On (Group) Strategy-proof
Mechanisms without Payment for Facility Location Games
Nguyen Kim Thang
WINE’10
Tuesday, December 14, 2010
Outline
Our main results Definitions Facility Games
Results
High level ideas.
Conclusion & Further directions
Facility Games
A network is represented by a graph minimum-length path.
G(V, E) d(u, v) =
1
10
agents, agent has location n i xi ∈ V
A det mechanism f : V n → V
x = !x1, . . . , xn" #→ P
f : V n → ∆(V )
A random mechanism Agent’s cost:
Agent’s cost: EF ∼P [d(xi, F )]
d(xi, F )
x = !x1, . . . , xn" #→ F
Tuesday, December 14, 2010
Facility Games
A network is represented by a graph minimum-length path.
G(V, E) d(u, v) =
1
10
agents, agent has location n i xi ∈ V
A det mechanism f : V n → V
x = !x1, . . . , xn" #→ P
f : V n → ∆(V )
A random mechanism
Agent’s cost: d(xi, F )
x = !x1, . . . , xn" #→ F
Facility Games
1
10
A mechanism is group strategy-proof if for all , there existsS ⊂ N i ∈ S
cost(xi, f (x!S, x−S)) > cost(xi, f (x))
A mechanism is strategy-proof if
cost(xi, f (x!i, x−i)) ≥ cost(xi, f (x))
∀i
Social objective functions: !
i∈N
cost(xi, F )
A mechanism is -approximation if α
cost(f ) ≤ α · OP T f
Tuesday, December 14, 2010
Results
Line graphs General graphs Det
Ran
n GSP (n − 1) SP 1 GSP
1 GSP
(upper bound) (lower bound)
Results
Line graphs General graphs Det
Ran
n GSP (n − 1) SP 1 GSP
1 GSP
(upper bound) (lower bound)
(2 − 2/n) SP, n GSP (u.b)
Tuesday, December 14, 2010
Results
Line graphs General graphs Det
Ran
n GSP (n − 1) SP
!
2 − 4 n1/3
"
SP, Ω(n1−!) GSP
1 GSP
1 GSP
(upper bound) (lower bound)
(2 − 2/n) SP, n GSP (u.b)
Results
Designing (G)SP mechanism without money is expensive
Line graphs General graphs Det
Ran
n GSP (n − 1) SP
!
2 − 4 n1/3
"
SP, Ω(n1−!) GSP
1 GSP
1 GSP
(upper bound) (lower bound)
(2 − 2/n) SP, n GSP (u.b)
Tuesday, December 14, 2010
Framework of lower bound
Instance 1 Instance 2 Instance m
Fix a graph
An instance differs from the previous one in some agents’ locations
Connect instances using strategy-proofness.
GSP mechanisms
Dictatorship: open the facility at location of some fixed agent.
GSP and -approximation.n
Deterministic lower bound by theorem of
Schummer and Vohra. (Any SP mechanism on a cycle graph is dictatorship.)
Tuesday, December 14, 2010
GSP mechanisms
Dictatorship: open the facility at location of some fixed agent.
GSP and -approximation.n
Deterministic lower bound by theorem of
Schummer and Vohra. (Any SP mechanism on a cycle graph is dictatorship.)
Thm: no randomized GSP mechanism is better than - approximation.
n1−3!/3
Deterministic GSP mechanisms
u1
u2 un
v1
v2 1
1 − ! vn
Tuesday, December 14, 2010
Deterministic GSP mechanisms
u1
u2 un
v1
v2 1
1 − ! vn
Case 1: facility is opened in U
All agents but the first one move to . The facility is not opened at
v1
(n − 1)(1 − !)
2 − ! ≈ n − 1 2 v1
Deterministic GSP mechanisms
u1
u2 un
v1
v2 1
1 − ! vn
Case 1: facility is opened in U
All agents but the first one move to . The facility is not opened at
v1
(n − 1)(1 − !)
2 − ! ≈ n − 1 2 v1
Tuesday, December 14, 2010
Deterministic GSP mechanisms
u1
u2 un
v1
v2 1
1 − ! vn
Case 1: facility is opened in U
All agents but the first one move to . The facility is not opened at
v1
(n − 1)(1 − !)
2 − ! ≈ n − 1 2 v1
Deterministic GSP mechanisms
u1
u2 un
v1
v2 1
1 − !
(n − 1)(1 − !)
2 − ! ≈ n − 1 2
Case 2: facility is opened in All agents but the last one move to .
V vn
Again,
vn
Case 1: facility is opened in U
All agents but the first one move to . The facility is not opened at
v1
(n − 1)(1 − !)
2 − ! ≈ n − 1 2 v1
Tuesday, December 14, 2010
Deterministic GSP mechanisms
u1
u2 un
v1
v2 1
1 − !
(n − 1)(1 − !) n − 1
Case 2: facility is opened in All agents but the last one move to .
V vn
vn
Case 1: facility is opened in U
All agents but the first one move to . The facility is not opened at
v1
(n − 1)(1 − !)
2 − ! ≈ n − 1 2 v1
Weakness
The argument does not carry.
u1
u2 un
v1
v2 1
1 − ! vn
An mechanism opens facility at with prob 1 − δ, δ
prevent agents
from collaborating.
2, . . . , n − 1 vn, u1
old cost of agent 2: 1 − !
new cost of agent 2: (1 − δ) · (1 − ") + δ · 1
Tuesday, December 14, 2010
Weakness
The argument does not carry.
u1
u2 un
v1
v2 1
1 − ! vn
An mechanism opens facility at with prob 1 − δ, δ
prevent agents
from collaborating.
2, . . . , n − 1 vn, u1
old cost of agent 2: 1 − !
new cost of agent 2: (1 − δ) · (1 − ") + δ · 1
Weakness
The argument does not carry.
u1
u2 un
v1
v2 1
1 − ! vn
An mechanism opens facility at with prob 1 − δ, δ
prevent agents
from collaborating.
2, . . . , n − 1 vn, u1
old cost of agent 2: 1 − !
new cost of agent 2: (1 − δ) · (1 − ") + δ · 1
Tuesday, December 14, 2010
Solution
u1 u2
un
v1
v2 1
vn β1
β2 βn−1 β1
β2
βn−1
Solution
u1 u2
un
v1
v2 1
vn β1
β2 βn−1 β1
β2
βn−1
Need: symmetry + asymmetry.
Tuesday, December 14, 2010
Solution
u1 u2
un
v1
v2 1
vn β1
β2 βn−1 β1
β2
βn−1
Need: symmetry + asymmetry.
modify brown edges.
2n−! > β1 > β2 > . . . > βn−1 > n−!
define the length in circular permutation
Solution
u1 u2
un
v1
v2 1
vn β1
β2 βn−1 β1
β2
βn−1
harder to prevent agents from collaborating.
Why it works?
amplify the gap by recursive construction
Need: symmetry + asymmetry.
modify brown edges.
2n−! > β1 > β2 > . . . > βn−1 > n−!
define the length in circular permutation
Tuesday, December 14, 2010
Lower bound construction
u1
u2 un
v1
v2 1
vn
Lemma (informal):
P[f ((x)) = vn] < 1 − !
! log n
log log n
"3!/2
Lower bound construction
u1
u2 un
v1
v2 1
vn
Lemma (informal):
P[f ((x)) = vn] < 1 − !
! log n
log log n
"3!/2
Recursive construction of multiple levels.
Tuesday, December 14, 2010
Lower bound construction
u1
u2 un
v1
v2 vn
u1
u2 un
v2 vn
u1
u2 un
v2 vn
Lemma (informal):
P[f ((x)) = vn] < 1 − !
! log n
log log n
"3!/2
Recursive construction of multiple levels.
Lower bound construction
u1
u2 un
v1
v2 vn
u1
u2 un
v1
v2 vn
u1
u2 un
v1
v2 vn
Lemma (informal):
Lemma (informal):
P[f ((x)) = vn] < 1 − n3!/2 P[f ((x)) = vn] < 1 − !
! log n
log log n
"3!/2
Recursive construction of multiple levels.
Tuesday, December 14, 2010
Randomized SP mechanisms
Mechanism (Random dictatorship): open the facility at with probability .
strategy-proof but not GSP 2-approximation
xi 1/n
Randomized SP mechanisms
Mechanism (Random dictatorship): open the facility at with probability .
strategy-proof but not GSP 2-approximation
xi 1/n
Can we do better?
Tuesday, December 14, 2010
Randomized SP mechanisms
Mechanism (Random dictatorship): open the facility at with probability .
strategy-proof but not GSP 2-approximation
xi 1/n
Can we do better?
Thm:
no randomized SP mechanism is better than approximation.
!
2 − 4 n1/3
"
Randomized SP mechanism
s t
1 1
Idea 2!
Tuesday, December 14, 2010
Randomized SP mechanism
s t
1 1
2!
Initially, “good” mechanism has prob. at 1/2 s, t
Idea
Randomized SP mechanism
s t
There exists an agent, let say blue, with cost at least 1 + !
1 1
2!
Initially, “good” mechanism has prob. at 1/2 s, t
Idea
Tuesday, December 14, 2010
Randomized SP mechanism
s t
There exists an agent, let say blue, with cost at least 1 + !
1 1
2!
Initially, “good” mechanism has prob. at 1/2 s, t
Idea
Randomized SP mechanism
s t
There exists an agent, let say blue, with cost at least 1 + !
1 1
2!
Initially, “good” mechanism has prob. at 1/2 s, t
Idea
Tuesday, December 14, 2010
Randomized SP mechanism
s t
There exists an agent, let say blue, with cost at least 1 + !
Move appropriate blue agents to , by strategy- proof argue that .
s ps ≈ pt ≈ 1/2
1 1
2!
Initially, “good” mechanism has prob. at 1/2 s, t
Idea
Randomized SP mechanism
s t
There exists an agent, let say blue, with cost at least 1 + !
Move appropriate blue agents to , by strategy- proof argue that .
s ps ≈ pt ≈ 1/2
1 1
2!
Initially, “good” mechanism has prob. at 1/2 s, t
Idea
Tuesday, December 14, 2010
Randomized SP mechanism
s t
There exists an agent, let say blue, with cost at least 1 + !
Move appropriate blue agents to , by strategy- proof argue that .
s ps ≈ pt ≈ 1/2
1 1
2!
Initially, “good” mechanism has prob. at 1/2 s, t
Idea
Randomized SP mechanism
s t
There exists an agent, let say blue, with cost at least 1 + !
Move appropriate blue agents to , by strategy- proof argue that .
s ps ≈ pt ≈ 1/2
1/2 · (1 + 2 + 2!) + 1/2 · (1 + 2!)
1 + 2! ≈ 2
Then:
1 + !
1 1
2!
Initially, “good” mechanism has prob. at 1/2 s, t
Idea
Tuesday, December 14, 2010
Conclusion
Open a constant facilities:
easy in term of optimization.
SP mechanism with bounded ratio?
Complete characterization of performance of randomized (G)SP mechanisms.
Thank you!
Tuesday, December 14, 2010