Routing games

Routing Games

By Javad Ghareh Chamani

1

2

Outlines

•Introduction

•Some Definitions

•Non-atomic Selfish Routing

•Atomic Selfish Routing

3

Introduction2 Routing models

◦Nonatomic selfish routing

Reasons of explain◦Simple but general◦Understandability of POA◦Different techniques needed for

analyze

◦Atomic selfish routing

4

Some DefinitionsPigou’s example

◦ Selfish players◦ Each player wants to minimize cost

5

Multicommodity Flow Network

Directed graph G = (V, E), may have parallel edgesCommodities:

set of source/sink pairs: (s1, t1),…,(sk , tk )Assign each player to a\some commoditynonnegative, continuous, non-decreasing cost

function ce

6

Multicommodity Flow Network (cnt)

Paths

Routes◦ Traffic described with flow f

◦ fP: amount of traffic of commodity i that chooses path to travel from si to ti

◦ Prescribed amount of traffic for commodity i : ri

7

Feasibility of FlowA flow f is feasible for a vector r if it

routes all of the traffic

8

Cost DefinitionCost of a path P with respect to a flow

f

fe amount of traffic using paths that contain the edge e.

9

Nonatomic EquilibriumA flow (f) is equilibrium if:

no commodity can increase its gain by changing its traffic

10

Braess’s Paradox

One commoditywants to route 1

unit of trafficfrom s to t

11

Braess’s Paradox(cont)

New equilibrium

Old cost: 1.5New cost: 2

12

Existence & Uniqueness of Equilibrium

There exists at least one equilibrium

All equilibriums are equivalent (of equal cost)

Using Potential Function

13

Potential FunctionsLet ce(x) be the cost of transporting x

over some edge eThe total amount spent on that edge

is x.ce(x)

14

Marginal Cost and Optimality

Follows from convex optimization problem nonnegativity constraints

15

Marginal Cost and Optimality(cnt)

How to minimize the total marginal cost?

Create a new network with the marginal cost as cost function, and find an equilibrium.

16

Back to Finding EquilibriumRemaining questions?

◦ Can we find equilibrium by finding optimal flows?

◦ What function should we minimize?

17

Potential FunctionWe want to minimize the functions

he(x) of all edges. The sum of those is called the potential function and should be minimized.

18

ConsequencesThe potential function is convex:

◦ there exists a minimum, and therefore an equilibrium

◦ and all equilibriums form a set◦ with the same cost

19

Atomic Selfish RoutingAlmost the same as the nonatomic

instance.Differences:

◦ Each commodity represents ONE player instead of a large number of players.

◦ Player i must route traffic amount ri on a

SINGLE path.Result: Each player must route a

significant amount of traffic instead of a negligible amount.

20

Feasibility of a flowA flow f is feasible if:

◦it routes all traffic◦it uses one path per player i to route

ri units

21

Atomic equilibrium flow

Also note:Different equilibrium flows can have different

costs

no uniqueness

22

Nonexistence of equilibria

Traffic amount: r1 = 1; r2 = 2 P1 s -> t P2 s -> v -> t P3 s -> w -> t P4 s -> v -> w -> t

1- If player 2 takes P1 or P2, player 1 takes P4.

2- If player 1 takes P4, player 2 takes P3. 3- If player 2 takes P3 or P4, player 1

takes P1. 4- If player 1 takes P1, player 2 takes P2. This does never end. . .

23

Difference of Atomic and NonDifferent EQ of an atomic

instance can have different costsNonatomic EQ have same costPOA in atomic instance can be

larger than nonatomic

24

Overcome nonexistence of equilibrium

To guarantee an equilibrium, additional restrictions are placed on atomic instances. For instance:

Give all players an equal amount of traffic which has to be routed.

25

Overcome nonexistence of equilibrium (cnt)

26

ProofIn atomic instances we can

discretize the potential function that was used for nonatomic instances:

#Possible flows is finite. One flow f is a global minimum of

27

Proof (cnt)Claim: flow f is an equilibrium

flow for instance (G,r,c).Suppose: Player i could strictly decrease cost

in equilibrium flow f by deviating from path P to P’ , yielding the new flow f’ :