Chapter 8: Network Design and Routing

Algorithmic Game Theory
Routing basicsSelfish routingWardrop equilibriumBraess's paradoxCongestion gamesRosenthal's potentialNetwork formation games
Sercan KÜLCÜ · Giresun University · Computer Engineering

Networks and Routing

Warm-up: Braess's Paradox

The rules

Each driver picks their own fastest route from S to T; travel time on a narrow road grows with the number of drivers on it, while a wide road always takes 45 minutes.

Think about it
  • If a zero-time shortcut from A to B opens, do trips get shorter?
  • When everyone takes their own fastest route, is the average time minimized?
40·x/N454540·x/N0SABTN drivers S → T; x: drivers on that roaddashed: new shortcut A → B

Networks and Routing

Learning Outcomes

Find a Wardrop equilibrium

Compute the equilibrium of a two-route network by equalizing delays; find the system optimum via marginal costs and compare.

Explain Braess's paradox

Show with numbers why a new road can make everyone slower.

Reason with potentials

Compute Rosenthal's potential; prove that best-response dynamics must stop.

Model network formation

Check which networks are equilibria when players build the links.

Networks and Routing

Network Design and Routing

  • Network design decides the nodes, links, capacities and costs.
  • Routing decides which paths the flow takes from source to destination.
  • Algorithmic game theory studies traffic and communication networks whose users each try to reduce their own cost.
  • Key applications: city traffic, Internet backbones, power grids.
sabctDesign: which edges, what capacity?Routing: which path for the flow? (red)

Networks and Routing

Basic Network Design Concepts

StarRingFull meshTree
Nodes and links

Node: an intersection, router or substation. Link: a road, fiber line or transmission line.

Capacity and bandwidth

The flow a link can carry per unit time; beyond it, delay rises sharply.

Cost

Building cost (design) and usage cost: delay, energy, tolls (routing).

Example

Tree/layered topology in a data center; a near-mesh road network in a city.

Networks and Routing

Classic Routing Algorithms

MethodWhat it doesNote
DijkstraSingle-source shortest paths with nonnegative weightsO((V + E) log V); the core of OSPF (Open Shortest Path First, a link-state protocol)
Bellman–FordAllows negative weights; updates by learning from neighborsO(V·E); its distributed version is RIP (a distance-vector protocol)
Dynamic routingRecomputes paths when link states and costs changeAdapts to traffic and failures
Load balancingSpreads traffic over several paths or serversE.g. ECMP (equal-cost multipath)
Classic algorithms assume fixed edge weights. In reality delay depends on how many people choose a path; when strategic users can deviate from the suggested route, the network equilibrium and the system optimum diverge. This turns routing into a game.

Networks and Routing

Delay Depends on Load: ℓe(x)

  • Each edge's delay is a nondecreasing function of its flow x: ℓe(x).
  • Constant: a wide highway, ℓ = b.
  • Linear: ℓ = a·x + b (the two-route and Braess examples).
  • BPR (US Bureau of Public Roads): ℓ = t0(1 + 0.15 (x/c)4); t0: free-flow time, c: road capacity. Standard in traffic planning.
  • Queueing delay in networks: ℓ = 1/(c − x), only for 0 ≤ x < c (single-server queue, M/M/1; c: packets served per second, x: packets arriving per second). As x approaches c, delay grows without bound; for x ≥ c the queue grows forever.
012300.511.5x / capacityℓ / free-flow timeconstlinearBPR

Selfish Routing

Selfish Routing

  • Each user picks the route that minimizes their own delay.
  • With many small users, flow is treated as splittable (nonatomic): a single user's effect is negligible.
  • Individual route choices can drift away from the system optimum, which minimizes total delay.
  • In Braess's paradox, adding a link can increase the equilibrium delay.
Modelr units of traffic s → t
fP: flow on path P
fe: total flow on edge e
ℓP(f) = Σe∈P ℓe(fe)
Total delay
C(f) = Σe fe·ℓe(fe)

A path's delay is the sum of its edges' delays; total delay is flow × delay summed over edges.

Selfish Routing

Wardrop's Principles (1952)

1st principle: user equilibrium

  • All used paths have equal, minimal delay.
  • Unused paths are no faster.
  • No single user can cut their travel time by switching.

2nd principle: system optimum

  • Average (total) delay is minimal.
  • The split a central planner would choose.
  • Some users may be “sacrificed” to a slower road.
In short: no used path is slower than any other path. Equilibrium explains users' individual choices, but it need not match the system optimum. With linear delays an equilibrium exists, and every equilibrium gives each edge the same delay.

Selfish Routing

Step by Step: Wardrop Equilibrium on Two Routes

ABRoute 1: ℓ₁ = 5 + x₁/100Route 2: ℓ₂ = 20 + x₂/2003,000 cars A → B01020304001,0002,0003,000x₁ (cars on Route 1)delay (min)ℓ₁ℓ₂

Setup: 3,000 cars travel from A to B. Route 1 is short but narrow (delay rises fast); Route 2 is long but wide.

ABRoute 1: ℓ₁ = 5 + x₁/100Route 2: ℓ₂ = 20 + x₂/200x₁ = 3,000 cars → 35 minx₂ = 0 cars → 20 min01020304001,0002,0003,000x₁ (cars on Route 1)delay (min)ℓ₁ℓ₂

Try 1: everyone on Route 1 → ℓ₁ = 35 min, while the empty Route 2 takes 20. Route 1 drivers want to switch: not an equilibrium.

ABRoute 1: ℓ₁ = 5 + x₁/100Route 2: ℓ₂ = 20 + x₂/200x₁ = 1,500 cars → 20 minx₂ = 1,500 cars → 27.5 min01020304001,0002,0003,000x₁ (cars on Route 1)delay (min)ℓ₁ℓ₂

Try 2: 1,500 / 1,500 → ℓ₁ = 20, ℓ₂ = 27.5 min. Now Route 2 drivers move to Route 1: still not an equilibrium.

ABRoute 1: ℓ₁ = 5 + x₁/100Route 2: ℓ₂ = 20 + x₂/2003,000 cars A → B01020304001,0002,0003,000x₁ (cars on Route 1)delay (min)ℓ₁ℓ₂

Equilibrium condition: if both routes are used, their delays must be equal: 5 + x₁/100 = 20 + (3000 − x₁)/200.

ABRoute 1: ℓ₁ = 5 + x₁/100Route 2: ℓ₂ = 20 + x₂/200x₁ = 2,000 cars → 25 minx₂ = 1,000 cars → 25 min01020304001,0002,0003,000x₁ (cars on Route 1)delay (min)ℓ₁ℓ₂eq.: 25 min

×200: 1000 + 2x₁ = 4000 + 3000 − x₁ → 3x₁ = 6000 → x₁ = 2,000, x₂ = 1,000; both take 25 min.

ABRoute 1: ℓ₁ = 5 + x₁/100Route 2: ℓ₂ = 20 + x₂/200x₁ = 2,000 cars → 25 minx₂ = 1,000 cars → 25 min01020304001,0002,0003,000x₁ (cars on Route 1)delay (min)ℓ₁ℓ₂eq.: 25 min

Check: a driver who switches still faces ≈ 25 min → nobody gains. Total delay 3,000 × 25 = 75,000 car·min.

Selfish Routing

Equilibrium Is Not the System Optimum

Total delay
C = x₁·ℓ₁(x₁) + x₂·ℓ₂(x₂)
60b70b80b90b100b110b01,0002,0003,000x₁total delay (car·min)eq. 75k

At the equilibrium x₁ = 2,000 the total is 75,000 car·min. Is this the bottom of the curve?

Total delay
C = x₁·ℓ₁(x₁) + x₂·ℓ₂(x₂)
Marginal cost: the extra total delay from adding one more car
Route 1: x₁·ℓ₁ = 5x₁ + x₁²/100 → derivative 5 + x₁/50
Route 2: x₂·ℓ₂ = 20x₂ + x₂²/200 → derivative 20 + x₂/100
60b70b80b90b100b110b01,0002,0003,000x₁total delay (car·min)eq. 75k

One more car adds its own delay plus the x·ℓ′ delay it imposes on everyone else.

Total delay
C = x₁·ℓ₁(x₁) + x₂·ℓ₂(x₂)
Marginal cost: the extra total delay from adding one more car
Route 1: x₁·ℓ₁ = 5x₁ + x₁²/100 → derivative 5 + x₁/50
Route 2: x₂·ℓ₂ = 20x₂ + x₂²/200 → derivative 20 + x₂/100
Equate: 5 + x₁/50 = 20 + (3000 − x₁)/100
→ 3x₁ = 4500 → x₁ = x₂ = 1,500
60b70b80b90b100b110b01,0002,0003,000x₁total delay (car·min)eq. 75koptimum 71.25k

At the optimum, marginal costs are equal: traffic splits 1,500 / 1,500, total 71,250.

x₁ / x₂ℓ₁ / ℓ₂ (min)Total
Equilibrium2,000 / 1,00025 / 2575,000
Optimum1,500 / 1,50020 / 27.571,250
60b70b80b90b100b110b01,0002,0003,000x₁total delay (car·min)eq. 75koptimum 71.25k

At the optimum, Route 1 drivers take 20 min and Route 2 drivers 27.5; Route 2 drivers want to leave → the optimum is not an equilibrium.

Selfish Routing

Equilibrium and Optimum: Two Different Minimizations

50b70b90b110b01,0002,0003,000x₁ (cars on Route 1)thousand car·minC minimal: 1,500Φ minimal: 2,000C(x₁)Φ(x₁)
Wardrop equilibrium

Minimizes the Beckmann potential (Beckmann, McGuire, Winsten, 1956):
Φ(f) = Σe ∫0fe ℓe(t) dt
Derivative ℓe: the delay seen by the “last driver in”.

System optimum

Minimizes total delay:
C(f) = Σe fe·ℓe(fe)
Derivative ℓe + fe·ℓ′e: the marginal cost.

The two curves bottom out in different places: Φ at x₁ = 2,000 (equilibrium, Φ = 52,500), C at x₁ = 1,500 (optimum, C = 71,250). The gap comes from one term: a driver ignores the x·ℓ′ externality their presence imposes on others.

Selfish Routing

Code Trace: find_eq(l1, l2, r)

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 0hi = 3,000width = 3,000

Call: the two-route network (ℓ₁ = 5 + x/100, ℓ₂ = 20 + x/200), 3,000 cars. Goal: the x with ℓ₁(x) = ℓ₂(r − x).

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 0hi = 3,000width = 3,000

Search interval [0, 3,000]: the equilibrium must lie here (at x = 0 Route 1 is faster, at x = r slower).

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 1,500hi = 3,000width = 1,500x = 1,500ℓ₁ = 20 · ℓ₂ = 27.5

Round 1: x = 1,500 → ℓ₁ = 20 < ℓ₂ = 27.5. Route 1 is still faster, so more cars move to it: lo = 1,500.

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 1,500hi = 2,250width = 750x = 2,250ℓ₁ = 27.5 · ℓ₂ = 23.75

Round 2: x = 2,250 → ℓ₁ = 27.5 ≥ ℓ₂ = 23.75. Route 1 is too crowded: hi = 2,250.

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 1,875hi = 2,250width = 375x = 1,875ℓ₁ = 23.75 · ℓ₂ = 25.63

Round 3: x = 1,875 → ℓ₁ = 23.75 < ℓ₂ = 25.63. Route 1 is still faster, so more cars move to it: lo = 1,875.

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 1,875hi = 2,062.5width = 187.5x = 2,062.5ℓ₁ = 25.63 · ℓ₂ = 24.69

Round 4: x = 2,062.5 → ℓ₁ = 25.63 ≥ ℓ₂ = 24.69. Route 1 is too crowded: hi = 2,062.5.

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 1,968.75hi = 2,062.5width = 93.75x = 1,968.75ℓ₁ = 24.69 · ℓ₂ = 25.16

Round 5: x = 1,968.75 → ℓ₁ = 24.69 < ℓ₂ = 25.16. Route 1 is still faster, so more cars move to it: lo = 1,968.75.

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 1,968.75hi = 2,015.63width = 46.88x = 2,015.63ℓ₁ = 25.16 · ℓ₂ = 24.92

Round 6: x = 2,015.63 → ℓ₁ = 25.16 ≥ ℓ₂ = 24.92. Route 1 is too crowded: hi = 2,015.63.

1020304001,0002,0003,000x₁minℓ₁(x₁)ℓ₂(r − x₁)lo = 1,968.75hi = 2,015.63width = 46.88x = 1,992.19ℓ₁ = 24.92 · ℓ₂ = 25.04

Width 46.88 ≤ 50 → stop. Result ≈ 1,992.19 (true equilibrium 2,000). Each round halves the interval: after k rounds the error is ≤ r/2k.

def find_eq(l1, l2, r, eps=50):
  if l1(0) >= l2(r): return 0   # route 1 unused
  if l1(r) <= l2(0): return r   # all on route 1
  lo, hi = 0, r
  while hi - lo > eps:
    x = (lo + hi) / 2
    if l1(x) < l2(r - x):
      lo = x      # route 1 faster: raise
    else:
      hi = x      # route 1 slower: lower
  return (lo + hi) / 2

Since ℓ₁(x) − ℓ₂(r − x) is nondecreasing, its zero is the equilibrium; the first two lines catch the corner cases. In general networks the Beckmann potential is minimized by convex optimization (e.g. Frank–Wolfe).

Braess's Paradox

Step by Step: Braess's Paradox (4,000 Cars)

x/1000 cars450 cars450 carsx/1000 carssABtPathcarstimes→A→t0—s→B→t0—s→A→B→t——4,000 cars s → tx: cars on the edge

Setup: 4,000 cars from s to t. s→A and B→t are narrow roads (x/100 min); A→t and s→B are wide highways (always 45 min).

x/1002,000 cars452,000 cars452,000 carsx/1002,000 carssABtPathcarstimes→A→t2,00065 mins→B→t2,00065 mins→A→B→t——Average: 65 minTotal: 260,000 car·min

Symmetry: 2,000 / 2,000 → both paths take 2000/100 + 45 = 65 min. Switching doesn't help: a Wardrop equilibrium (and the optimum).

x/1002,000 cars452,000 cars452,000 carsx/1002,000 cars0sABtPathcarstimes→A→t2,00065 mins→B→t2,00065 mins→A→B→t0—Average: 65 minTotal: 260,000 car·min

The city opens a very short, 0-minute link road from A to B. At first nobody uses it.

x/1002,000 cars452,000 cars452,000 carsx/1002,000 cars0sABtPathcarstimes→A→t2,00065 mins→B→t2,00065 mins→A→B→t040 minZigzag: 20 + 0 + 20= 40 < 65 min

One driver tries s→A→B→t: 20 + 0 + 20 = 40 min < 65. Very profitable individually → they switch.

x/1002,500 cars451,500 cars451,500 carsx/1002,500 cars01,000 carssABtPathcarstimes→A→t1,50070 mins→B→t1,50070 mins→A→B→t1,00050 minZigzag still fastest50 < 70 min

1,000 drivers have switched: 2,500 cars on the narrow roads → 25 min. Old paths 70, zigzag 50 min: switching continues.

x/1004,000 cars450 cars450 carsx/1004,000 cars04,000 carssABtPathcarstimes→A→t085 mins→B→t085 mins→A→B→t4,00080 minAverage: 80 minTotal: 320,000 car·min

Everyone zigzags: 40 + 0 + 40 = 80 min. Going back to an old path costs 40 + 45 = 85 min → nobody goes back: a new equilibrium.

x/1004,000 cars450 cars450 carsx/1004,000 cars04,000 carssABtPathcarstimes→A→t085 mins→B→t085 mins→A→B→t4,00080 min65 → 80 min260k → 320k car·min

Result: a road was added and everyone lost 15 min. More capacity + selfish choice = a worse equilibrium. The system optimum with the shortcut is ≈ 64.7 min (only 500 cars on the zigzag).

Braess's Paradox

Why Does Everyone Zigzag?

3040506070809001,0002,0003,0004,000z: cars on the zigzagtravel time (min)old equilibrium 65s→A→t, s→B→tzigzag
If the rest of the traffic splits evenly over the two old paths, they take 65 + z/200 and the zigzag 40 + z/100 min. The zigzag is faster for every z (B→t is at most 40 min < 45), so it is a dominant choice for everyone. But each switching driver loads the narrow roads: at z = 4,000 everyone takes 80 min.

Braess's Paradox

Braess's Paradox: Lessons and Examples

  • Adding a link to a network can increase the delay of all users because of selfish route choice (Braess, 1968).
  • The new road looks attractive individually, but the traffic pattern users create together moves to a worse equilibrium.
  • Conversely, closing a road can sometimes speed everyone up.
Stuttgart (1969)

Traffic that worsened after a new road investment is often said to have improved once a street was closed (Knödel, 1969).

New York (1990)

When 42nd Street was closed on Earth Day, the expected gridlock reportedly never came (Kolata, 1990).

Mechanical / electrical

A similar paradox appears in networks of springs and resistors (Cohen and Horowitz, 1991).

Tolls, access restrictions or central routing can reduce the effect.

Braess's Paradox

Back to the Warm-up: Braess's Paradox

The theory's answer: no
Average timeSelfish equilibriumOptimum
No shortcut20 + 45 = 6565
With shortcut40 + 0 + 40 = 80≈ 64.7

With the shortcut everyone picks S→A→B→T. Going back to an old road costs 40 + 45 = 85, so nobody does: equilibrium 80. The optimum with the shortcut gives the zigzag a small share (1,750 / 1,750 / 500 cars → 258,750 / 4,000 ≈ 64.7).

In practice

In laboratory experiments, too, subjects drifted over time toward the slower equilibrium the paradox predicts once the shortcut was added (Rapoport et al., 2009).

In cities: besides Stuttgart and New York, removing the elevated highway over Seoul's Cheonggyecheon in the early 2000s is often said not to have worsened traffic as feared. Real networks are far more complex; these cases are not proof, only signs that the paradox is possible.

Building or closing a road should be judged by the equilibrium drivers will reach, not by capacity alone.

Braess's Paradox

Code Trace: best_subnet(net, r)

x/1004545x/1000sABtremovedeq. time—s→AA→ts→BB→tA→B

Call: the Braess network, 5 edges, 4,000 cars. Question: closing which edge cuts the equilibrium time the most?

x/1004545x/1000sABtremovedeq. time—80 mins→AA→ts→BB→tA→Bbest80 mincut = None

Equilibrium of the full network: everyone zigzags, 80 min. Removing nothing is the initial candidate.

✕ x/1004545x/1000sABtremovedeq. time—80 mins→A85 minA→ts→BB→tA→Bbest80 mincut = None

Close s→A: only s→B→t remains: 40 + 45 = 85 min.

✕ x/1004545x/1000sABtremovedeq. time—80 mins→A85 minA→ts→BB→tA→Bbest80 mincut = None

85 < 80 is false → candidate unchanged.

x/100✕ 4545x/1000sABtremovedeq. time—80 mins→A85 minA→t80 mins→BB→tA→Bbest80 mincut = None

Close A→t: the zigzag and the lower path remain; the zigzag dominates: 80 min.

x/100✕ 4545x/1000sABtremovedeq. time—80 mins→A85 minA→t80 mins→BB→tA→Bbest80 mincut = None

80 < 80 is false → candidate unchanged.

x/10045✕ 45x/1000sABtremovedeq. time—80 mins→A85 minA→t80 mins→B80 minB→tA→Bbest80 mincut = None

Close s→B: the upper path and the zigzag remain; the zigzag dominates: 80 min.

x/10045✕ 45x/1000sABtremovedeq. time—80 mins→A85 minA→t80 mins→B80 minB→tA→Bbest80 mincut = None

80 < 80 is false → candidate unchanged.

x/1004545✕ x/1000sABtremovedeq. time—80 mins→A85 minA→t80 mins→B80 minB→t85 minA→Bbest80 mincut = None

Close B→t: only s→A→t remains: 40 + 45 = 85 min.

x/1004545✕ x/1000sABtremovedeq. time—80 mins→A85 minA→t80 mins→B80 minB→t85 minA→Bbest80 mincut = None

85 < 80 is false → candidate unchanged.

x/1004545x/100✕ 0sABtremovedeq. time—80 mins→A85 minA→t80 mins→B80 minB→t85 minA→B65 minbest80 mincut = None

Close A→B: the pre-Braess network: 2,000 / 2,000, 20 + 45 = 65 min.

x/1004545x/100✕ 0sABtremovedeq. time—80 mins→A85 minA→t80 mins→B80 minB→t85 minA→B65 minbest65 mincut = A→B

65 < previous best → update: new best subnetwork “A→B closed”.

x/1004545x/100✕ 0sABtremovedeq. time—80 mins→A85 minA→t80 mins→B80 minB→t85 minA→B65 minbest65 mincut = A→B

Result: close A→B; equilibrium 80 → 65 min. The designer asks “which edges should stay open?”; in general networks this choice is hard.

def best_subnet(net, r):
  best, cut = eq_time(net, r), None
  for e in net:
    sub = [k for k in net if k != e]
    d = eq_time(sub, r)   # inf if no path
    if d < best:
      best, cut = d, e
  return cut, best

eq_time(net, r): the common time of the used paths in the subnetwork's Wardrop equilibrium (computed in the script by minimizing the Beckmann potential). The code only tries removing one edge; a full search needs 2m subnetworks.

Braess's Paradox

Network Design for Selfish Users

Series-parallel: no paradoxsutbuilt from series/parallel joinssABtWheatstone (Braess) substructure
Problem

Which subnetwork of a given network, left open, minimizes the Wardrop equilibrium time? Closing a road, not only building one, is a design decision.

Further reading: hardness

Choosing the best edges to close is hard in general: with linear delays, closing nothing is at most 4/3 times worse than the best choice, and guaranteeing anything better efficiently is NP-hard (Roughgarden, 2006).

Further reading: topology

In networks built from parallel paths and their series connections (series-parallel), Braess's paradox never occurs; other networks can show it with suitable delays (Milchtaich, 2006).

Congestion Games

Congestion Games

  • n players, a set of resources E (edges). Each player's strategy is a subset of resources (e.g. an s–t path).
  • A resource's cost ce(ne) depends only on the number of players using it and grows with that load.
  • A player's cost: the sum of the costs of the chosen resources.
  • As players try to cut their own delay, the total network cost can rise.
AtomicNonatomic
Playersfinite, n playerscontinuous flow
Effectone player adds 1 to the loadone player is negligible
Equilibriumpure NashWardrop
PotentialRosenthal (sum)Beckmann (integral)
Rosenthal (1973): every finite congestion game has a potential function, so a pure Nash equilibrium always exists. No mixed strategies needed.

Congestion Games

Rosenthal's Potential Function

Φ(s) = Σe [ ce(1) + ce(2) + … + ce(ne) ]
On each edge, add the costs from 1 up to the number of users ne (this is not the total cost).
Example: all 4 drivers on s→A→t. s→A: ce(k) = k → 1 + 2 + 3 + 4 = 10; A→t: ce(k) = 4 → 16; other edges empty. Φ = 26 (total cost 4 × 8 = 32).
1k=12k=23k=34k=4s→A: ce(k) = kΣ = 104k=14k=24k=34k=4A→t: ce(k) = 4Σ = 16
Key identity. When driver i leaves an edge, the last term ce(ne) of that edge's sum disappears (that is what i paid there); entering a new edge adds ce(ne + 1) (i's new cost). So the change in Φ is exactly i's change in cost: ΔΦ = ci(new) − ci(old).
Consequence. Every improving move strictly decreases Φ; Φ takes finitely many values → the dynamics stop. Where they stop, nobody can improve → a pure Nash equilibrium exists (Rosenthal, 1973).

Congestion Games

Step by Step: Best-Response Dynamics and Φ

k [4]4 [4]4 [0]k [0]0 [0]sABtdriverpathcost1top82top83top84top8Φ (potential)26

Edge costs: k (number of users) on s→A and B→t, 4 on s→B and A→t, 0 on A→B. Everyone on s→A→t: cost 4 + 4 = 8, Φ = 26.

k [3]4 [3]4 [1]k [1]0 [0]sABtdriverpathcost1bottom52top73top74top7Φ (potential)2623

Driver 1's options: top 8, bottom 4 + 1 = 5, zig 4 + 0 + 1 = 5. Ties go to the first → bottom (8 → 5). Φ: 26 → 23 (Δ = −3).

k [3]4 [2]4 [1]k [2]0 [1]sABtdriverpathcost1bottom62zig53top74top7Φ (potential)262321

Driver 2: top 7, bottom 4 + 2 = 6, zig 3 + 0 + 2 = 5 → zig (7 → 5). Φ: 23 → 21 (Δ = −2).

k [3]4 [1]4 [1]k [3]0 [2]sABtdriverpathcost1bottom72zig63zig64top7Φ (potential)26232120

Driver 3: top 7, bottom 4 + 3 = 7, zig 3 + 0 + 3 = 6 → zig (7 → 6). Φ: 21 → 20 (Δ = −1).

k [3]4 [1]4 [1]k [3]0 [2]sABtdriverpathcost1bottom72zig63zig64top7Φ (potential)26232120

Driver 4 (top, 7): bottom 4 + 4 = 8, zig 3 + 0 + 4 = 7 → nothing better, stays. Each drop in Φ equaled the mover's gain.

k [3]4 [1]4 [1]k [3]0 [2]sABtdriverpathcost1bottom72zig63zig64top7Φ (potential)26232120

In the second round nobody can improve → pure Nash equilibrium (Φ = 20). Total 7 + 6 + 6 + 7 = 26; without A→B a 2+2 split would give 24: atomic Braess!

Congestion Games

Potential Games: What Do We Gain?

Existence

A state minimizing Φ is a pure Nash equilibrium; an equilibrium always exists.

Convergence

No sequence of improving moves is infinite (the finite improvement property): best-response dynamics with strict improvements always stop (a player already best-responding stays put).

Speed (further reading)

Stopping is guaranteed, speed is not: for some congestion games no fast method for finding an equilibrium is known (Fabrikant, Papadimitriou, Talwar, 2004).

Generality

Every (exact) potential game is equivalent to a congestion game (Monderer and Shapley, 1996).

Learning

If players best-respond in turn, they reach an equilibrium.

Efficiency

The equilibrium minimizing Φ is one way to select a good equilibrium among many, but it is no guarantee.

Network Formation

Network Externalities: Negative and Positive

  • Externality: a user's decision changes others' costs or benefits, but the user ignores this in their own calculation.

Negative: congestion

  • Each extra driver imposes a total delay of x·ℓ′(x) on the x others on the same road.
  • In the two-route example, with 2,000 cars on Route 1: 2,000 × 1/100 = 20 car·min.
  • Result: the busy road is overused; the remedy is a congestion toll.

Positive: network effects

  • The value of joining grows with the number of members: phones, messaging, social networks.
  • n users have n(n − 1)/2 possible connections; the claim that value grows like n² (Metcalfe) is a heuristic.
  • A new user ignores the benefit they bring others: adoption stays slow until a critical mass is reached.
Network formation games have both: a link shortens distances for others as well as its builder (a positive externality), but only the buyer pays for it.

Network Formation

Network Formation Games

  • So far the network was given; now the players build it themselves: Internet peering, social networks, P2P networks.
  • Every link has a building cost; being well connected (short distances) is valuable. The tension: link cost ↔ reach.
Fabrikant et al. (2003): unilateral

Player i buys links to any nodes (α each). Cost: α·(links bought) + Σj dist(i, j). Equilibrium: Nash.

Jackson–Wolinsky (1996): bilateral

A link needs both sides' consent to form; either side can cut it. Solution concept: pairwise stability.

Questions: Which networks are equilibria? Which is socially best? How far apart can they be?

Network Formation

Bilateral Links: Pairwise Stability (n = 4)

  • Benefit: 4 from each player at distance 1, 2.2 at distance 2, 1 at distance 3.
  • Each link costs 2 to each of its two endpoints.
  • Pairwise stable: (i) no player gains by cutting one of their links; (ii) no missing link would benefit both endpoints together.
ABCDpathABCDcycleABCDstar (center A)ABCDcomplete graph
NetworkPayoffsTotalPairwise stable?
path5.2/6.2/6.2/5.222.8No: A–D is added (+3 benefit, −2 cost each)
cycle6.2 (all)24.8Yes
star6/6.4/6.4/6.425.2Yes; highest total
complete graph6 (all)24No: cutting a link gains 2, loses 1.8

In the star, two leaves linking would each gain 1.8 but pay 2, so they don't.

There can be several pairwise stable networks, and they need not match the one with the highest total: the cycle is stable but trails the star by 0.4.

Network Formation

Step by Step: A Network Formation Game (n = 4, α = 1.5)

12344 players, link price α = 1.5cost_i = α · (links bought) + Σ_j dist(i, j)Social cost = Σ cost_iDisconnected → cost ∞

Each player chooses which links to buy themselves; everyone may use a link, whoever bought it.

17.525.535.546Path 1–2–3–4 (i buys i+1)costs 7.5 · 5.5 · 5.5 · 6social cost 24.5

Path network: 1 buys one link (1.5) and its distances are 1 + 2 + 3 = 6 → 7.5. The end players are far from everyone.

1725.535.544If 1 buys a link to 4:2 · 1.5 + (1 + 2 + 1) = 77 < 7.5 → gainThe path is not an equilibrium

If player 1 buys a second link to 4, its distances drop from 6 to 4: 3 + 4 = 7 < 7.5 → a profitable deviation.

1234Complete graph: 6 linkssocial: 6 · 1.5 + 12 = 21Dropping a link: −1.5 + 1→ gains, not an equilibrium

The complete graph is socially cheapest (21). But a player who bought a link saves 1.5 by dropping it and loses only 1 in distance → not an equilibrium.

17.5253545Center 1 buys three linkscenter 4.5 + 3 = 7.5; leaf 5A leaf buying a link: 1.5 + 4 = 5.5Center drops a link → split→ Nash equilibrium (22.5)

Star: a new link saves a leaf only 1 in distance (1 < 1.5); if the center drops a link the network splits → nobody deviates.

15.525.535.545.54-cycle (one link each): 5.5social cost 22Script: all 4,096 profiles checkedequilibria: star 22.5 · cycle 22optimum: complete graph 21

Brute-force search: at α = 1.5 the equilibria are the star (22.5) and the 4-cycle (22); the optimum (21) is not an equilibrium.

Network Formation

Code Trace: cost(i, links, alpha)

1d=0234player i = 1queue: [1]dist: 1:0Σ dist = 0arrow: buyer → other end · red: new link

Path network 1–2–3–4; 1 has bought only link (1, 2). BFS starts at i = 1: dist = {1: 0}.

1d=02d=134player i = 1queue: [2]dist: 1:0 2:1Σ dist = 1arrow: buyer → other end · red: new link

u = 1 leaves the queue; new neighbor 2 → distance 1.

1d=02d=13d=24player i = 1queue: [3]dist: 1:0 2:1 3:2Σ dist = 3arrow: buyer → other end · red: new link

u = 2 leaves the queue; new neighbor 3 → distance 2.

1d=02d=13d=24d=3player i = 1queue: [4]dist: 1:0 2:1 3:2 4:3Σ dist = 6arrow: buyer → other end · red: new link

u = 3 leaves the queue; new neighbor 4 → distance 3.

1d=02d=13d=24d=3player i = 1queue: []dist: 1:0 2:1 3:2 4:3Σ dist = 6arrow: buyer → other end · red: new link

u = 4 leaves the queue; all its neighbors are already visited.

1d=02d=13d=24d=3player i = 1queue: []dist: 1:0 2:1 3:2 4:3Σ dist = 6arrow: buyer → other end · red: new link

Everyone is reachable. bought = 1 → cost 1.5·1 + 6 = 7.5. Now let's try buying one more link.

1d=0234player i = 1queue: [1]dist: 1:0Σ dist = 0arrow: buyer → other end · red: new link

Deviation: 1 also buys link (1, 4) (orange). Same computation again. BFS starts at i = 1: dist = {1: 0}.

1d=02d=134d=1player i = 1queue: [2, 4]dist: 1:0 2:1 4:1Σ dist = 2arrow: buyer → other end · red: new link

u = 1 leaves the queue; new neighbors 2, 4 → distance 1.

1d=02d=13d=24d=1player i = 1queue: [4, 3]dist: 1:0 2:1 3:2 4:1Σ dist = 4arrow: buyer → other end · red: new link

u = 2 leaves the queue; new neighbor 3 → distance 2.

1d=02d=13d=24d=1player i = 1queue: [3]dist: 1:0 2:1 3:2 4:1Σ dist = 4arrow: buyer → other end · red: new link

u = 4 leaves the queue; all its neighbors are already visited.

1d=02d=13d=24d=1player i = 1queue: []dist: 1:0 2:1 3:2 4:1Σ dist = 4arrow: buyer → other end · red: new link

u = 3 leaves the queue; all its neighbors are already visited.

1d=02d=13d=24d=1player i = 1queue: []dist: 1:0 2:1 3:2 4:1Σ dist = 4arrow: buyer → other end · red: new link

Everyone is reachable. bought = 2 → cost 1.5·2 + 4 = 7. 7 < 7.5 → a profitable deviation: the path is not a Nash equilibrium.

def cost(i, links, alpha):
  dist = {i: 0}; queue = [i]
  while queue:
    u = queue.pop(0)
    for v in nbrs(links, u):
      if v not in dist:
        dist[v] = dist[u] + 1
        queue.append(v)
  bought = sum(1 for (a, b) in links if a == i)
  return alpha * bought + sum(dist.values())

nbrs(links, u): neighbors regardless of who bought the link (everyone uses links); the network is assumed connected. A player pays only for the links they bought.

Network Formation

Equilibrium and Optimum as the Link Price Changes

10152025300123α (link price)social coststarcompleteα = 2
αOptimumEquilibria (script)
0.5compl. (15)compl. (15)
1.5compl. (21)cycle 22, star 22.5
2.5star (25.5)star 25.5 · path 27.5

For n = 4 the star costs 3α + 18 and the complete graph 6α + 12: for α < 2 the complete graph is cheaper, for α > 2 the star. For general n the optimum is likewise the complete graph for α ≤ 2 and the star for α ≥ 2 (Fabrikant et al., 2003).

α < 1: links are cheap, everyone connects → the complete graph is both equilibrium and optimum. α ≥ 1: the star is always an equilibrium, but costlier equilibria can appear too (at α = 2.5, the path, with the middle players buying the links).

Control and Incentives

Routing in Networks: Centralized or Distributed?

Centralized routingDistributed (selfish) routing
Who decides?A traffic control center, an SDN (software-defined networking) controller, ISP traffic engineeringEach driver / each router (navigation apps, BGP: the inter-domain routing protocol)
GoalMinimum total delay C(f) (Wardrop's 2nd principle)Minimum own delay (Wardrop's 1st principle)
Two-route example1,500 / 1,500 → 71,250 car·min2,000 / 1,000 → 75,000 car·min
RequiresFull demand and delay information; users must complyOnly local information; no compliance problem
WeaknessAt the optimum some are on the slower road → urge to deviate; a single point of failureInefficient equilibria such as Braess's
Hybrids are common in networks: the center controls only prices (congestion tolls), information (route suggestions) or part of the traffic (Stackelberg routing: the center routes that part itself); users choose the rest.

Control and Incentives

Incentives, Load Balancing and Real Life

Right incentives

  • Steer users toward balanced use of network resources → efficiency improves.
  • Traffic: tolls or warnings reduce congestion.
  • Smart grid: time-of-use pricing balances load at peak hours.

Wrong incentives

  • Strategic, selfish behavior grows → congestion and low efficiency.
  • An app sending everyone the same “fastest” route can flood side streets.
  • A badly chosen toll shifts traffic to another bottleneck.
Internet

ISPs route data packets; BGP picks routes between networks.

Traffic management

Route optimization, smart signals, navigation apps.

Energy

Energy distribution and load shifting in smart grids.

Control and Incentives

Efficiency vs. Fairness: Sharing Bandwidth

Two links have capacity 1. Flow A uses both, B only link 1, C only link 2: xA + xB ≤ 1, xA + xC ≤ 1. Which allocation should we pick?

CriterionWhat is optimized?(xA, xB, xC)Total
Maximum total flowΣ xi(0, 1, 1)2 (A starves)
Max-min fairness (egalitarian)Raise the smallest share as far as possible, then the next(1/2, 1/2, 1/2)1.5
Proportional fairness (Nash social welfare)Maximize the product of shares (Kelly, 1997)(1/3, 2/3, 2/3)≈ 1.67
Under proportional fairness A, which uses both links, gets less, but nobody drops to zero: a zero share zeroes the product. Maximizing xA(1 − xA)² is the same as maximizing log xA + 2 log(1 − xA); since the derivative of log x is 1/x, 1/xA = 2/(1 − xA) ⇒ xA = 1/3. Kelly et al. (1998) interpret TCP-like congestion control as a distributed algorithm that converges to this allocation via per-link congestion prices.

Control and Incentives

Discussion Questions

  • When navigation apps suggest the “fastest” route to everyone, do we approach the Wardrop equilibrium or the system optimum? Why?
  • A city plans a new bridge. To foresee a Braess effect, how do we compute the equilibrium drivers will reach once it opens?
  • Are Internet routers (BGP) selfish agents? Whose cost do they minimize?

Review

Summary

ConceptCore ideaExample
Selfish routingEveryone minimizes their own delay3,000 cars, two routes
Wardrop equilibriumUsed routes are equal and fastest2,000 / 1,000, 25 min
System optimumMarginal costs are equal1,500 / 1,500, 71,250 car·min
Braess's paradoxNew road → everyone slower65 → 80 min
Rosenthal's potentialΔΦ = the mover's cost changeΦ: 26 → 23 → 21 → 20
Network formationLink cost ↔ reachα = 1.5: star / cycle equilibria, complete graph optimal
Equilibrium and optimum often diverge: a driver looks at their own delay, a planner at marginal cost.

Review

Sample Problems

  1. 6,000 cars; Route 1: ℓ₁ = 10 + x₁/200, Route 2: ℓ₂ = 25 + x₂/100. Wardrop equilibrium?
  2. System optimum of the same network?
  3. Braess network (x/100 and 45 min), 6,000 cars: equilibrium time before and after the new road?
  4. Φ at the atomic example's equilibrium (bottom, zig, zig, top)? Change if driver 4 switches to zig?
  5. n = 4, α = 0.5: is the complete graph an equilibrium? The star?
  6. Problem 1: which x does find_eq try in two rounds, and the interval?
Show answers
  1. 10 + x₁/200 = 25 + (6000 − x₁)/100 → 3x₁ = 15,000 → x₁ = 5,000, x₂ = 1,000; both 35 min.
  2. Marginal: 10 + x₁/100 = 25 + x₂/50 → x₁ = 4,500, x₂ = 1,500 (ℓ₁ = 32.5, ℓ₂ = 40).
  3. Before: 3,000 / 3,000: 30 + 45 = 75 min. After: 3,000 zigzag, 1,500 per old path; s→A load 4,500 → every path 90 min.
  4. (1 + 2 + 3) + (1 + 2 + 3) + 4 + 4 + 0 = 20. Driver 4's cost goes 7 → 7, so ΔΦ = 0.
  5. Complete: dropping a link saves 0.5, costs 1 → equilibrium. Star: a leaf buying a link pays 0.5, gains 1 → not.
  6. x = 3,000: ℓ₁ = 25 < ℓ₂ = 55 → lo = 3,000; x = 4,500: 32.5 < 40 → lo = 4,500. Interval [4,500, 6,000] (equilibrium 5,000).

Review

References

  • T. Roughgarden, Selfish Routing and the Price of Anarchy, MIT Press, 2005.
  • N. Nisan, T. Roughgarden, É. Tardos, V. Vazirani (eds.), Algorithmic Game Theory, Cambridge University Press, 2007 (Chapters 18–19).
  • J. G. Wardrop, “Some Theoretical Aspects of Road Traffic Research”, Proc. Institution of Civil Engineers, 1952.
  • D. Braess, “Über ein Paradoxon aus der Verkehrsplanung”, Unternehmensforschung, 1968.
  • R. W. Rosenthal, “A Class of Games Possessing Pure-Strategy Nash Equilibria”, Int. J. Game Theory, 1973.
  • D. Monderer, L. Shapley, “Potential Games”, Games and Economic Behavior, 1996.
  • A. Fabrikant et al., “On a Network Creation Game”, PODC, 2003. · M. Jackson, A. Wolinsky, “A Strategic Model of Social and Economic Networks”, J. Economic Theory, 1996.

Review

References: Papers, Books and Courses

  • A. Fabrikant, C. Papadimitriou, K. Talwar, “The Complexity of Pure Nash Equilibria”, STOC, 2004. · M. Beckmann, C. B. McGuire, C. B. Winsten, Studies in the Economics of Transportation, Yale University Press, 1956.
  • A. Rapoport, T. Kugler, S. Dugar, E. J. Gisches, “Choice of Routes in Congested Traffic Networks: Experimental Tests of the Braess Paradox”, Games and Economic Behavior 65(2), 2009. · W. Knödel, Graphentheoretische Methoden und ihre Anwendungen, Springer, 1969. · G. Kolata, “What if They Closed 42d Street and Nobody Noticed?”, The New York Times, December 25, 1990.
  • I. Milchtaich, “Network Topology and the Efficiency of Equilibrium”, Games and Economic Behavior 57, 2006. · T. Roughgarden, “On the Severity of Braess's Paradox”, J. Computer and System Sciences 72, 2006.
  • F. Kelly, “Charging and Rate Control for Elastic Traffic”, European Trans. Telecommunications 8, 1997. · F. Kelly, A. Maulloo, D. Tan, J. Operational Research Society 49, 1998.
  • D. Easley, J. Kleinberg, Networks, Crowds, and Markets, Cambridge University Press, 2010 (Chapter 8: traffic networks and Braess; Chapter 17: network effects).
  • A. R. Karlin, Y. Peres, Game Theory, Alive, American Mathematical Society, 2017 (Price of Anarchy and traffic routing).
  • Course pages: T. Roughgarden, CS364A (Stanford) · M. Hoefer, AGT (Frankfurt) · M. Dinitz, AGT (Johns Hopkins)

The End

Chapter 8: Network Design and Routing
Previous: Chapter 7: Social Choice and Voting · Next: Chapter 9: The Price of Anarchy
1 / 1 Sercan KÜLCÜ. All rights reserved.