Lecturer: Moni Naor

Algorithmic Game Theory

Uri Feige Robi Krauthgamer Moni Naor

Lecture 10: Mechanism Design

Announcements

• January: course will be 1300:-15:00 – The meetings on Jan 7th, 14th and 21st 2009

Recap social choice • Social choice: collectively choosing among

outcomes or aggregate preferences • Arrow’s Impossibility Theorem• Gibbard-Satterthwaite Theorem: There exists no

social choice function f for more than 2 alternatives that is simultaneously:– Onto: for every candidate, there are some votes that make

the candidate win– Nondictatorial– Incentive compatible

Change must happen at some profile i*

•Where voter i* changed his opinion

Proof of Arrow’s Theorem: Find the Dictator

Claim: For any a,b 2 A consider sets of profiles

ab ba ba … ba ab ab ba … ba ab ab ab … ba … … … ab ab ab ba

a Á b b Á a

Claim: this i* is the dictator!

Hybrid argumentVoters

1

2

n

Profiles

0 1 2

…

n

Single-peaked preferences [Black 48]• Suppose alternatives are ordered on a line• Every voter prefers alternatives that are closer to

her most preferred alternative

a1 a2 a3 a4 a5

v1v2 v3v4

v5

• Choose the median voter’s peak as the winner

• Strategy-proof!

What about Probabilistic Voting Schemes?

Electing the Doge in the Republic of Venice 1268-1797

• A sequence of electoral colleges, where at each stage: – A sub-college is selected at random

(lottery)– The sub college elects the next

electoral college by approval voting.

• Final college elects the Doge

Lottery

Approval

Probabilistic Voting Schemes

Can do something ``non trivial” to get truthful voting• Elect a random leader/dictator • Choose at random a pair of alternatives and see

which one is preferred by the majority.But this all we can do:Any scheme has to be a combination of such rules

Range Voting

• Each voter ranks the candidates in a certain range (say 0-99)

• The votes for all candidates are summed up and the one with highest total score wins

Can be considered as a generalization of approval voting from the range 0-1

No incentive for voter to rate a candidate lower than a candidate they like less.

Mechanism Design

• Mechanisms• Recall: We want to implement a social choice function

– Need to know agents’ preferences – They may not reveal them to us truthfully

• Example:– One item to allocate:

• Want to give it to the participant who values it the most– If we just ask participants to tell us their preferences: may lie

• Can use payments result is also a payment vector p=(p1,p2, … pn)

The setting

• Set of alternatives A– Who wins the auction– Which path is chosen– Who is matched to whom

• Each participant: a value function vi:A R

• Can pay participants: valuation of choice a with payment pi is vi(a)+pi

Quasi linear preferences

Example: Vickrey’s Second Price Auction

• Single item for sale• Each player has scalar value wi – willingness to pay• If he wins item and has to pay p: utility wi-p• If someone else wins item: utility 0Second price auction: Winner is the one with the highest

declared value wi. Pays the second highest bid p*=maxj i wj

Theorem (Vickrey): for any every w1, w2,…,wn and every wi’. Let ui be i’s utility if he bids wi and u’i if he bids wi’. Then ui ¸ u’i..

Despite private information and selfish behavior compute “reliably” the max function!

Direct Revelation MechanismA direct revelation mechanism is a social choice function

f: V1 V2 … Vn A

and payment functions pi: V1 V2 … Vn R

• Participant i pays pi(v1, v2, … vn)

A mechanism (f,p1, p2,… pn) is incentive compatible if for every v=(v1, v2, …,vn), i and vi’ 2 V1: if a = f(vi,v-i) and a’ = f(v’i,v-i) then

vi(a)-pi(vi,v-i) ¸ vi(a’) -pi(v’i,v-i) Prefer telling the truth about vi

v=(v1, v2,… vn)

v-i=(v1, v2,… vi-1 ,vi+1 ,… vn)

Vickrey Clarke Grove MechanismA mechanism (f,p1, p2,… pn ) is called Vickrey-

Clarke-Grove (VCG) if • f(v1, v2, … vn) maximizes i vi(a) over A

– Maximizes welfare

• There are functions h1, h2,… hn where

hi: V1 V2 … Vn R does not depend on vi

we have that:pi(v1, v2, … vn) = hi(v-i) - j i vj(f(v1, v2,…

vn)) v=(v1, v2,… vn)

v-i=(v1, v2,… vi-1 ,vi+1 ,… vn)

Depends only on chosen alternative

Does not depend on vi

Example: Second Price AuctionRecall: f assigns the item to one participant andvi(j) = 0 if j i and vi(i)=wi

• f(v1, v2, … vn) = i s.t. wi =maxj(w1, w2,… wn)

• hi(v-i) = maxj(w1, w2, … wi-1, wi+1 ,…, wn)

• pi(v) = hi(v-i) - j i vj(f(v1, v2,… vn))

If i the winner pi(vi) = hi(v-i) = maxj i wj

and for j i pj(vi)= wi – wi = 0

A={i wins|I 2 I}

maximizes i vi(a) over A

VCG is Incentive CompatibleTheorem: Every VCG Mechanism (f,p1, p2,… pn) is

incentive compatibleProof: Fix i, v-i, vi and v’i. Let a=f(vi,v-i) and a’=f(v’i,v-i). Have to show

vi(a)-pi(vi,v-i) ¸ vi (a’) -pi(v’i,v-i) Utility of i when declaring vi: vi(a) + j i vj(a) - hi(v-i)

Utility of i when declaring v’i: vi(a’)+ j i vj(a’)- hi(v-i)Since a maximizes social welfare

vi(a) + j i vj(a) ¸ vi(a’) + j i vj(a’)

Social welfare (of others) when he participates

Clarke Pivot RuleWhat is the “right”: h?

Individually rational: participants always get non negative utilityvi(f(v1, v2,… vn)) - pi(v1, v2,… vn) ¸ 0

No positive transfers: no participant is ever paid money pi(v1, v2,… vn) ¸ 0

Clark Pivot rule: Choosing hi(v-i) = maxb 2 A j i vj(b)

Payment of i when a=f(v1, v2,…, vn):

pi(v1, v2,… vn) = maxb 2 A j i vj(b) - j i vj(a)

i pays an amount corresponding to the total “damage” he causes other players: difference in social welfare caused by his participation

Social welfare when he does not participate

maximizes i vi(a) over A

Rationality of Clarke Pivot RuleTheorem: Every VCG Mechanism with Clarke pivot payments

makes no positive Payments. If vi(a) ¸ 0 then it is Individually rational

Proof: Let a=f(v1, v2,… vn) maximizes social welfareLet b 2 A maximize j i vj(b)

Utility of i: vi(a) + j i vj(a) - j i vj(b)

¸ j vj(a) - j vj(b) ¸ 0

Payment of i: j i vj(b) - j i vj(a) ¸ 0 from choice of b

Examples: Second Price AuctionSecond Price auction: hi(v-i) = maxj(w1, w2,…, wi-1, wi+1,…, wn)

= maxb 2 A j i vj(b)

Multiunit auction: if k identical items are to be sold to k individuals. A={S wins |S ½ I, |S|=k} and

vi(S) = 0 if i2S and vi(i)=wi if i 2 S Allocate units to top k bidders. They pay the k+1th priceClaim: this is

maxS’ ½ I\{i} |S’| =k j i vj(S’)-j i vj(S)

Generalized Second Price AuctionsMultiunit auction: if k identical items are to be sold to k

individuals. A={S wins |S ½ I, |S|=k} and

vi(S) = 0 if i2S and vi(i)=wi if i 2 S

Allocate units to top k bidders. The jth highest bidder pays bid j+1.

Common in web advertising

Claim: this is not incentive compatible

Examples: Public ProjectWant to build a bridge:

– Cost is C (if built) – Value to each individual vi

– Want to built iff i vj ¸ C

Player with vj ¸ 0 pays only if pivotal

j i vj < C but j vj ¸ C

in which case pays pj = C- j i vj

In general: i pj < C

Payments do not cover project cost’s• Subsidy necessary!

A={build, not build}

Equality only when

i vj = C

Set A of alternatives: all s-t paths

Buying a (Short) Path in a GraphA Directed graph G=(V,E) where each edge e is

“owned” by a different player and has cost ce.Want to construct a path from source s to destination t.• How do we solicit the real cost ce?

– Set of alternatives: all paths from s to t– Player e has cost: 0 if e not on chosen path and –ce if on– Maximizing social welfare: finding shortest s-t path:

minpaths e 2 path ce

A VCG mechanism that pays 0 to those not on path p: pay each e0 2 p: e 2 p’ ce - e 2 p\{e0} ce

where p’ is shortest path without eo

Clarke mechanism is not perfect• Requires payments & quasilinear utility functions• In general money needs to flow away from the system

– Strong budget balance = payments sum to 0– Impossible in general [Green & Laffont 77]

• Vulnerable to collusions• Maximizes sum of players’ utilities (social welfare)

– not counting payments) But: sometimes the center is not interested in maximizing

social welfare:– E.g. the center may want to maximize revenue

Games with Incomplete Information

Game defined by having for every player i2 I • A set of actions Xi

• A set of types Ti. The value ti 2 Ti is the private information i knows.

• A utility function ui: Ti X1 X2 … Xn R where ui(ti, x1, x2, … xn) is the utility obtained by i if his private information is ti and the profile of actions taken by all players is (x1, x2, … xn).

Player i chooses his action knowing ti but not other values

…Games with Incomplete Information

A strategy for player i2 I is si:Ti X1

A strategy si is (weakly) dominant if for all ti 2 Ti we have that si(ti) is a dominant strategy in the full information game defined by the ti’s: for all ti’s and all

x=(x1, x2, xi-1, x’i, xi+1 … xn) we have that

ui(ti, si(ti), x-i) ¸ ui(ti, x)

Alternative play

Games and MechanismsA mechanism is given by• Types T1, T2, … Tn

• Actions X1, X2, …, Xn • An alternative set A and outcome function a: X1 X2 … Xn A • Player’s valuation functions vi: T1 A R • Payment functions pi: X1 X2 … Xn R • The utility of player i ui(ti, x1, x2, … xn) = ui(ti, a(x1, x2, … xn)) - pi(x1, x2, … xn)

A mechanism implements a social choice function ff: T1 T2 … Tn A in dominant strategies if for some dominant strategies s1, s2, … sn (of the

induced game) for all t1, t2, … tn

f(t1, t2, … tn ) = a(s1(t1), s2(t2), … sn(tn))

Quasi linear preferences

The Revelation Principle

Theorem: if there exists an arbitrary mechanism implementing a social choice function f in dominant strategies, then there exists an incentive compatible mechanism that implements f

The payments of the players in the incentive compatible mechanism are identical to those obtained at equilibrium in the original mechanism

Proof: by simulation

Revelation Principle: Intuition

Player 1: t1

Player n: tn

...

Strategy

Strategy

Strategys1(t1) Original

“complex”“indirect”mechanism

Outcomea,p1,…,pn

Constructed “direct revelation” mechanism

Strategysn(tn)

..

.

Revelation Principle: Proof• Since si is dominant for player i, then for all ti, x:

vi(ti, a(si(ti), x-i)) - pi(si(ti), x-i)

¸ vi(ti,a(x))-pi(x)

• In particular for all x-i = s-i (t-i ) and xi = si (t’i )

To understand mechanism: can think of the equivalent direct revelation mechanism

Direct Characterization• A mechanism is incentive compatible iff the following hold for

all i and all vi

• The payment pi does not depend on vi but only on the alternative chosen f(vi, v-i) – the payment of alternative a is pa

• The mechanism optimizes for each player: f(vi, v-i) 2 argmaxa (vi(a)-pa)

Bayesian Nash Implementation• There is a distribution Di on the types Ti of Player i• It is known to everyone• The value ti 2Di

Ti is the private information i knows

• A profile of strategis si is a Bayesian Nash Equilibrium if for i all ti and all x’i

Ed-i[ui(ti, si(ti), s-i(t-i) )] ¸ Ed-i

[ui(ti, s-i(t-i)) ]

Bayesian Nash: First Price Auction• First price auction for a single item with two players.• Each has a private value t1 and t2 in T1=T2=[0,1]

• Does not make sense to bid true value – utility 0.• There are distributions D1 and D2

• Looking for s1(t1) and s2(t2) that are best replies to each other

• Suppose both D1 and D2 are uniform.

Claim: In the strategies s1(t1) = ti/2 are in Bayesian Nash Equilibrium t1

Cannot winWin half the time

Expected RevenuesExpected Revenue:

– For first price auction: max(T1/2, T2/2) where T1 and T2 uniform in [0,1]

– For second price auction min(T1, T2) – Which is better? – Both are 1/3.– Coincidence?

Theorem [Revenue Equivalence]: under very general conditions, every two Bayesian Nash implementations of the same social choice function if for some player and some type they have the same expected

payment then– All types have the same expected payment to the player– If all player have the same expected payment: the expected revenues

are the same