Chapter 9: The Price of Anarchy

Algorithmic Game Theory
PoA and PoSPigou's networkThe 4/3 boundSmoothnessPoA in the Braess networkLoad balancingPrice of stabilityCongestion tolls
Sercan KÜLCÜ · Giresun University · Computer Engineering

Basic Concepts

Warm-up: The Load Balancing Game

The rules

Jobs arrive one by one and each joins server A, B or C; a job's completion time is the server's total load divided by its speed.

Think about it
  • If every job minimizes its own completion time, is the system's total cost minimized too?
  • How much worse than the best assignment can an equilibrium of selfish jobs be?
A (speed 2)time 4 / 2 = 2B (speed 1)time 2 / 1 = 2C (speed 1)time 1 / 1 = 1new job: A, B, C?Completion time = server's total load ÷ speed

Basic Concepts

Learning Outcomes

Define

Define the Price of Anarchy (PoA) and the Price of Stability (PoS); show the difference with an example.

Compute

Find equilibrium and optimal costs in the Pigou, Braess and load balancing examples and compute the ratio.

Explain the bound

Explain the intuitive proof of PoA ≤ 4/3 for linear latencies and why it is tight.

Improve

Show how a congestion toll turns the optimum into an equilibrium in the two-route network and Pigou's network.

Basic Concepts

The Price of Anarchy: The Bill for Selfishness

  • The Price of Anarchy (PoA) measures the efficiency lost between the worst equilibrium that selfish behavior produces and the social optimum.
  • In cost problems, the social optimum minimizes total cost.
  • In a traffic network, total delay can rise while each user tries to cut their own travel time.
  • Koutsoupias and Papadimitriou (1999) introduced the idea as the “coordination ratio”; Papadimitriou (2001) named it the “Price of Anarchy”: the cost of everyone deciding alone (anarchy) instead of a central “dictator”.
Example: in the two-route network the Wardrop equilibrium costs 75,000 and the system optimum 71,250 car·min (ratio 75,000 / 71,250 ≈ 1.053); in the Braess network the new road raised travel time from 65 to 80 min. The Price of Anarchy measures and bounds such gaps with a single number.

Basic Concepts

Nash Equilibrium and Social Optimum

Nash equilibrium (individual)

  • No player can improve their outcome by changing strategy unilaterally.
  • Reflects individual incentives.
  • There may be several, good and bad.

Social optimum (system)

  • Minimizes total cost or maximizes social welfare.
  • Reflects the system's common goal.
  • Usually not an equilibrium: someone wants to deviate.
Why do they differ? Individual rationality (everyone looks after their own payoff), externalities (costs imposed on others are ignored) and missing information. The Price of Anarchy measures the ratio between the two.

Basic Concepts

Definitions: PoA and PoS

Price of Anarchy
PoA = maxs ∈ NE C(s) / C(OPT)
worst equilibrium: a pessimistic guarantee
Price of Stability
PoS = mins ∈ NE C(s) / C(OPT)
best equilibrium: the optimistic scenario

NE: the game's Nash equilibria; C(s): total cost of state s; OPT: the state of minimum total cost (social optimum). PoA = cost of the worst equilibrium ÷ cost of the optimum; PoS = cost of the best equilibrium ÷ cost of the optimum.

OPTsocial optimumbest NE→ PoSworst NE→ PoAcost →
  • With C(OPT) > 0, always 1 ≤ PoS ≤ PoA. With a unique equilibrium they are equal.
  • In payoff (welfare) problems the ratio is flipped: SW(OPT) / SW(equilibrium) (SW: social welfare), so a larger value still means worse efficiency.

Basic Concepts

Measuring and Reading Efficiency

  • Total cost: the sum of all users' delay or resource-use costs.
  • In cost problems, the PoA is the ratio of the worst equilibrium cost to the social optimum cost.
  • A PoA of 1 means no efficiency loss.
  • The larger the ratio, the more strategic behavior costs.
PoAMeaningExample
1no lossBraess, before the new road
4/333% extra costPigou, the linear bound
2 − 2/(m+1)< 2 timesload balancing
unboundedno guaranteesteep latency (xp)
Why the worst case? The PoA is a guarantee that holds without knowing which equilibrium players reach: “whatever happens, the loss is at most this much.” It is the game-theoretic counterpart of an algorithm's approximation ratio.

Basic Concepts

Two Worst-Case Ratios: What Is Missing?

RatioWhat blocks the optimumNumerator / denominator
Approximation ratiocomputational power: a problem not solvable in polynomial time (NP-hard)solution found by a fast algorithm ÷ optimum
Price of Anarchycoordination: no central authority, everyone pursues their own interestworst equilibrium / optimum
Same problem, two losses

Makespan on m identical machines: Graham's greedy list algorithm is at most 2 − 1/m times worse; a pure equilibrium of selfish jobs at most 2 − 2/(m+1) times. For m = 2: 3/2 and 4/3.

Limits of the analysis

Real players are not purely selfish: trust, social norms, reputation and repeated interaction can foster cooperation. So the PoA is a pessimistic guarantee, not a prediction.

Pigou's Network

Step by Step: The Price of Anarchy in Pigou's Network

stℓ(x) = xℓ(x) = 11 unit of traffic0.50.75100.250.50.751x (share on top road)total cost C(x)

Setup (Pigou, 1920): 1 unit of traffic, many small drivers. Top road: delay x (the share on top); bottom road: always 1.

stℓ(x) = xℓ(x) = 1top: 1bottom: 00.50.75100.250.50.751x (share on top road)total cost C(x)eq.: 1

Equilibrium: the top road is never slower than the bottom (x ≤ 1) → everyone switches to the top. x = 1, everyone takes 1 min: C(eq) = 1.

stℓ(x) = xℓ(x) = 1top: 1bottom: 00.50.75100.250.50.751x (share on top road)total cost C(x)C(x) = x² − x + 1eq.: 1

Total cost: share x on top takes x min, share 1 − x on the bottom takes 1 min → C(x) = x·x + (1 − x)·1 = x² − x + 1.

stℓ(x) = xℓ(x) = 1top: 0.5bottom: 0.50.50.75100.250.50.751x (share on top road)total cost C(x)C(x) = x² − x + 1eq.: 1optimum: 3/4

Minimum: C′(x) = 2x − 1 = 0 → x = 1/2. Half the traffic is “sacrificed” to the bottom road.

stℓ(x) = xℓ(x) = 1top: 0.5bottom: 0.50.50.75100.250.50.751x (share on top road)total cost C(x)C(x) = x² − x + 1eq.: 1optimum: 3/4

C(1/2) = 1/4 + 1/2 = 3/4. Top drivers take 1/2 min, bottom drivers 1 min; the bottom drivers want to switch → the optimum is not an equilibrium.

stℓ(x) = xℓ(x) = 1top: 1bottom: 00.50.75100.250.50.751x (share on top road)total cost C(x)C(x) = x² − x + 1eq.: 1optimum: 3/4

PoA = C(eq) / C(opt) = 1 / (3/4) = 4/3 ≈ 1.33: selfish routing produces 33% more total delay.

Pigou's Network

Why the Difference? Private vs. Marginal Cost

00.511.5200.51x (share on top road)mineq.optimum2x (marginal)x (private)bottom road: 1
  • A driver looks at their private cost: compares x with 1 → x = 1.
  • Society looks at marginal cost: the top road's total cost is x·x with derivative 2x; 2x = 1 → x = 1/2.
  • The gap is x·ℓ′(x): the delay one driver imposes on the others (an externality).
General principle: at equilibrium the delays of used paths are equal. At the optimum, “my own delay + the delay I add for others” is equalized. With f cars on a road and delay ℓ(f), the added delay is f·ℓ′(f) (ℓ′: the derivative of the delay). This gap is the basis for congestion tolls.

Linear Latency and Braess

Steeper Latency, Larger PoA

00.5100.51xℓ(x)xx¹⁶
p (ℓ = xp)xOPTC(OPT)PoA
10.50.751.333
20.5770.6151.626
40.6690.4652.151
80.760.3253.081
160.8380.2124.727
The equilibrium always sends everyone to the top road (cost 1). As latency gets steeper (larger p) the optimum improves and the PoA grows slowly but without bound: about 4.7 at p = 16. So 4/3 is not universal; it is specific to linear latency.

Linear Latency and Braess

Latency Exponent and PoA in Pigou's Network

Top road xp, bottom road 1  ·  C(x) = xp+1 + (1 − x)  ·  C′(x) = 0 ⇒ x* = (p + 1)−1/p  ·  PoA = 1 / C(x*)

As the latency function gets steeper, the optimum improves, the equilibrium cost stays at 1 and the PoA grows without bound; with linear latency (p = 1), PoA = 4/3.

Linear Latency and Braess

Code Trace: pigou_poa(p)

00.5100.51x →C(x) = x2 + (1 − x)eq.p = 1xOPT = 0.5

p = 1 (Pigou): C(x) = x² + 1 − x, C′(x) = 2x − 1 = 0 ⇒ xOPT = 2−1 = 0.5.

00.5100.51x →C(x) = x2 + (1 − x)eq.p = 1xOPT = 0.5C(OPT) = 0.75PoA = 1 / C(OPT) = 1.333

C(OPT) = 0.25 + 0.5 = 0.75 ⇒ PoA = 4/3: Pigou's classic result.

00.5100.51x →C(x) = x3 + (1 − x)eq.p = 2xOPT = 0.577

p = 2: C(x) = x·x2 + (1 − x)·1, C′(x) = 3x2 − 1 = 0 ⇒ xOPT = 3−1/2 ≈ 0.577.

00.5100.51x →C(x) = x3 + (1 − x)eq.p = 2xOPT = 0.577C(OPT) = 0.615PoA = 1 / C(OPT) = 1.626

The curve got steeper: the optimum fell to 0.615 while the equilibrium is still 1 ⇒ PoA 1.626.

00.5100.51x →C(x) = x5 + (1 − x)eq.p = 4xOPT = 0.669

p = 4: C(x) = x·x4 + (1 − x)·1, C′(x) = 5x4 − 1 = 0 ⇒ xOPT = 5−1/4 ≈ 0.669.

00.5100.51x →C(x) = x5 + (1 − x)eq.p = 4xOPT = 0.669C(OPT) = 0.465PoA = 1 / C(OPT) = 2.151

p = 4: at the optimum 67% of the traffic is on top; PoA 2.151 > 2.

00.5100.51x →C(x) = x9 + (1 − x)eq.p = 8xOPT = 0.76

p = 8: C(x) = x·x8 + (1 − x)·1, C′(x) = 9x8 − 1 = 0 ⇒ xOPT = 9−1/8 ≈ 0.76.

00.5100.51x →C(x) = x9 + (1 − x)eq.p = 8xOPT = 0.76C(OPT) = 0.325PoA = 1 / C(OPT) = 3.081

PoA 3.081: slow growth, but unbounded.

00.5100.51x →C(x) = x17 + (1 − x)eq.p = 16xOPT = 0.838

p = 16: C(x) = x·x16 + (1 − x)·1, C′(x) = 17x16 − 1 = 0 ⇒ xOPT = 17−1/16 ≈ 0.838.

00.5100.51x →C(x) = x17 + (1 − x)eq.p = 16xOPT = 0.838C(OPT) = 0.212PoA = 1 / C(OPT) = 4.727

p = 16: PoA 4.727. The 4/3 guarantee holds only for linear latency.

def pigou_poa(p):          # top x**p, bottom 1
  x = (p + 1) ** (-1 / p)  # where C'(x) = 0
  opt = x ** (p + 1) + (1 - x)
  return 1 / opt           # eq. cost is 1

for p in [1, 2, 4, 8, 16]:
  print(p, round(pigou_poa(p), 3))

Linear Latency and Braess

The 4/3 Upper Bound for Linear Latency

Theorem (Roughgarden–Tardos, 2002). If latencies are linear (ℓe(x) = aex + be, a, b ≥ 0), then in every network with every demand: C(eq) ≤ 4/3 · C(opt).
Notation. fe: equilibrium traffic on edge e; feOPT: optimal traffic. Total delay sums “traffic × delay” over edges: C(f) = Σ ℓe(fe)·fe.
1. Equilibrium condition. At equilibrium everyone uses a shortest path. Freeze the delays and move drivers to their optimal paths: nobody finds a shorter one: Σ ℓe(fe)·fe ≤ Σ ℓe(fe)·feOPT.
The left side is the equilibrium's real cost C(f). The right side is a thought experiment: roads as slow as at equilibrium, but traffic spread as in the optimum. It remains to bound this right side by C(fOPT).

Linear Latency and Braess

The 4/3 Bound: Edge Inequality and Summation

2. Edge inequality. For each edge: ℓe(fe)·feOPT ≤ ℓe(feOPT)·feOPT + ¼·ℓe(fe)·fe. The trial cost exceeds the optimum's cost by at most ¼ of the equilibrium's.
Why is it true? Put ℓ = ax + b. The b terms: b·fOPT ≤ b·fOPT + ¼·b·f. The a terms need a·f·fOPT ≤ a·(fOPT)² + a·f²/4, i.e. a·(fOPT − f/2)² ≥ 0: a square.
Example (Pigou, the ℓ = x road): f = 1, fOPT = ½. Left: 1·½ = ½. Right: ½·½ + ¼·1·1 = ½. Equal.
3. Sum. Sum step 2 over edges, use step 1: C(f) ≤ C(fOPT) + ¼·C(f). Move ¼·C(f) left: ¾·C(f) ≤ C(fOPT) → C(f) / C(fOPT) ≤ 4/3.
Pigou attains it: the bound is tight. With constant latencies in the class, the worst case is always a two-link network like Pigou's (Roughgarden, 2003): topology does not raise the PoA.

Linear Latency and Braess

Testing the 4/3 Bound

ExampleEquilibriumOptimumObserved ratioUpper bound
Pigou (x and 1)13/44/3 (tight)4/3
Braess (4,000 cars)320,000258,75080 / 64.69 ≈ 1.2374/3
Braess normalized (x, 1, 1, x, 0)23/24/3 (tight)4/3
Two routes, 3,000 cars75,00071,2501.0534/3
20,000 random parallel linear networks——max 1.334/3
Load Balancing Game, Level 21312.51.045/2 (atomic)
An upper bound is the worst case over a class (e.g. all linear-latency networks); single instances usually do better. In 20,000 random networks (2–4 links, random a, b, demand) the ratio never exceeded 4/3 and neared it in Pigou-like cases.

Linear Latency and Braess

One Recipe: Smoothness

Thought experiment. Let s be a state (each player has a strategy) and sOPT the optimum. What would player i pay if only i switched to their sOPT strategy while the others stay in s? This is i's “trial cost”.
Notation. s−i: the strategies of players other than i. Ci(siOPT, s−i): i's trial cost. C(s) = Σi Ci(s): total cost.
Definition (Roughgarden, 2009). Fix constants λ > 0 and 0 ≤ μ < 1. A game is (λ, μ)-smooth if Σi Ci(siOPT, s−i) ≤ λ·C(sOPT) + μ·C(s) for every s and sOPT.
In words: the trial costs add up to at most λ times the optimum's cost plus μ times the cost of s. The smaller λ and μ, the better the resulting bound.

Linear Latency and Braess

From Smoothness to a PoA Bound: Four Steps

1. Equilibrium. s is a Nash equilibrium: nobody gains by deviating alone. For every i, Ci(s) ≤ Ci(siOPT, s−i).
2. Sum. Add these inequalities over all players: C(s) ≤ Σi Ci(siOPT, s−i).
3. Apply smoothness. The right side is at most λ·C(sOPT) + μ·C(s), so C(s) ≤ λ·C(sOPT) + μ·C(s).
4. Rearrange. Move μ·C(s) left: (1 − μ)·C(s) ≤ λ·C(sOPT). Divide both sides by (1 − μ)·C(sOPT): C(s) / C(sOPT) ≤ λ / (1 − μ). Since s was any equilibrium, PoA ≤ λ / (1 − μ).
Familiar cases. The edge inequality in the 4/3 proof means λ = 1, μ = ¼: 1 / (1 − ¼) = 4/3. In atomic (indivisible-player) linear congestion games λ = 5/3, μ = 1/3: (5/3) / (2/3) = 5/2.
Why does it matter? The proof uses only “nobody deviates alone”, so the same bound holds for weaker equilibrium notions too (e.g. mixed equilibria): the robust PoA.

Linear Latency and Braess

Atomic Congestion Games: PoA ≤ 5/2

z (equilibrium load)y(sOPT)01234500149162512002612214965693362822181616468574841363351109684746660

Cell: 3 × (right − left), avoiding fractions. No negative cells: it holds (checked for y, z ≤ 60); orange 0s mark equality.

  • Each player is one unit of load; edge cost ce(x) = aex + be.
  • ze, ye: players on e at equilibrium and at the optimum. A lone switcher to their optimal path meets at most ze + 1 players on e.
  • Key inequality (integers): y(z + 1) ≤ 5/3·y² + 1/3·z². With c(x) = x: left is the trial cost; y², z² the optimal and equilibrium costs.
  • ⇒ (5/3, 1/3)-smooth ⇒ PoA ≤ (5/3)/(2/3) = 5/2; the bound is tight (Awerbuch–Azar–Epstein; Christodoulou–Koutsoupias, 2005).
Splittable traffic gives 4/3, indivisible players 5/2: the delay a player adds itself (the "+1") enlarges the loss.

Linear Latency and Braess

Step by Step: PoA in the Braess Network

x/1004,000 cars450 cars450 carsx/1004,000 cars04,000 carssABtEquilibrium with the new roadall s→A→B→t: 80 minC(eq) = 4,000 · 80= 320,000 car·min

Once A→B opens everyone zigzags, 80 min each. Total 320,000 car·min.

x/1002,000 cars452,000 cars452,000 carsx/1002,000 cars00 carssABtBefore the new road: 65 min80 / 65 ≈ 1.231This is the “Braess loss”;not the PoA: other network

A common comparison is 80/65 ≈ 1.23. But the PoA's denominator is the optimum of the same network; can the new network do better than 65?

x/1002,000 cars452,000 cars452,000 carsx/1002,000 cars00 carssABtMarginal costs at z = 0:zigzag: 2·20 + 2·20 = 80old path: 2·20 + 45 = 85→ some zigzag helps!

On s→A with 2,000 cars the delay is 20 min; marginal cost ℓ + x·ℓ′ = 20 + 2,000·(1/100) = 40 min. The zigzag uses two such edges: 40 + 0 + 40 = 80; an old path 40 + 45 = 85. 80 < 85: even a planner sends some cars on the zigzag.

x/1002,250 cars451,750 cars451,750 carsx/1002,250 cars0500 carssABtSymmetric: a cars per old pathC(a) = 2(4000 − a)²/100 + 90aC′ = 0 → a = 1,750, z = 500C(opt) = 258,750

Optimum: 1,750 on each old path, 500 on the zigzag (marginal costs equal at 90). Average ≈ 64.69 min; the script searched all splits.

x/1002,250 cars451,750 cars451,750 carsx/1002,250 cars0500 carssABtPoA = 320,000 / 258,750≈ 1.237≤ 4/3 ✓normalized: exactly 4/3

PoA = 320,000 / 258,750 ≈ 1.237 ≤ 4/3. In the classic normalized Braess network (x, 1, 1, x, 0) the ratio is exactly 2 / (3/2) = 4/3.

Load Balancing

The Load Balancing Game

  • n jobs (players), m identical machines; job j has size wj.
  • Each job picks a machine; its cost is that machine's total load (its completion time).
  • Social cost: the makespan (maximum load), i.e. when all jobs are done.
  • Applications: server clusters, cloud task placement, processor cores. Koutsoupias–Papadimitriou first defined the PoA in this model (1999).
i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2makespan = 4i3, i4 cost: 2
This game always has a pure Nash equilibrium: when a job moves to a less loaded machine, the load list sorted in decreasing order gets lexicographically “smaller” (a smaller number at the first difference). This cannot go on forever; the process stops at a Nash equilibrium.

Load Balancing

Step by Step: PoA in the Load Balancing Game

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 24 jobs: sizes 2, 2, 1, 12 identical machinesJob cost: its machine's loadSocial cost: makespan

The load balancing game: each job (player) picks a machine and wants to minimize its completion time, i.e. its machine's total load.

i1: 2i2: 2i3: 1i4: 1M1: 3M2: 3makespan 3Optimum: {2, 1} | {2, 1}loads 3 and 3makespan = 3

Social optimum: total work 6 on two machines → best possible 3 and 3. Makespan 3.

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2makespan 4Assignment: {2, 2} | {1, 1}loads 4 and 2makespan = 4

Another assignment: both big jobs on M1 (load 4), both small jobs on M2 (load 2). Can this be an equilibrium?

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2makespan 4If i1 (2) moves to M2:2 + 2 = 4, now 4→ no gainIf i3 (1) moves to M1: 4 + 1 = 5 > 2

Deviation check: a big job moving to M2 still waits 4; a small job moving to M1 waits 5. Nobody gains → a Nash equilibrium.

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2makespan 4Worst equilibrium: 4Best equilibrium: 3 (optimum is an eq.)PoA = 4/3, PoS = 1Script: 6 of 16 assignments are eq.

Here PoA = 4/3 and PoS = 1: the optimum is an equilibrium, but so is a bad assignment.

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2makespan 4m identical machines, pure eq.:PoA = 2 − 2/(m + 1)m = 2 → 4/3 (this example is tight)Mixed equilibria: grows very slowly in m

General result (Finn–Horowitz, 1979; Koutsoupias–Papadimitriou, 1999): for pure equilibria PoA = 2 − 2/(m+1) < 2. For mixed equilibria the PoA grows very slowly but without bound in m (Koutsoupias–Mavronicolas–Spirakis, 2003; Czumaj–Vöcking, 2002).

Load Balancing

Code Trace: is_nash(sizes, assign, m)

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2i1, i2 → M1; i3, i4 → M2

Call: job sizes and each job's machine (0 = M1, 1 = M2).

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load = [0, 0]

The load array is zeroed.

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load = [2, 0]

j = 0: i1 (size 2) is added to M1 → load = [2, 0].

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load = [4, 0]

j = 1: i2 (size 2) is added to M1 → load = [4, 0].

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load = [4, 1]

j = 2: i3 (size 1) is added to M2 → load = [4, 1].

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load = [4, 2]

j = 3: i4 (size 1) is added to M2 → load = [4, 2].

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load[1] + sizes[0] = 4load[0] = 4 → stays

j = 0, h = 1: if i1 moves to M2, 2 + 2 = 4; now 4. 4 < 4? No.

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load[1] + sizes[1] = 4load[0] = 4 → stays

j = 1, h = 1: if i2 moves to M2, 2 + 2 = 4; now 4. 4 < 4? No.

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load[0] + sizes[2] = 5load[1] = 2 → stays

j = 2, h = 0: if i3 moves to M1, 4 + 1 = 5; now 2. 5 < 2? No.

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2load[0] + sizes[3] = 5load[1] = 2 → stays

j = 3, h = 0: if i4 moves to M1, 4 + 1 = 5; now 2. 5 < 2? No.

i1: 2i2: 2i3: 1i4: 1M1: 4M2: 2makespan 4is_nash → True

No job could escape to a better machine → True: {2, 2} | {1, 1} is a Nash equilibrium (makespan 4).

def is_nash(sizes, assign, m):
  load = [0] * m
  for j, k in enumerate(assign):
    load[k] += sizes[j]
  for j, k in enumerate(assign):
    for h in range(m):
      if load[h] + sizes[j] < load[k]:
        return False      # j moves
  return True             # nobody moves

For h = k (its own machine) the condition is always false; the steps show only h ≠ k.

Load Balancing

Brute Force: All 16 Assignments

Split (M1 | M2, symmetric)AssignmentsMakespanNash?
{2, 1} | {2, 1}43✓ eq.
{2, 1, 1} | {2}44✗
{2, 2} | {1, 1}24✓ eq.
{2, 2, 1} | {1}45✗
{2, 2, 1, 1} | { }26✗
The script tried all 24 = 16 assignments and every job's deviation in each. The equilibria are those with makespan 3 and 4 → PoA = 4/3 (worst equilibrium), PoS = 1 (best equilibrium = optimum).

Best-response dynamics in this game always stop at an equilibrium: every profitable move shrinks the sorted load list. Which equilibrium is reached depends on the start and the order of moves.

Load Balancing

Code Trace: PoA and PoS by Trying All Assignments

i1:2i2:2i3:1i4:1M1load 6M2load 01/16 assign = (0, 0, 0, 0)makespan 6 · Nash? ✗opt = 6 · eqs = []

(0, 0, 0, 0): all on M1, loads [6, 0]; i1 moving to M2 gets 2 < 6 ⇒ not an equilibrium. opt is 6 for now.

i1:2i2:2i3:1i4:1M1load 5M2load 12/16 assign = (0, 0, 0, 1)makespan 5 · Nash? ✗opt = 5 · eqs = []

(0, 0, 0, 1): loads [5, 1]; i1 moving to the other machine gets 3 < 5 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 5M2load 13/16 assign = (0, 0, 1, 0)makespan 5 · Nash? ✗opt = 5 · eqs = []

(0, 0, 1, 0): loads [5, 1]; i1 moving to the other machine gets 3 < 5 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 4M2load 24/16 assign = (0, 0, 1, 1)makespan 4 · Nash? ✓opt = 4 · eqs = [4]

(0, 0, 1, 1): loads [4, 2], nobody can escape ⇒ equilibrium; makespan 4 is added to the list.

i1:2i2:2i3:1i4:1M1load 4M2load 25/16 assign = (0, 1, 0, 0)makespan 4 · Nash? ✗opt = 4 · eqs = [4]

(0, 1, 0, 0): loads [4, 2]; i3 moving to the other machine gets 3 < 4 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 3M2load 36/16 assign = (0, 1, 0, 1)makespan 3 · Nash? ✓opt = 3 · eqs = [4, 3]

(0, 1, 0, 1): loads [3, 3], nobody can escape ⇒ equilibrium; makespan 3 is added to the list.

i1:2i2:2i3:1i4:1M1load 3M2load 37/16 assign = (0, 1, 1, 0)makespan 3 · Nash? ✓opt = 3 · eqs = [4, 3, 3]

(0, 1, 1, 0): loads [3, 3], nobody can escape ⇒ equilibrium; makespan 3 is added to the list.

i1:2i2:2i3:1i4:1M1load 2M2load 48/16 assign = (0, 1, 1, 1)makespan 4 · Nash? ✗opt = 3 · eqs = [4, 3, 3]

(0, 1, 1, 1): loads [2, 4]; i3 moving to the other machine gets 3 < 4 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 4M2load 29/16 assign = (1, 0, 0, 0)makespan 4 · Nash? ✗opt = 3 · eqs = [4, 3, 3]

(1, 0, 0, 0): loads [4, 2]; i3 moving to the other machine gets 3 < 4 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 3M2load 310/16 assign = (1, 0, 0, 1)makespan 3 · Nash? ✓opt = 3 · eqs = [4, 3, 3, 3]

(1, 0, 0, 1): loads [3, 3], nobody can escape ⇒ equilibrium; makespan 3 is added to the list.

i1:2i2:2i3:1i4:1M1load 3M2load 311/16 assign = (1, 0, 1, 0)makespan 3 · Nash? ✓opt = 3 · eqs = [4, 3, 3, 3, 3]

(1, 0, 1, 0): loads [3, 3], nobody can escape ⇒ equilibrium; makespan 3 is added to the list.

i1:2i2:2i3:1i4:1M1load 2M2load 412/16 assign = (1, 0, 1, 1)makespan 4 · Nash? ✗opt = 3 · eqs = [4, 3, 3, 3, 3]

(1, 0, 1, 1): loads [2, 4]; i3 moving to the other machine gets 3 < 4 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 2M2load 413/16 assign = (1, 1, 0, 0)makespan 4 · Nash? ✓opt = 3 · eqs = [4, 3, 3, 3, 3, 4]

(1, 1, 0, 0): loads [2, 4], nobody can escape ⇒ equilibrium; makespan 4 is added to the list.

i1:2i2:2i3:1i4:1M1load 1M2load 514/16 assign = (1, 1, 0, 1)makespan 5 · Nash? ✗opt = 3 · eqs = [4, 3, 3, 3, 3, 4]

(1, 1, 0, 1): loads [1, 5]; i1 moving to the other machine gets 3 < 5 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 1M2load 515/16 assign = (1, 1, 1, 0)makespan 5 · Nash? ✗opt = 3 · eqs = [4, 3, 3, 3, 3, 4]

(1, 1, 1, 0): loads [1, 5]; i1 moving to the other machine gets 3 < 5 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 0M2load 616/16 assign = (1, 1, 1, 1)makespan 6 · Nash? ✗opt = 3 · eqs = [4, 3, 3, 3, 3, 4]

(1, 1, 1, 1): loads [0, 6]; i1 moving to the other machine gets 2 < 6 ⇒ not an equilibrium.

i1:2i2:2i3:1i4:1M1load 0M2load 616/16 assign = (1, 1, 1, 1)makespan 6 · Nash? ✗opt = 3 · eqs = [4, 3, 3, 3, 3, 4]

Equilibria [4, 3, 3, 3, 3, 4] ⇒ PoA = 4/3, PoS = 3/3 = 1. There are mn assignments: exponential for large games.

from itertools import product
def poa_pos(sizes, m):
  opt, eqs = float("inf"), []
  for assign in product(range(m), repeat=len(sizes)):
    load = [0] * m
    for j, k in enumerate(assign):
      load[k] += sizes[j]
    opt = min(opt, max(load))
    if is_nash(sizes, assign, m):    # eq. test
      eqs.append(max(load))
  return max(eqs) / opt, min(eqs) / opt
r = poa_pos([2, 2, 1, 1], 2)
print(round(r[0], 3), r[1])

Load Balancing

Machines with Different Speeds (Related Machines)

  • Machine i has speed si; a job's completion time is its machine's total load divided by si. A pure Nash equilibrium still always exists.
  • Example: speeds A = 2, B = 1, C = 1; job sizes 2, 1, 1. Social cost: the maximum completion time.
AssignmentTimes (A, B, C)MaxNash?
A {2}, B {1}, C {1}1, 1, 11yes (optimum)
A {1, 1}, B {2}, C { }1, 2, 02yes

In the second assignment, the size-2 job would wait (2 + 2)/2 = 2 on A or 2/1 = 2 on C; a size-1 job moving to C still waits 1: nobody gains. Ratio 2/1 = 2, above the bound 2 − 2/4 = 1.5 for three identical machines.

On identical machines the loss of a pure equilibrium stays below 2 (2 − 2/(m+1)). With different speeds the loss grows without bound (if very slowly) as machines are added (Czumaj–Vöcking, 2002). Levels 2 and 3 of the Load Balancing game use this model.

Load Balancing

Back to the Warm-up: The Load Balancing Game

The game measures something else

Here social cost is the sum of completion times, not the makespan. Level 2, CPU Scheduling: 7 units, speeds 2, 1, 1.

Loads (A, B, C)Total timeNash?
(3, 2, 2)3·1.5 + 2·2 + 2·2 = 12.5yes (optimum)
(4, 2, 1)4·2 + 2·2 + 1·1 = 13yes (worst)

PoA = 13 / 12.5 = 1.04. In (4, 2, 1), a job moving from A to C still waits 2: nobody gains. The mirror (4, 1, 2) costs 13 too.

Where does selfish placement lead?

If each arriving job picks its fastest server (ties alphabetical), the order is A, A, B, C, A, A, B: exactly the worst equilibrium (4, 2, 1). The game's selfish view uses this greedy placement.

Level 1 (6 jobs, identical servers): the only equilibrium (2, 2, 2) is optimal, PoA = 1. Loss appears once servers differ.

The 4/3 for {2, 2, 1, 1} measures makespan, the game total time: the PoA depends on the social cost used.

The Price of Stability

PoA ≠ PoS in the Same Game

stA: 2 TLB: 6 TL1234link cost is split equallyamong its users

4 players rent a link from s to t: A costs 2 TL, B costs 6 TL; users split the cost equally (fair cost sharing). Thickness: number of players on the link.

stA: 2 TLB: 6 TL1234all on A: 0.5 TL eachmoving alone to B: 6 TLtotal 2 = optimum

All on A: each pays 0.5; one moving alone to B would pay 6 ⇒ nobody moves. Equilibrium, total 2 = optimum.

stA: 2 TLB: 6 TL1234all on B: 1.5 TL eachmoving alone to A: 2 TLtotal 6

All on B: each pays 1.5; one moving alone to A would pay 2 (2 > 1.5) ⇒ nobody moves. Also an equilibrium, total 6.

stA: 2 TLB: 6 TL1234the 1 on A: 2 TLthe 3 on B: 2 TLB to A: 1 TL ⇒ moves

Split profile (1 A, 3 B): a player on B moving to A pays 1 < 2 ⇒ not an equilibrium. The script checked all 16 profiles: only "all A" and "all B" are equilibria.

stA: 2 TLB: 6 TL1234PoA = 6/2 = 3PoS = 2/2 = 1

PoA = 6/2 = 3 (worst equilibrium), PoS = 1 (best equilibrium). A coordination problem: if players can be steered to the good equilibrium there is no loss; the PoS measures this optimistic scenario.

The Price of Stability

Step by Step: The Price of Stability (The Hn Example)

11/21/31/41.11234vtplayerpathpays1shared0.2752shared0.2753shared0.2754shared0.275total1.1

Each player reaches t by one of two paths: a private path of cost 1/i used only by them, or a shared path of cost 1 + ε = 1.1 (ε: a small extra cost) whose cost users split equally. Optimum: everyone shares, total 1.1, 0.275 each.

11/21/31/41.11234vtplayerpathpays1shared0.3672shared0.3673shared0.3674own0.25total1.35

Player 4: own path 1/4 = 0.25 < 0.275 → leaves. The remaining three now pay 1.1/3 ≈ 0.367.

11/21/31/41.11234vtplayerpathpays1shared0.552shared0.553own0.3334own0.25total1.683

Player 3: 1/3 ≈ 0.333 < 0.367 → leaves. The remaining two pay 0.55.

11/21/31/41.11234vtplayerpathpays1shared1.12own0.53own0.3334own0.25total2.183

Player 2: 1/2 = 0.5 < 0.55 → leaves. Player 1 pays for the shared path alone: 1.1.

11/21/31/41.11234vtplayerpathpays1own12own0.53own0.3334own0.25total2.083

Player 1: own path 1 < 1.1 → leaves. Everyone on their own path: 1 + 1/2 + 1/3 + 1/4 = 25/12 ≈ 2.083.

11/21/31/41.11234vtplayerpathpays1own12own0.53own0.3334own0.25total2.083

Script: this is the only equilibrium of the 16 profiles → PoS = PoA = (25/12) / 1.1 = 125/66 ≈ 1.894. As ε → 0 with n players, PoS → Hn = 1 + 1/2 + … + 1/n.

The Price of Stability

Why PoS ≤ Hn? Potential and Cost

Notation. ce: cost of link e; ne: number of players using e (each pays ce/ne). H(k) = 1 + 1/2 + … + 1/k; e.g. H(1) = 1, H(4) = 25/12 ≈ 2.08. Hn = H(n). Total cost C(s): the sum of the costs of the links in use.
Rosenthal's potential. Imagine players joining a link one at a time: the 1st writes ce, the 2nd ce/2, …, the kth ce/k. The link contributes ce·H(ne): Φ(s) = Σe ce·H(ne).
Example (the Hn example, four players). All on the shared path: C = 1.1 and Φ = 1.1·H(4) = 1.1·25/12 ≈ 2.29. All on their own paths: each path has 1 user, H(1) = 1, so Φ = C = 25/12 ≈ 2.08.
The potential is between 1 and Hn times the cost. On every link in use, 1 ≤ H(ne) ≤ Hn (there are at most n users). Multiply each term by ce and sum: for every s, C(s) ≤ Φ(s) ≤ Hn·C(s).

The Price of Stability

PoS ≤ Hn: Walking from the Optimum to an Equilibrium

1. Start at the optimum. Run best-response dynamics from sOPT: at each step one player switches to a path that is cheaper for them.
2. Φ drops at every move. In a potential game the deviator's change in cost equals the change in Φ; if the player gains, Φ falls. There are finitely many profiles, so the dynamics stop at an equilibrium s with Φ(s) ≤ Φ(OPT).
3. Chain it. C(s) ≤ Φ(s) (potential is at least the cost) ≤ Φ(OPT) (step 2) ≤ Hn·C(OPT) (potential is at most Hn times the cost). In the example: 2.08 ≤ 2.08 ≤ 2.29 ≤ 2.29.
Some equilibrium costs at most Hn times the optimum: PoS ≤ Hn (Anshelevich et al., 2004). Hn grows very slowly (H100 ≈ 5.19; roughly ln n). The Hn example (a directed network) shows the bound is tight. The PoA, by contrast, can be as large as n.

Reducing the PoA

Step by Step: A Congestion Toll Shifts the Equilibrium

1020304001,0002,0003,000x₁minℓ₁ + τℓ₂x₁ = 2,000 carstotal delay C (thousand car·min)τ = 075τ = 2.5τ = 5τ = 7.5τ = 10optimum 71.25(bars start at 70,000)

Two-route network: ℓ₁ = 5 + x₁/100, ℓ₂ = 20 + x₂/200, 3,000 cars in total. No toll (τ = 0): equilibrium ℓ₁ = ℓ₂ → x₁ = 2,000, total 75,000 car·min. Drivers count the toll as an extra cost in minutes.

1020304001,0002,0003,000x₁minℓ₁ + τℓ₂x₁ ≈ 1,833 carstotal delay C (thousand car·min)τ = 075τ = 2.572.92τ = 5τ = 7.5τ = 10optimum 71.25(bars start at 70,000)

τ = 2.5: the Route 1 curve shifts up and the crossing moves left: x₁ ≈ 1,833, total ≈ 72,917. A toll is a transfer; it is not added to total delay.

1020304001,0002,0003,000x₁minℓ₁ + τℓ₂x₁ ≈ 1,667 carstotal delay C (thousand car·min)τ = 075τ = 2.572.92τ = 571.67τ = 7.5τ = 10optimum 71.25(bars start at 70,000)

τ = 5: x₁ ≈ 1,667, total ≈ 71,667. Each increase pushes some drivers onto the wide Route 2.

1020304001,0002,0003,000x₁minℓ₁ + τℓ₂x₁ = 1,500 carstotal delay C (thousand car·min)τ = 075τ = 2.572.92τ = 571.67τ = 7.571.25τ = 10optimum 71.25(bars start at 70,000)

τ = 7.5: x₁ = 1,500 → the system optimum (71,250) is now an equilibrium. Marginal-cost tolls are x₁ℓ′₁ = 15 min on Route 1 and x₂ℓ′₂ = 7.5 on Route 2; what matters is the difference: 7.5.

1020304001,0002,0003,000x₁minℓ₁ + τℓ₂x₁ ≈ 1,333 carstotal delay C (thousand car·min)τ = 075τ = 2.572.92τ = 571.67τ = 7.571.25τ = 1071.67optimum 71.25(bars start at 70,000)

τ = 10: too high a toll overshoots: x₁ ≈ 1,333, total again ≈ 71,667. The right toll equals the externality; it is not “the more the better”.

Reducing the PoA

Step by Step: A Congestion Toll Fixes Pigou

stℓ(x) = xℓ(x) = 11 unit of traffic00.511.5200.250.50.751x (share on top road)cost seen by a driverlow: 1xold eq. x = 1

Start: no toll, everyone on top (x = 1), total delay 1. Goal: make the optimum (x = 1/2) an equilibrium.

stℓ(x) = x + toll xℓ(x) = 11 unit of traffic00.511.5200.250.50.751x (share on top road)cost seen by a driverlow: 1xx + toll

Marginal-cost toll τe = fe·ℓ′e(fe): on top ℓ′ = 1 → τ = x; on the bottom ℓ′ = 0 → no toll. A driver on top sees x + x = 2x.

stℓ(x) = x + toll xℓ(x) = 1top: 0.5bottom: 0.500.511.5200.250.50.751x (share on top road)cost seen by a driverlow: 1xx + tollnew eq. x = 1/2

New equilibrium: 2x = 1 → x = 1/2. Selfish drivers choose the optimal split on their own; delay 3/4.

stℓ(x) = x + toll xℓ(x) = 1top: 0.5bottom: 0.500.511.5200.250.50.751x (share on top road)cost seen by a driverlow: 1xx + tollnew eq. x = 1/2

Tolls collected: 1/2 · 1/2 = 1/4. Drivers pay 3/4 + 1/4 = 1 in total: unless the revenue is returned, drivers gain nothing and the gain goes to society.

stℓ(x) = x + toll xℓ(x) = 1top: 0.5bottom: 0.500.511.5200.250.50.751x (share on top road)cost seen by a driverlow: 1xx + tollnew eq. x = 1/2

A fixed toll gives the same result: τ = 1/2 on top → x + 1/2 = 1 → x = 1/2. General theorem: marginal-cost tolls make the optimum an equilibrium in every network.

Reducing the PoA

Other Ways to Reduce the PoA

Congestion pricing

Brings the external cost users impose on others into their decision (London, Stockholm, Singapore).

Central (Stackelberg) routing

A center routes part of the traffic, the rest choose selfishly; as the controlled share grows, the ratio approaches 1.

Capacity planning

Roughgarden–Tardos: the selfish equilibrium for r units of traffic costs no more than the optimal routing of 2r units. Adding capacity can substitute for tolls.

Route advice and information

Recommendations can steer the equilibrium to a lower total cost, but compliance is voluntary.

Mechanism design

Set the rules so that the good outcome is the unique equilibrium.

Counting the cost

Weigh an intervention's information, enforcement and fairness costs against the efficiency it gains.

Reducing the PoA

Real Life and Discussion

Toll lanes in traffic networks, inter-domain route selection on the Internet, time-of-use tariffs in power grids: the question is the same in all three. How bad is the selfish equilibrium, and which intervention brings it closer to the optimum?

Discussion questions

  • Why don't individual interests always reach the social optimum? What externality is at work in Pigou's example?
  • Which incentives could reduce the PoA (tolls, bonuses, central routing)? Which is fairer?
  • Where might Braess's paradox show up in daily life? When does closing a road make sense?

Review

Summary

ExampleEquilibriumOptimumRatio
Two routes, 3,000 cars75,00071,250PoA ≈ 1.053
Pigou (x and 1)13/4PoA = 4/3
Linear latency, any network——PoA ≤ 4/3 (tight)
Pigou, ℓ = xp1→ 0PoA → ∞
Atomic congestion, linear——PoA ≤ 5/2 (tight)
Braess, 4,000 cars320,000258,750PoA ≈ 1.237
Load balancing {2, 2, 1, 1}4 (worst)3PoA = 4/3, PoS = 1
Fair sharing, Hn example25/121.1PoS ≈ 1.894 (≤ Hn)
Pigou + toll τ = x3/43/41

Review

Sample Problems

  1. In Pigou's network let the top road's delay be x². What are the equilibrium cost, the optimum and the PoA?
  2. Jobs {3, 3, 2, 2, 2}, 2 identical machines. What is the optimal makespan? Find a Nash equilibrium with makespan 7.
  3. What is the PoA in the normalized Braess network (s→A: x, A→t: 1, s→B: 1, B→t: x, A→B: 0; 1 unit of traffic)?
  4. Build the Hn example for n = 3 (small ε). What does the PoS approach?
  5. In a Pigou network with top road 2x and bottom road 1, what is the PoA, and what fixed toll makes the optimum an equilibrium?
Show answers
  1. Equilibrium 1 (everyone on top). C(x) = x³ + 1 − x → xOPT = 1/√3 ≈ 0.577, C(OPT) ≈ 0.615 → PoA ≈ 1.626.
  2. Optimum 6 ({3, 3} | {2, 2, 2}). Equilibrium: {3, 2, 2} | {3, 2} → 7 and 5; a size-2 job moving gets 5 + 2 = 7, no gain; a size-3 job gets 8 → equilibrium. Ratio 7/6.
  3. Equilibrium: everyone zigzags, 1 + 0 + 1 = 2. Optimum: half and half, 1/2 + 1 = 3/2. PoA = 4/3.
  4. H3 = 1 + 1/2 + 1/3 = 11/6 ≈ 1.83; optimum 1 + ε → PoS → 11/6.
  5. Equilibrium x = 1/2 (2x = 1), C = 1. Optimum C = 2x² + 1 − x → x = 1/4, C = 7/8 → PoA = 8/7. Toll: 2·(1/4) + τ = 1 → τ = 1/2.

Review

Exercises: Traces and New Topics

  1. What does pigou_poa(3) return? (xOPT = 4−1/3)
  2. Jobs (4, 3, 3, 2) start on M1; each step, the first job that gains moves to the least loaded machine. Final makespan? Optimum?
  3. (1, ½)-smooth: what PoA bound? Compare with Pigou's (1, ¼).
  4. PoA ≠ PoS game, A costs 3 TL, B 8 TL, 4 players: equilibria, PoA, PoS?
  5. Verify y(z + 1) ≤ 5/3·y² + 1/3·z² for y = 2, z = 3.
  6. What does poa_pos([3, 2, 2, 1], 2) return?
Show answers
  1. xOPT ≈ 0.63, C(OPT) ≈ 0.528 ⇒ PoA ≈ 1.896 (between p = 2 and 4).
  2. The size-4 job moves to M2 (loads 8, 4), then a size-3 job (5, 7); now nobody gains: makespan 7. Optimum 6 ({4, 2} | {3, 3}) ⇒ ratio 7/6.
  3. PoA ≤ 1/(1 − ½) = 2; Pigou's (1, ¼) gives 4/3. The smaller μ, the better the bound.
  4. All on A (0.75 each; moving to B costs 8) and all on B (2 each; moving to A costs 3). PoA = 8/3, PoS = 1.
  5. Left 8, right 20/3 + 3 = 29/3 ≈ 9.67 ⇒ holds (gap 5/3).
  6. (1.25, 1.0): optimum 4 ({3, 1} | {2, 2}), but {3, 2} | {2, 1} (loads 5 and 3) is also an equilibrium ⇒ PoA = 5/4, PoS = 1.

Review

References

  • T. Roughgarden, Twenty Lectures on Algorithmic Game Theory, Cambridge University Press, 2016 (Lectures 11–15).
  • N. Nisan, T. Roughgarden, É. Tardos, V. Vazirani (eds.), Algorithmic Game Theory, Cambridge University Press, 2007 (Chapters 17–20).
  • E. Koutsoupias, C. Papadimitriou, “Worst-case Equilibria”, STACS, 1999.
  • T. Roughgarden, É. Tardos, “How Bad Is Selfish Routing?”, Journal of the ACM, 2002.
  • T. Roughgarden, “The Price of Anarchy Is Independent of the Network Topology”, J. Computer and System Sciences, 2003.
  • E. Anshelevich et al., “The Price of Stability for Network Design with Fair Cost Allocation”, FOCS, 2004.
  • T. Roughgarden, “Intrinsic Robustness of the Price of Anarchy”, STOC, 2009.
  • G. Christodoulou, E. Koutsoupias, “The Price of Anarchy of Finite Congestion Games”, STOC, 2005.
  • B. Awerbuch, Y. Azar, A. Epstein, “The Price of Routing Unsplittable Flow”, STOC, 2005.
  • A. C. Pigou, The Economics of Welfare, Macmillan, 1920.
  • A. R. Karlin, Y. Peres, Game Theory, Alive, AMS, 2017 (Chapter 8: Price of Anarchy); D. Easley, J. Kleinberg, Networks, Crowds, and Markets, Cambridge, 2010 (Chapter 8: traffic and Braess).
  • Course page: T. Roughgarden, CS364A (Stanford, 2013), notes for Lectures 11–15.

Review

References (2/2)

  • C. Papadimitriou, “Algorithms, Games, and the Internet”, STOC, 2001.
  • R. L. Graham, “Bounds for Certain Multiprocessing Anomalies”, Bell System Technical Journal 45(9), 1966.
  • G. Finn, E. Horowitz, “A Linear Time Approximation Algorithm for Multiprocessor Scheduling”, BIT 19, 1979.
  • E. Koutsoupias, M. Mavronicolas, P. Spirakis, “Approximate Equilibria and Ball Fusion”, Theory of Computing Systems 36(6), 2003.
  • A. Czumaj, B. Vöcking, “Tight Bounds for Worst-Case Equilibria”, SODA, 2002.
  • T. Roughgarden, “Stackelberg Scheduling Strategies”, SIAM Journal on Computing 33(2), 2004.

The End

Chapter 9: The Price of Anarchy
Previous: Chapter 8: Network Design and Routing · Next: Chapter 10: Multi-Agent Systems
1 / 1 Sercan KÜLCÜ. All rights reserved.