Decision Theories of Ambiguity Gijs van de Kuilen, Theo Offerman,
PSY 5018H: Math Models Hum Behavior, Prof. Paul Schrater, Spring 2005 Rational Decision Making.
Maximum likelihood (ML)
A new tool is presented for measuring beliefs/likelihoods under uncertainty. It will simplify:  preference axiomatizations;  quantitative measurements;
Definitions In statistics, a hypothesis is a claim or statement about a property of a population. A hypothesis test is a standard procedure for testing.
Definition and Properties of the Production Function Lecture II.
Lecture II-2: Probability Review
Combining Bayesian Beliefs and Willingness to Bet to Analyze Attitudes towards Uncertainty by Peter P. Wakker, Econ. Dept., Erasmus Univ. Rotterdam (joint.
Statistics 11 Hypothesis Testing Discover the relationships that exist between events/things Accomplished by: Asking questions Getting answers In accord.
Sociology 5811: Lecture 7: Samples, Populations, The Sampling Distribution Copyright © 2005 by Evan Schofer Do not copy or distribute without permission.
Ambiguity will concern Dow Jones & Nikkei indexes today: Peter P. Wakker & Enrico Diecidue & Marcel Zeelenberg Lecture + paper are on my homepage Domain:
STOCHASTIC DOMINANCE APPROACH TO PORTFOLIO OPTIMIZATION Nesrin Alptekin Anadolu University, TURKEY.
Thinking and Decision Making
Health State Unable to perform some tasks at home and/or at work Able to perform all self care activities (eating, bathing, dressing) albeit with some.
Adapting de Finetti's Proper Scoring Rules for Measuring Bayesian Subjective Probabilities when Those Probabilities Are not Bayesian Peter P. Wakker (&
On the smooth ambiguity model and Machina’s reflection example Robert Nau Fuqua School of Business Duke University.
New Views on Risk Attitudes Peter P. Wakker Economics University of Amsterdam € 100 € 0€ 0 ½ ½ or € 50 for sure What would you rather have? Such gambles.
1 TABLE OF CONTENTS PROBABILITY THEORY Lecture – 1Basics Lecture – 2 Independence and Bernoulli Trials Lecture – 3Random Variables Lecture – 4 Binomial.
Modern Ways to Model Risk and Uncertainty Peter P. Topic: prospect theory (  classical expected utility) for modeling risk/uncertainty/ambiguity.
Ambiguity Made Operational by Peter P. Wakker, Econ. Dept., Erasmus Univ. Rotterdam (joint with Mohammed Abdellaoui & Aurélien Baillon) ESA, Tucson, Oct.
Ellsberg’s paradoxes: Problems for rank- dependent utility explanations Cherng-Horng Lan & Nigel Harvey Department of Psychology University College London.
Notes: Use this cover page for internal presentations Dynamic Reference Points: Investors as Consumers of Uncertainty Greg B Davies
PROBABILITY AND STATISTICS FOR ENGINEERING Hossein Sameti Department of Computer Engineering Sharif University of Technology Principles of Parameter Estimation.
Lecture 3 on Individual Optimization Uncertainty Up until now we have been treating bidders as expected wealth maximizers, and in that way treating their.
Statistical Decision Theory Bayes’ theorem: For discrete events For probability density functions.
Information Theory for Mobile Ad-Hoc Networks (ITMANET): The FLoWS Project Competitive Scheduling in Wireless Networks with Correlated Channel State Ozan.
Expected Value, Expected Utility & the Allais and Ellsberg Paradoxes
Allais Paradox, Ellsberg Paradox, and the Common Consequence Principle Then: Introduction to Prospect Theory Psychology 466: Judgment & Decision Making.
1 Definitions In statistics, a hypothesis is a claim or statement about a property of a population. A hypothesis test is a standard procedure for testing.
Transient Unterdetermination and the Miracle Argument Paul Hoyningen-Huene Leibniz Universität Hannover Center for Philosophy and Ethics of Science (ZEWW)
Adapting de Finetti's proper scoring rules for Measuring Subjective Beliefs to Modern Decision Theories ofAmbiguity Peter P. Wakker (& Gijs van de Kuilen,
12. Principles of Parameter Estimation
Hypothesis Testing and Confidence Intervals (Part 1): Using the Standard Normal Lecture 8 Justin Kern October 10 and 12, 2017.
Ascertaining certain certainties in choices under uncertainty
Peter P. Wakker & Enrico Diecidue & Marcel Zeelenberg
12. Principles of Parameter Estimation
Presentation transcript:

Making Choquet Expected Utility Tractable for the Study of Ambiguity Peter P. Wakker, December 20, 2005 Tel-Aviv University Aim:Make Choquet-expected utility, "rank-dependent utility," tractable for a general public and specialists alike, in particular for studying ambiguity. Tool:Ranks. Spinoff:Some changes of minds (?)

2 Expected utility: x1x1 xnxn p1p1 pnpn  p 1 U(x 1 ) p n U(x n ) (p 1 :x 1,…,p n :x n ) = x1x1 xnxn p1p1 pnpn is prospect yielding $x j with probability p j, j=1,…,n. 1.Expected Utility for Risk

x2x2 xnxn p2p2 pnpn p1p1 y2y2 ymym q2q2 qmqm p1p1 r r r r 3 Well-known implication: Independence from common consequence ("sure-thing principle"):  go to p. 28, where RDU = EU for risk next p. rank- moderately-  y2y2 ymym q2q2 qmqm p1p1 x2x2 xnxn p2p2 pnpn p1p1   

w w bb.10 bb ww 4 Well-known violation of expected utility: Allais paradox. M: million $ .10 1M M M.01 5M Is the certainty effect. next p. 1M  5M M OK for RDU, with "pessimism" or "inverse-S." go to p. 27, where RDU  EU for risk > < EU

Other well-known violation of expected utility, for unknown probabilities: Ellsberg paradox. Known urn: 100 bals, 50 red, 50 black. Unknown urn: 100 ball, ? red, 100–? black. Common preferences: (Red known : $100)  (Red unknown : $100) & (Black known : $100)  (Black unknown : $100). Violate expected utility. 5

Question 1 to audience: From what can we best infer that people deviate from expected utility for risk (given prob s )? a. Nash equilibria. b. Allais paradox. c. Ellsberg paradox. 6

Question 2 to audience: From what can we best infer that people deviate from expected utility for uncertainty (unknown probabilities)? a. Nash equilibria. b. Allais paradox. c. Ellsberg paradox. 7

Question 3 to audience: Assume rank-dependent utility for unknown probabilities ("Choquet Expected utility"). From what can we best infer that nonadditive measures are convex (= superadditive, "uncertainty-averse," "pessimistic")? a. Nash equilibria. b. Allais paradox. c. Ellsberg paradox. 8

After this lecture: Answer to Question 1 ("nonEU for risk") is: Allais paradox. Answer to Question 3 ("capacities convex in rank-dependent utility") is: Allais paradox! Answer to Question 2 ("nonEU for uncertainty") is: both Allais and Ellsberg paradox. P.s.: I do think that the Ellsberg paradox has more content than the Allais paradox. Explained later. 9

Related change of mind: The inequality Decision under risk  Decision under uncertainty in the strict sense of [ Decision under risk  Decision under uncertainty =  ] is incorrect! 10 Decision under risk  Decision under uncertainty. That's how it is!

Idea that Allais paradox speak only to risk, and not to uncertainty, does not even make sense to me! 11

Outline of lecture: 1.Expected Utility for Risk (almost done). 2.Expected Utility for Uncertainty. 3.Rank-Dependent Utility for Risk, Defined through Ranks. 4.Where Rank-Dependent Utility Differs from Expected Utility for Risk. 5.Where Rank-Dependent Utility Agrees with Expected Utility for Risk, and some properties. 6.Rank-Dependent Utility for Uncertainty, Defined through Ranks. 7.Where Rank-Dependent Utility Differs from Expected Utility for Uncertainty as it Did for Risk. 8.Where Rank-Dependent Utility Agrees with Expected Utility for Uncertainty. 9.Where Rank-Dependent Utility Differs from Expec- ted Utility for Uncertainty Differently than for Risk. 10. Applications of Ranks. 12

13 In preparation for rank-dependence and decision under uncertainty, remember: In (p 1 :x 1,…,p n :x n ), we have the liberty to number outcomes/probs such that x 1 ...  x n.

Wrong start for DUU: Let S = {s 1,…,s n } denote a finite state space. x : S   is an act, also denoted as an n- tuple x = (x 1,…,x n ). Is didactical mistake for rank-dependence! Why wrong? Later, for rank-dependence, ranking of outcomes will be crucial. Should use numbering of x j for that purpose; as under risk! Should not have committed to a numbering of outcomes for other reasons. So, start again: Expected Utility for Uncertainty

S: state space, or universal event (finite or infinite). Act is function x : S   with finite range. x = (E 1 :x 1, …, E n :x n ): yields x j for all s  E j, with: x 1,…,x n   are outcomes, E 1, …, E n are events partitioning S. No commitment to a numbering of outcomes! As for risk. This is an important notational point for rank- dependence (that model will come later). P.s.: If E 1,…,E n understood, we may write (x 1, …,x n ). 15

16 (E 1 :x 1,…,E n :x n ) = x1x1 xnxn E1E1 EnEn Subjective Expected utility:   (E 1 )U(x 1 )  (E 1 )U(x n ) x1x1 xnxn E1E1 EnEn U: subjective index of utility.  : subjective probability.

Monotonicity:      E x   E x; Notation:  E x is (x with outcomes on E replaced by  ): E 1 x = (E 1 :10,E 2 :x 2,.., E n :x n );  E n x = (E 1 :x 1,.., E n - 1 :x n - 1, E n :  ); etc. R and ranking position of E is R next p.

18 Well-known implication: sure-thing principle:  E x   E y R R   E x   E y R R rank- next p.go to p.42, RDU=EU

19 (Not-so-well-known) violation (MacCrimmon & Larsson '79; here Tversky & Kahneman '92). Within-subjects expt, 156 money managers. d: DJ tomorrw –DJ today. L: d 35; K: $1000.  (77%) L M H K 25K L M H 0 Certainty-effect & Allais hold for uncertainty in general, not only for risk! w b H w b H  (66%) 25K L M H 75K 0 25K L M H H b w H b w next p.go to p.36, RDU  EU Almost-unknown implication > < SEU OK for RDU: pes- simism / inverse-S.

3. Rank-Dependent Utility for Risk, Defined through Ranks Empirical findings: nonlinear treatment of probabilities. Hence RDU. Two insights needed for getting the theory: Insight 1. Deviations from expected value: also caused by nonlinear perception / processing of probability, through w(p). 20 Insight 2. Turn this into decision theory through rank-dependence. go to p. with ut.curv, if treated.

21 First rank-order x 1 > … > x n. Decision weight of x j will depend on: 1.p j ; 2.p j–1 + … + p 1, the probability of receiving something better. The latter will be called a rank. I will recommend using ranks instead of comonotonicity … Rank-dependent utility of x1x1 xnxn p1p1 pnpn ?

22 First rank-order x 1 > … > x n. Then rank-dependent utility is  1 U(x 1 ) + … +  n U(x n ) where  j = w(p j + p j–1 + … + p 1 ) – w(p j–1 + … + p 1 ). The decision weight  j depends on p j and on p j–1 + … + p 1 : p j–1 + … + p 1 is the rank of p j,x j, i.e. the probability of receiving something better. So, rank-dependent utility of x1x1 xnxn p1p1 pnpn ?

Ranks and ranked probabilities (formalized hereafter) are proposed as central concepts in this lecture. Were introduced by Abdellaoui & Wakker (2005). With them, rank-dependent life will be easier than it was ever before! 23

In general, pairs p r, also denoted p \r, with p+r  1 are called ranked probabilities. r is the rank of p.  (p r ) = w(p+r) – w(r) is the decision weight of p r. 24

25 Again, rank-dependent utility of x1x1 xnxn p1p1 pnpn with rank-ordering x 1  …  x n :  (p 1 r 1 )U(x 1 ) + … +  (p n r n )U(x n ) with r j = p j–1 + … + p 1 (so r 1 = 0). The smaller the rank r in p r, the better the outcome. The best rank, 0, is also denoted b, as in p b = p 0. The worst rank for p, 1–p, is also denoted w, as in p w = p 1–p.

There is a duality: For today: never mind. Rank = goodness-rank (probability of receiving an outcome ranked better). The, dual, badness-rank (probability of receving an outcome ranked worst) is sometimes more natural. 26

Allais paradox explained by rank dependence. Now the expression "rank dependence" can be taken literally! Where Rank-Dependent Utility Differs from Expected Utility for Risk go to p. 4, with Allais

Sure-thing principle of EU goes through completely for RDU if we replace probability by ranked probability Where Rank-Dependent Utility Agrees with Expected Utility for Risk, and some Properties go to p. 3, with risk-s.th.pr Now see Fig. of w-shaped.doc Some properties, suggested by Allais paradox, follow now (more to come later).

w convex (pessimism): r < r´  w(p+r) – w(r)  w(p+r´) – w(r´) Equivalent to:  (p r ) increasing in r. Remember: big rank is bad outcome. "Decision weight is increasing in rank." w concave (optimism) is similar.  2 more pessimistic than  1, i.e. w 2 more convex than w 1 : r < r´,  1 (p r ) =  1 (q r´ )   2 (p r )   2 (q r´ ) 29

30 Inverse-S: How define formally? Specify b, then require concavity below, convexity above? We prefer an alternative formalization w p insensitivity- region 1.Specify insensitivity region in the middle. 2. Specify through inequalities that marginal w is smaller there than in extreme regions. 3. Avoid comparisons between two extreme regions (by restricting domains of inequalities).

*: restricting domains to avoid comparisons between two extreme regions. 1 w 1 p In insensitivity region, marginal w smaller than in extreme regions. 31 (i)  (p b )   (p r ) r+p bwbw worst-rank region b best-rank region insensitivity- region  (p r ) rp  (p b ) 1–p  (p w ) best-rank overweighting worst-rank overweighting (ii)  (p w )   (p r ) on * [0,b w ] (r+p  b w ) on * [b b,1] (r  b b ). Inverse-S, or likelihood insensitivity, holds on region [b b,b w ], if (i) and (ii) below hold.

x = (E 1 :x 1, …, E n :x n ) was act, with E j 's partitioning S. Rank-dependent utility for uncertainty: (also called Choquet expected utility) W is capacity, i.e. (i) W(  ) = 0; (ii) W(S) = 1 for the universal event S; (iii) If A  B then W(A)  W(B) (monotonicity with respect to set inclusion) Rank-Dependent Utility for Uncertainty, Defined through Ranks

E R, with E  R = , is ranked event, with R the rank.  (E R ) = W(E  R) – W(R) is decision weight of ranked event. RDU of x = (E 1 :x 1, …, E n :x n ), with rank-ordering x 1  …  x n, is:  j  n  (E j R j )U(x j ) with R j = E j–1  …  E 1 (so R 1 =  ). Compared to SEU, ranks R j have now been added, expressing rank-dependence. 33

The smaller the rank R in E R, the better the outcome. The best (smallest) rank, , is also denoted b, as in E b = E . The worst (biggest) rank for E, E c, is also denoted w, as in E w = E E c. 34

Difficult notation in the past: S = {s 1,…,s n }. For RDU(x 1,…,x n ), take a rank-ordering  of s 1,...,s n such that x  1 ...  x  n. For each state s  j,   j = W(s  j,…, s  1 ) – W(s  j-1,…, s  1 ) RDU =   1 U(x  1 ) + … +   n U(x  n ) Due to  -notation, difficult to handle.  (2) = 5: Is state s 2 fifth-best, or is state s 5 second-best? I can never remember! 35

Convexity of W follows from Allais paradox! Easily expressable in terms of ranks:  (E R ) increasing in R Where Rank-Dependent Utility Differs from Expected Utility for Uncertainty as It Did for Risk go to p. 19, Allais for uncertainty

The whole measurement of utility, and preference characterization, of RDU for uncertainty is just the same as SEU, if we simply use ranked events instead of events! Where Rank-Dependent Utility Agrees with Expected Utility for Uncertainty go to p. on TO measuement etc. if presented

Allais: deviations from EU. Pessimism/convexity of w/W, or insensitivity/inverse-S. For risk and uncertainty alike. Deviations from EU in an absolute sense. Ellsberg: more deviations from EU for uncertainty than for risk. More pessimism/etc. for uncertainty than for risk. Deviations from EU in an relative sense. Deviations from EU: byproduct Where Rank-Dependent Utility Differs from Expected Utility for Uncertainty Differently than It Did for Risk

Historical coincidence: Schmeidler (1989) assumed EU for risk, i.e. linear w. Then: more pessimism/convexity for uncertainty than for risk (based on Ellsberg),  pessimism/convexity for uncertainty. Voilà source of numerous misunderstandings. 39

Big idea to infer of Ellsberg is not, I think, ambiguity aversion. Big idea to infer from Ellsberg is, I think, within-person between-source comparisons. Not possible for risk, because risk is only one source. Typical of uncertainty, where there are many sources. Uncertainty is a rich domain, with no patterns to be expected to hold in great generality. In this rich domain, many phenomena are present and are yet to be discovered. 40

General technique for revealing orderings  (A R )   (B R´ ) from preferences: Abdellaoui & Wakker (2005). Thus, preference foundations can be given for everything written hereafter Applications of Ranks

What is null event? Important for updating, equilibria, etc. E is null if W(E) = 0? E is null if W(E c ) = 1? E is null if (H: , E: , L:  ) ~ (H: , E: , L:  )? For some   , H …? Or for all   , H …? ? We: Wrong question! Better refer to ranked events! 42  (E b ) = 0.  (E w ) = 0.  (E H ) = 0. Plausible condition is null-invariance: independence of nullness from rank.

W convex:  increases in rank. W concave:  decreases in rank. W symmetric:  (E b ) =  (E w ). Inverse-S: (1) Specify (“event-”)insensitivity-region; (2) Specify through inequalities that marginal W is smaller there than in extreme regions. (3) Avoid comparisons between two extreme regions (by restricting domains of inequalities). 43 (2)  (E b )   (E R )  (E w )   (E R ) (3) on event-interval [ ,B w ] (E  R  B w ) on event-interval [B b,S] (R  B b ) (1) Insensitivity-region is [B b,B w ].

Theorem.  2 is more ambiguity averse than  1 in sense that W 2 is more convex than W 1 iff  1 (B R´ ) =  1 (A R ) with R´  R   2 (B R´ )   2 (A R ). 44

Theorem. Probabilistic sophistication holds  [  (A R )   (B R )   (A R´ )   (B R´ )]. In words: ordering of likelihoods is independent of rank. 45

Updating on A given B, with A  B. What is W(A) if B is observed? Gilboa (1989a,b): 46 W(A) W(B). Dempster & Shafer: 1 – W(B c ) W(A  B c ) – W(B c ). W(A) + 1 – W ( (B\A) c ) W(A). Jaffray, Denneberg:  (B b )  (A b )  (B w )  (A B c )  (A b ) +  ( (B\A) w )  (A b ) Gilboa & Schmeidler (1993): depends on optimism / pessimism. Ranks formalize this. Cohen,Gilboa,Jaffray,&Schmeidler (2000): lowest one did best.

Conclusion: With ranks and ranked probabilities or events, rank-dependent-utility/Choquet- expected-utility becomes considerably more tractable. 47

The end! 48

Psychology: 1 = w(.10)100 Psychology: x = w(p)100 Here is graph of w(p): U(1) = 0.10U(100) U(0) = Assume following data deviating from expected value 49 $ ~ $1$1 0 (a) $ 81 $ ~ $ 9$ 9 $ ~ $ 25 $ ~ $ ~ $ (b)(d)(c) (e) EU: EU: U(9) = 0.30U(100) = EU: U(x) = pU(100) = p. Here is graph of U(x): next p. go to p.27,RDU U(100) = 1, U(0) = (c) $ p $0$0 $ 100 $ 70 $ (a) (b) (d) (e) $ (e) (d) (c) p $0$ $ 70 $ 30 $ (a) (b)