Local Smoothness and the Price of Anarchy in Atomic Splittable

Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Local Smoothness and the
Price of Anarchy in
Atomic Splittable Congestion Games
Tim Roughgarden
Florian Schoppmann
January 23, 2011
1
Introduction and Model
Local Smoothness
The Price of Anarchy
I
Conclusion
(Papadimitriou & Koutsoupias 1999)
Loss due to lack of central control
PoA =
I
Tight Bounds Via Local Smoothness
SC(equilibrium)
equilibrium SC(optimum)
sup
Analogs:
I
I
Approximation Ratio (lack of
computing power)
Competitive Ratio (lack of
knowledge about future)
(Nash) Equilibrium:
I
No player can improve by unilateral change of strategy
2
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
How Efficient Is a System?
Autonomous decision makers:
I Local optimization,
distributed
Central Control:
I Global optimization
I
Not efficient
I
Efficient (if it works)
I
No regret
I
Unfair
I
Simple, “fault-tolerant”
I
Hard to implement
If PoA is low, no need to tell people what to do
I
Allows for a simple system design
If PoA is high, need to re-design system or give incentives
I
Mechanism Design, e.g., toll pricing
3
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Key Points
I
Main Definition: A general framework (“local
smoothness”) to bound the price of anarchy in games
with convex strategy sets
I
Theorem: Every PoA bound proved via local smoothness
holds even for correlated equilibria (i.e., with vanishing
per-step swap regret)
I
I
But does not apply to coarse-correlated equilibria
Theorem: Local smoothness provides optimal bounds for
atomic splittable congestion games
I
solves problem that had been open for several years
(Nisan et al. 2007, Algorithmic Game Theory )
4
Introduction and Model
Local Smoothness
Congestion Games
Tight Bounds Via Local Smoothness
(Rosenthal 1973)
Model: Finite set E of resources (e.g., network edges)
I
Finite set of players N
I
Player i’s strategy set = collection Pi of
subsets of E (e.g., paths)
I
cost `e (xe ) per resource, xe = load on e
I
Player i’s cost = sum of resource costs
I
Objective: Sum of player costs
E
Conclusion
5
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Atomic Splittable Congestion Games
Individ.
1
Each player i:
I
I
I
has weight wi
i
chooses ~x s.t.
0
P
i
p∈Pi xp
1
= wi
i
e∈E xe
has cost ci (~x ) =
· `e (xe )
where xei is i’s load on e
0
1
Coalition
1
0
x .5 2
.5
1
P
Conclusion
0
1
SC(�x ) = 3.25
Objective:
SC(~x ) :=
X
e∈E
xe · `e (xe ) =
X
i∈[n]
ci (~x )
6
Introduction and Model
Local Smoothness
Review: Smoothness
Tight Bounds Via Local Smoothness
(Roughgarden 2009)
Game (λ, µ)-smooth if ∀ outcomes ~x , ~y :
(*)
n
X
i=1
ci (~y i , ~x −i ) ≤ λ · SC(~y ) + µ · SC(~x )
Fact: The PoA of any (λ, µ)-smooth game is ≤
λ
1−µ
Conclusion
7
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Smoothness Is Not Strong Enough
For convex games, we derive better upper bounds by requiring
smoothness only for outcomes ~y “arbitrarily close to” ~x :
(*) ci (~y i , ~x −i ) − ci (~x ) ≥ ∇i ci (~x )T (~y i − ~x i )
~x is NE iff variational inequality holds:
∇i ci (~x )T (~y i − ~x i ) ≥ 0
Local Smoothness: Replace lhs of (*) by rhs
�x i �y i
8
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Local Smoothness
Key Definition: Game is (λ, µ)-locally smooth w.r.t. ~y if ∀~x ,
(*)
n
X
ci (~x ) + ∇i ci (~x )T (~y i − ~x i ) ≤ λ · SC(~y ) + µ · SC(~x )
i=1
Assuming (λ, µ)-local smooth w.r.t. ~y and ~x pure NE:
SC(~x ) =
X
≤
X
ci (~x )
[Def. of SC]
i
i
ci (~x ) + ∇i ci (~x )T (~y i − ~x i ) [Var. ineq.]
≤ λ · SC(~y ) + µ · SC(~x )
Hence: PoA of pure NE ≤
λ
1−µ
[(*)]
9
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Intermezzo: Equilibrium Concepts
I
Pure Nash = outcome ~x :
∀i, ~y i ∈ Si : ci (~x ) ≤ ci (~y i , ~x −i )
I
Conclusion
coarse correlated
correlated
mixed Nash
pure
Nash
Mixed Nash = product distribution P:
∀i, ~y i ∈ Si : E~x ∼P [ci (~x )] ≤ E~x ∼P [ci (~y i , ~x −i )]
I
Correlated = distribution P:
∀i, δi : Si → Si : E~x ∼P [ci (~x )] ≤ E~x ∼P [ci (δi (~x i ), ~x −i )]
I
Coarse-correlated = distribution P:
∀i, ~y i ∈ Si : E~x ∼P [ci (~x )] ≤ E~x ∼P [ci (~y i , ~x −i )]
10
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Why important?
Tight connection to learning:
I
Correlated equilibria ∼ empirical distribution of sequences
with vanishing per-step swap regret
I
Coarse-correlated equilibria ∼ empirical distribution of
sequences with vanishing per-step regret
Reduced assumptions:
I
Bound applies to repeated-play sequences that do not
converge in any sense
I
Nash equilibria in general hard to compute
I
Simple algorithms for regret-minimization exist
11
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Robustness of (Local) Smoothness Bounds
I
Roughgarden (2009):
Smoothness bounds all
coarse-correlated equilibria
I
Theorem: Local smoothness
bounds all correlated equilibria
I
Theorem: Local smoothness
does not bound
coarse-correlated equilibria
coarse correlated
correlated
mixed Nash
pure
Nash
12
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
13
Local Smoothness Bounds All Correlated Equilibria
Key Claim:
(*)
E~x ∼P ∇i ci (~x )T (~y i − ~x i ) ≥ 0
Proof of Key Claim (by contradiction, assume lhs of (*) < 0):
I
Define ~x := ((1 − ) · ~x i + · ~y i , ~x −i )
I
By dominated convergence theorem:
Z
Z
ci (~x ) − ci (~x )
lim
dP(~x ) = ∇i ci (~x )T (~y i − ~x i ) dP(~x )
&0
= E~x ∼P [∇i ci (~x )T (~y i − ~x i )] < 0
I
For small enough , E~x ∼P [ci (~x )] < E~x ∼P [ci (~x )]
E
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Local Smoothness Does Not Bound All CCEs
Example:
I
I
n = 2, strat. sets = [0, 1], c1 (~x ) = c2 (~x ) = (x1 − x2 )2 + P(0, α) = P(1, 1 − α) =
1
2
defines a CCE (for α ≤ 14 )
1
1–α
α
0
0
1
14
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Tight Bounds for Atomic Splittable CGs
Let γ(L) := smallest upper bound on PoA provable via local
smoothness, for set of cost functions L
Theorem: The PoA in every atomic splittable congestion game
is at most γ(L). This bound is tight.
Previous work:
I
Cominetti et al. (2006): Special case of our local
smoothness framework (λ fixed at 1)
I
Harks (2007): Tight upper bounds (tight by our results)
15
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Core of upper-bound proof
Find minimal
(*)
λ
1−µ
y · `(x) +
s.t. ∀` ∈ L, x ≥ 0, y > 0 with `(y ) > 0:
y2 0
· ` (x) ≤ λ · y · `(y ) + µ · x · `(x)
4
We define:
2
y · `(x) + y4 · `0 (x) − µ · x · `(x)
g`,x,y (µ) :=
y · `(y ) · (1 − µ)
Then (*) is equivalent to g`,x,y (µ) ≤
λ
.
1−µ
Best upper bound provable using local smoothness:
γ(L) = inf
µ∈[0,1)
sup
`∈L
x≥0,y >0,`(y )>0
g`,x,y (µ)
16
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Plots of Functions g`,x,y (µ) When ` Fixed
λ
1−µ
γ({�})
x · �(x)
2
= y · �(x) + y4 · �� (x)
µ
0
0.5
1
17
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Lower-Bound Construction for Special Case (I)
Here: Special case `, x, y with x · `(x) = y · `(x) +
y2
4
· `0 (x)
For concreteness: `(z) = z 3 , x = 32 , y = 1, q := 2x − 1 ∈ N
Let λ, µ arbitrary s.t. g`,x,y (µ) =
λ
.
1−µ
Family of instances:
I
n players with weight ≈ 1
I
n resources
I
Each player: 1 “optimal”,
n−1
“non-opt.” strategies
q
I
In NE ~x , players put α on
opt, β on non-opt.
“non-optimal”
strategies
“optimal”
strategy
18
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Lower-Bound Construction for Special Case (II)
For Nash equilibrium ~x , we need:
I
The load on each resource is exactly x:
α + (n − 1) · β = x
I
Marginal costs are equal on used strategies:
`(x) + α · `0 (x) = q · (`(x) + β · `0 (x))
I
On each resource, one player puts load 12 ± o(1) and all
others put load o(1). Solving the above equations gives:
α=
n − 1 q · x n→∞ y
+
−−−→
2n
n
2
and β =
x − α n→∞
−−−→ 0
n−1
19
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Lower-Bound Construction for Special Case (III)
For the optimum ~y , we need:
I
If every player only uses its “optimal” strategy, each
resource has load y + o(1). The weight of any player is:
n−1
x − α n→∞
−−−→ 1
α+
·β =α+
q
2x − 1
h
i
n→∞ 1
x−α
q = 2x − 1, β = n−1 , α −−−→ 2
In NE ~x , social cost contributed by each resource is
x · `(x) = λ · y · `(y ) + µ · x · `(x),
i.e.,
λ
1−µ
times as in optimum ~y . =⇒ PoA = ( 32 )4
[x · `(x) = y · `(x) +
y2
4
· `0 (x), g`,x,y (µ) =
λ
]
1−µ
20
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Comparison (Polynomials)
Degree
Atomic
splittable
1
2
3
4
5
6
7
8
d
(Roughg. 2003, Aland et al. 2006)
Atomic
unsplittable
(weighted)
1.500
2.618
2.549
9.909
5.063
47.82
11.09
277.0
26.32
1,858
66.88
14,099
180.3
118,926
512.0
1,101,126
√
( 1+
d+1 d+1
)
2
Conclusion
Θ( logd d )d+1
Nonatomic
1.333
1.626
1.896
2.151
2.394
2.630
2.858
3.081
Θ( logd d )
21
Introduction and Model
Local Smoothness
Tight Bounds Via Local Smoothness
Conclusion
Take-Home Points
I
Market power can be harmful (“price of collusion”)
I
Essentially all PoA results for congestion games conform
to the (local) smoothness framework
I
Local smoothness can also be used to reprove the PoA for
resource allocation games (Johari & Tsitsiklis 2004)
Remaining open problems:
I
PoA for coarse-correlated equilibria
I
PoA in symmetric games (different player weights)
22