Mediators in Position Auctions
Itai Ashlagi
Dov Monderer Moshe Tennenholtz
Technion
Talk Outline
• Mediators in games with complete information.
• Mediators and mediated equilibrium in games with incomplete
information.
• Apply the theory to position auctions.
Mediators- Complete Information
Monderer & Tennenholtz 06
• A mediator is defined to be a reliable entity, which can ask the agents for
the right to play on their behalf, and is guaranteed to behave in a pre-
specified way based on messages received from the agents.
• However, a mediator can not enforce behavior; agents can play in the game directly without the mediator's help.
Mediators – Complete Information
c
d 5,0 1,1 0,5 4,4
c d
Mediator:
If both use the mediator services – (c,c)
If a single player chooses the mediator, the mediator plays d on behalf of this player.
c
d 5,0 1,1 0,5 4,4
c d
0,5 1,1
m
Mediated
1,1 3,3 2,8 3,6
Games with Incomplete Information
s1 s2
t1
t2
1,4 7,2
0,5 6,4
1,5
5,1
4,2
2,4 5,0
6,0 5,2 0,2
1 2
3 4
Games with Incomplete Information
s1 s2
t1
t2
0,5 3,6 7,2 1,4 2,8 5,1
1,5 6,4
5,0 2,4 4,2 3,3 0,2 5,2
1,1 6,0
1 2
3 4
Ex–post equilibrium -
Implementing an Outcome Function by Mediation
2,2 0,0
3,0 5,2
5,2 3,0
0,0 2,2
A B
a b
a b a b
No ex-post
equilibrium in G
G
Implementing an Outcome Function by Mediation
2,2 0,0
3,0 5,2
5,2 3,0
0,0 2,2
A B
a b
a b a b
No ex-post
equilibirum in G
G
M{1,2}(m,m-A)=(a,a) M{1,2}(m,m-B)=(b,b)
M{1} =a
M{2}(m-A)=b, M{2}(m-B)=a
Mediator M
Implementing an Outcome Function by Mediation
2,2 0,0
3,0 5,2
5,2 3,0
0,0 2,2
A B
a b
a b a b
No ex-post
equilibirum in G
G
M{1,2}(m,m-A)=(a,a) M{1,2}(m,m-B)=(b,b)
M{1} =a
M{2}(m-A)=b, M{2}(m-B)=a
Mediator M
5,2 3,0
2,2 5,2
A
3,0 5,2
3,0 5,2
0,0
2,2 0,0 2,2
m
b
m-A m-B a b
a a
b
2,2 0,0
5,2 2,2
0,0 2,2
0,0 2,2
3,0
5,2 3,0 5,2
m
m-A m-B a b
B
GM
Implementing an Outcome Function by Mediation (cont.)
5,2 3,0
2,2 5,2
A
3,0 5,2
3,0 5,2
0,0
2,2 0,0 2,2
m
b
m-A m-B a b
a a
b
2,2 0,0
5,2 2,2
0,0 2,2
0,0 2,2
3,0
5,2 3,0 5,2
m
m-A m-B a b
B
GM
M{1,2}(m,m-A)=(a,a) M{1,2}(m,m-B)=(b,b)
M{1} =a
M{2}(m-A)=b, M{2}(m-B)=a
Implementing an Outcome Function by Mediation (cont.)
5,2 3,0
2,2 5,2
A
3,0 5,2
3,0 5,2
0,0
2,2 0,0 2,2
m
b
m-A m-B a b
a a
b
2,2 0,0
5,2 2,2
0,0 2,2
0,0 2,2
3,0
5,2 3,0 5,2
m
m-A m-B a b
B
GM
The mediator implements the following outcome function:
A)=(a,a) (B)=(b,b)
M{1,2}(m,m-A)=(a,a) M{1,2}(m,m-B)=(b,b)
M{1} =a
M{2}(m-A)=b, M{2}(m-B)=a
Mediators & Mechanism Design
Mechanism design – find a game to implement
Mediators – find a mediator to implement for a given game.
Position Auctions - Model
• k – #positions, n - #players n>k
• vi - player i’s valuation per-click
• j- position j’s click-through rate 1>2>> Allocation rule – jth highest bid to jth highest position
Tie breaks - fixed order priority rule (1,2,…,n) Payment scheme
pj(b1,…,bn) – position j’s payment under bid profile (b1,…,bn) Quasi-linear utilities: utility for i if assigned to position j
and pays qi per-click isj(vi-qi)
Outcome(b) = (allocation(b), position payment vector(b))
Some Position Auctions
• VCG pj(b)=l¸j+1b(l)(k-1-k)/j
• Self-price pj(b)=b(j)
• Next –price pj(b)=b(j+1)
There is no (ex-post) equilibrium in the self-price and next-price position auctions.
In which position auctions can the VCG outcome function be implemented? Why should we do it?
Example
self-price, single slot auction
1=1, n=2
c-mediator v1
v2
v2 0 v1¸ v2
Example
self-price, single slot auction
1=1, n=2
c-mediator v1
v2
v2 0 v1¸ v2
c-mediator
vi cv
i
For every c¸1 vcg can be
implemented in the single- slot self-price auction.
c>1 can lead to negative utilities for players who trust the mediator.
Example
self-price, single slot auction
1=1, n=2
c-mediator v1
v2
v2 0 v1¸ v2
c-mediator
vi cv
i
For every c¸1 vcg can be
implemented in the single- slot self-price auction.
c>1 can lead to negative utilities for players who trust the mediator.
Example
self-price, single slot auction
1=1, n=2
c-mediator v1
v2
v2 0 v1¸ v2
c-mediator
vi cv
i
For every c¸1 vcg can be
implemented in the single- slot self-price auction.
Valid Mediators – players who trust the mediator never loose money
Self-Price Position Auctions
n=3, k=2
v1=5, v2=5, v3=10
The VCG outcome function can not be implemented in the self-price position auction unless k=1.
Self-Price Position Auctions
n=3, k=2
v1=5, v2=5, v3=10
The VCG outcome function can not be implemented in the self-price position auction unless k=1.
VCG
player 3, pays 5 player 1, pays 5 player 2, pays 0
Self-Price Position Auctions
n=3, k=2
v1=5, v2=5, v3=10
The mediator must submit 5 on behalf of both players 1 and 3.
But then player 3 will not be assigned to the first position!
The VCG outcome function can not be implemented in the self-price position auction unless k=1.
VCG
player 3, pays 5 player 1, pays 5 player 2, pays 0
Theorem: There exists a valid mediator that implements vcg in the next-price position auction
Next-price Position Auctions
Edelman, Ostrovsky and Schwarz provided a mechanism that can be viewed as a “simplified” form of a mediator where participation is mandatory.
1+p1vcg(v)
p2vcg(v) p1vcg(v)
pk-1vcg(v)
pkvcg(v) pkvcg(v)/2 pkvcg(v)/2 Position s
accordi ng to v If all players
choose the mediator:
MN(v}=
Mediator for the next-
price auction
1+p1vcg(v)
p2vcg(v) p1vcg(v)
pk-1vcg(v)
pkvcg(v) pkvcg(v)/2 pkvcg(v)/2 Position s
accordi ng to v
If some players play directly:
If all players choose the mediator:
MN(v}=
Mediator for the next-
price auction
Proof:
1. pj-1vcg(v)¸ pjvcg(v) for every j¸ 2
where equality holds if and only if v(j)=…=v(k+1)
Proof:
1. pj-1vcg(v)¸ pjvcg(v) for every j¸ 2
where equality holds if and only if v(j)=…=v(k+1) 2.Reporting untruthfully to the mediator
is non-beneficial.
Proof:
1. pj-1vcg(v)¸ pjvcg(v) for every j¸ 2
where equality holds if and only if v(j)=…=v(k+1) 2.Reporting untruthfully to the mediator
is non-beneficial.
3. pjvcg(v) · v(j+1) for every j
h - i’s position without deviation h’ – i’s position after deviation
Proof:
1. pj-1vcg(v)¸ pjvcg(v) for every j¸ 2
where equality holds if and only if v(j)=…=v(k+1) 2.Reporting untruthfully to the mediator
is non-beneficial.
3. pjvcg(v) · v(j+1) for every j
h - i’s position without deviation h’ – i’s position after deviation VCG utility in
h position ¸ VCG utility in h’ position
Proof:
1. pj-1vcg(v)¸ pjvcg(v) for every j¸ 2
where equality holds if and only if v(j)=…=v(k+1) 2.Reporting untruthfully to the mediator
is non-beneficial.
3. pjvcg(v) · v(j+1) for every j
h - i’s position without deviation h’ – i’s position after deviation VCG utility in
h position ¸ VCG utility in
h’ position next-price utility in ¸ h’ position
Proof:
1. pj-1vcg(v)¸ pjvcg(v) for every j¸ 2
where equality holds if and only if v(j)=…=v(k+1) 2.Reporting untruthfully to the mediator
is non-beneficial.
3. pjvcg(v) · v(j+1) for every j
h - i’s position without deviation h’ – i’s position after deviation VCG utility in
h position ¸ VCG utility in
h’ position next-price utility in ¸ h’ position
Existence of Valid
Mediators for Position Auctions
Theorem:
Let G be a position auction. If the following conditions hold then there exists a valid mediator that implements vcg in G:
C1: position payment depends only on lower position’s bids.
C2: VCG cover – any VCG outcome can be obtained by some bid profile.
C3: G is monotone
Each one of these conditions are necessary.
*assumption – players don’t pay more than their bid.
The Mediator
b(v) – a “good” profile for v (obtains the desired outcome for v).
vi = (v-i, Z) - i has the “largest” value
MN(v)=b(v)
MN\{i}(v)=b-i(vi)
MS(vs)=vS (other subsets S)
Existence of Valid Mediators for Position Auctions (cont.)
Corollaries
1. Suppose pj(b)=wjb(j+1) , 0· wj· 1.
Valid mediators exist if and only if for every j, wj· wj+1
2. Valid mediators exist in k-price position auctions
Quality effect
Valid mediators exist in the existing (Google, Yahoo) position auctions, where the click-
through rate for player i in position j is ®ij
Related Work
Mediators in Incomplete Information Games Collusive Bidder Behavior at Single-Object
Second-Price and English Auctions (Graham and Masrshall 1987)
Bidding Rings (McAfee and McMillan 1992)
Bidding Rings Revisited (Bhat, Leyton-Brown, Shoham and Tennenholtz 2005)
Position Auctions
Internet Advertising and the Generalized
Second Price Auction (Edelman, Ostrovsky and Schwarz 2005)
Conclusions
• Introduced the study of mediators in games with incomplete information.
• Applied mediators to the context of position auctions.
• Characterization of the position
auctions in which the VCG outcome function can be implemented.
Future Work
• Stronger implementations in position auctions (2-strong, k- strong).
• Mediator in other applications.
• Mediators and Learning.