• Aucun résultat trouvé

GamesandOptimization,November2016 Ga¨etanFOURNIER LocationGamesonNetworks

N/A
N/A
Protected

Academic year: 2022

Partager "GamesandOptimization,November2016 Ga¨etanFOURNIER LocationGamesonNetworks"

Copied!
32
0
0

Texte intégral

(1)

Location Games on Networks

Ga¨etan FOURNIER

IAST, TSE

Games and Optimization, November 2016

(2)

Uniform distribution General distributions Probabilistic Hotelling game

The model and examples Existence result Efficiency result

Hypothesis on buyers

1 Non-strategic continuum of buyers, distributed on a network generated by a metric graph.

2 They buy a given quantity of a good whose price is fixed: they shop to the closest location.

e1

e2

e3

e4

e5

e6 e7

e8

(3)

Hypothesis on sellers (= players)

1 A fixed number of strategic sellers simultaneously choose their locations.

2 They want to sell as much as possible.

e1

e2

e3

e4 e5

e6

e7

e8

1 3

(4)

Uniform distribution General distributions Probabilistic Hotelling game

The model and examples Existence result Efficiency result

Hypothesis on sellers (= players)

1 A fixed number of strategic sellers simultaneously choose their locations.

2 They want to sell as much as possible.

e1 e2

e3

e4 e5

e6

e7

e8

1 3

(5)

The unit interval [Eaton-Lipsey,1975]

1 For n=2, there exists a pure Nash equilibrium.

2 For n=3, there is no pure Nash equilibrium.

3 For n≥4,there exists a pure Nash equilibrium.

The starSk(r)

1 For n≤k, there exists a pure Nash equilibrium.

2 For n∈ [k+1,3k−2], there is no pure Nash equilibrium.

3 For n≥3k−1, there exists a pure Nash equilibrium.

(6)

Uniform distribution General distributions Probabilistic Hotelling game

The model and examples Existence result Efficiency result

Results with uniform density (Fournier-Scarsini[2015])

ä Existence of pure Nash equilibrium for any graph when the number of player is large enough.

Theorem: Fournier-Scarsini [2015]

On any finite graph, Hotelling games admit a pure Nash equilibrium, provided the number of players is larger than

N∶=3 card(E) + ∑

e∈E

⌈ 5λ(e)

λ ⌉ whereλ=min

E λ( the length of the shortest edge).

(7)

Extensive analysis (also with small number of players): not easy.

Counterexample [D. Palvolgyi]

1 For n=2, there exists a pure Nash equilibrium.

2 For n∈ {3,4}, there is no pure Nash equilibrium.

3 For n∈ {5,6}, there exists a pure Nash equilibrium.

4 For n∈ {7, . . . ,16}, there is no pure Nash equilibrium.

5 For n=17, there exists a pure Nash equilibrium.

(8)

Uniform distribution General distributions Probabilistic Hotelling game

The model and examples Existence result Efficiency result

ä Existence of pure Nash equilibrium for any graph when the number of player is large enough.

ä Efficiency of these equilibria: not for the players (sellers) but for the consumers.

(9)

Efficiency of the equilibria

Traveling distances of consumers, in equilibrium and in social optimum.

Equilibrium social cost: ? Optimum social cost: ?

(10)

Uniform distribution General distributions Probabilistic Hotelling game

The model and examples Existence result Efficiency result

Social costs in equilibrium and in social optimum.

Equilibrium social cost: 18 Optimum social cost: 161

(11)

For x ∈Sn, thesocial costσ(x) is given by:

σ(x) ∶= ∫

S min

i∈{1,...,n}d(xi,y)dy The price of anarchy is given by:

PoA(n) ∶= maxx∈En(H)σ(x) minx∈Snσ(x)

,

The price of stability is given by:

PoA(n) ∶=

minx∈En(H)σ(x) minx∈Snσ(x) , whereEn(H)is the set of equilibrium withn players.

(12)

Uniform distribution General distributions Probabilistic Hotelling game

The model and examples Existence result Efficiency result

On the unit interval, we have:

PoA(n) =

⎧⎪

⎪⎪

⎪⎪

⎪⎩

2 ifn is even, 2( n

n+1) ifn>3 is odd.

Forn≥4

PoS(n) = n n−2

(13)

ä Efficiency of these equilibria: not for the players (sellers) but for the consumers.

Theorem

Suppose that the gameH(n,G) has an equilibrium. Then PoA(n) Ð→

n→+∞2 PoS(n) Ð→

n→+∞1

(14)

Uniform distribution General distributions Probabilistic Hotelling game

The model Counter-example The result

General distributions of consumers [Fournier 2015]

For a subset A ⊂ G, the quantity of consumers lo- cated inAis:

Ag(x)dL(x) where L is the Lebesgue measure, and g >0.

(15)

Counterexamples with general distribution of consumers

Counterexamples

(1) For any>0, the function g ∶ [0,1] →R+: g ∶=x ↦1+x

is arbitrary close to1but the game H(n,[0,1],g)doesn’t admit any Nash equilibrium in pure strategies forn>2.

(2) In fact, when the number of player is small, the class of distributions such that the corresponding Hotelling games admit an exact pure Nash equilibrium is small.

Suppose that x = (x1, . . . ,xn) is an equilibrium (x1≤ ⋅ ⋅ ⋅ ≤xn) in the game H(n,[0,1],f). We first claim that all players are coupled, i.e. thatx1=x2<x3=x4< ⋅ ⋅ ⋅ <xn−1 =xn.

(16)

Uniform distribution General distributions Probabilistic Hotelling game

The model Counter-example The result

→ Suppose that 3 players (or more) share the same location xk =xk+1=xk+2

∣ ∣

xk−1 xk xk+3

pk(x1, . . . ,xn) = 1 3∫

(1+x)dx Either:

∫ (1+x)dx>pk(x1, . . . ,xn) or:

∫ (1+x)dx>pk(x1, . . . ,xn)

(17)

→ Suppose now that there exists a locationxk ∈ [0,1]with a single player k. His payoff is equal to a right trapezoid’s area:

∣ ∣

1+xk−12+xk 1+xk+x2k+1

xk−1

xk−1+xk

2

xk

xk+xk+1

2

xk+1

(18)

Uniform distribution General distributions Probabilistic Hotelling game

The model Counter-example The result

All players are coupled.

0 x1

x2

x3

x4 xn−2

xn−1 xn

xn+1 xn+2 xn+3 A1 A2 A3 A4 . . .

An−1An+1 . . .

p1(x) =p2(x) = A1+A2 2 =A1=A2

◻ A1>A2⇒ player 1 has a profitable deviation: x1−δ.

◻ A2>A1⇒ player 1 has a profitable deviation: x1+δ.

Same: p3(x) =p4(x) =A3=A4

A2=A3

◻ A2<A3⇒ player 2 has a profitable deviation: x3−δ.

◻ A2>A3⇒ player 3 has a profitable deviation: x2+δ.

(19)

Counterexample 2

There exists a pure Nash equilibrium on the unit interval with 4 players and with densityg if and only if g satisfiesQ1

2

=

Q1 4

+Q3 4

2

0 x x+y

2 y 1

A B C D

A=B=C =D⇒x =Q1

4, y=Q3

4, x+y 2 =Q1

2

(20)

Uniform distribution General distributions Probabilistic Hotelling game

The model Counter-example The result

Definition

A profile of actionsx∶= (x1, . . . ,xn)is a pure additive -equilibrium ofH(n,G,g) if and only if for alli ∈ {1, . . . ,n}and all y∈G, we have:

pi(x1, . . . ,xi−1,y,xi+1, . . . ,xn) −pi(x) ≤

Definition

A profile of actionsx ∶= (x1, . . . ,xn) is a pure multiplicative -equilibrium of H(n,G,g) if and only if for alli ∈ {1, . . . ,n} and for ally ∈G we have:

pi(x1, . . . ,xi−1,y,xi+1, . . . ,xn) ≤ (1+)pi(x)

(21)

Asymptotic existence of-equilibrium.

Suppose that:

Ê g isK-Lipschitz

Ë There exist m andM such that for allx, 0<m≤g(x) ≤M Then:

∀>0, ∃N() ∈N, ∀n≥N(),

there exists an −pure equilibrium in the game with n players and density distribution g.

N() ∼ 1

(22)

Uniform distribution General distributions Probabilistic Hotelling game

The model Counter-example The result

Sketch of the proof

1/ We approximate the density functiong by a step function ˆg(1), where1 is a parameter playing a role in the length of the steps.

Becauseg isK-Lipschitz, the step function ˆg(1) is such that

∥g−gˆ(1)∥

≤when 1 is small enough.

2/ We prove that there exists an (exact) equilibrium in pure strategies in the gameH(n,G,gˆ(1)), when the number of players n is larger than a lower-boundN(1). This lower bound increases when1 goes to zero.

(23)

3/ If1 is small enough, the equilibrium constructed in the previous step is a multiplicative-equilibrium in the game H(n,S,g). We obtain therefore a lower boundN()on the number of players n that guarantees the existence of a pure multiplicative

-equilibrium in H(n,G,g).

(24)

Uniform distribution General distributions Probabilistic Hotelling game

The model Results

Probabilistic Hotelling game [Fournier 2016]

Motivation: find a continuous version of the Hotelling game.

Ê n players chose a location in [0,1].

Ë The probability that consumer t∈ [0,1] shops to locationxk

is equal to:

f(∣xk−t∣)

ni=1f(∣xi −t∣)

for a given positive and decreasing function f. Ì The payoff of a player k is equal to:

πk(x1, . . . ,xn) = ∫

1 0

f(∣xk−t∣)

ni=1f(∣xi−t∣) g(t)dt

(25)

Theorem

Suppose thatf is C2, symmetric, strictly positive, decreasing and concave. Then there exists a symmetric equilibrium in pure strategies. This equilibrium is(x, . . . ,x), where x satisfies:

1 0

f(x−t)g(t)

f(x−t) dt =0 (1)

i.e:

[ f

f ∗g] (x) =0 (2)

Remark:x doesn’t depend on the number of players.

(26)

Uniform distribution General distributions Probabilistic Hotelling game

The model Results

è Simple cases: ifg =1, then x= 1

2

è In general: either compute ff ∗g, or because πk(x)is strictly concave inxk, the best response dynamics is well defined.

(27)

Open questions:

(1) Can we approximate the standard Hotelling game with a sequence of continuous games? We can’t with concave functionf.

(2) Impact of price competition?

(3) Voting models

(4) What if consumers do not always buy?

(28)

Uniform distribution General distributions Probabilistic Hotelling game

The model Results

Theorem

Suppose that consumers are distributed on the real line according toN (0, σ) and that they buy iif v−p−f(d) ≥0⇔d ≤δ. This game always has at least one equilibrium in pure strategies:

(1) σδ <

ln 2 2 ,

NE = {(tA,tA+2δ) ∣ tA∈ [α, β] ∈ [−2δ,0]}

(2)

ln 2 2 < δ

σ <

2 ln 2 then

NE = {(−tA,tA) ∣ tA

2 ln 2−δ∈ [0, δ]}

(3)

2 ln 2< δ

σ then

NE = {(0,0)}

(29)

Thank you

(30)

Uniform distribution General distributions Probabilistic Hotelling game

The model Results

Stochastic dominance / Majorization

For a vectorz = (z1, . . . ,zn), we denotez[1]≥ ⋅ ⋅ ⋅ ≥z[n] its decreasing rearrangement.

Definition

Letx,y ∈ [0,1]n be such

n

i=1

xi =

n

i=1

yi

if, for allk ∈ {1, . . . ,n}

k

i=1

x[i]

k

i=1

y[i]. then we say thatx is majorizedbyy (x ≺y).

(31)

Definition

A functionφ∶Rn→R is saidSchur-convex ifx≺yimplies φ(x) ≤φ(y).

Proposition

Ifψ∶R→Ris a convex function, φ(x1, . . . ,xn) =

n

i=1

ψ(xi), thenφis Schur-convex.

(32)

Uniform distribution General distributions Probabilistic Hotelling game

The model Results

Suppose thatf is C2, strictly concave, and bounded below by >0.

Then the functionxk ↦πk(xk,x−k) is concave and (jointly) continuous.

It follows from theNash-Glicksberg theorem that the game admits an equilibrium in pure strategies.

Moreover, due to the symmetry of the problem, there exists at least one symmetric pure Nash equilibrium.

Références

Documents relatifs

In induced live IFT140 RNAi cells, a bright YFP signal was present along the length of the flagellum especially in the distal region whereas the pool of IFT material at

It is interesting to note that for the final project, compared to the control group, users in the experimental group used around the same number of blocks and components, but a

Further, the cost function used made marginal cost dependent on the capacity limit and would not easily lend itself to the discussion of what happens when just the number of

Mean field games (MFG) study Nash equilibria in differential games with infinitely many in- distinguishable agents.. In this note, we investigate the classical question of

To sum up, doubly-stochastic extra-gradient convergence is controlled with a better rate than stochastic extra-gradient (EG) with non-smooth losses; with smooth losses, DSEG

Although the theoretical values of the probabilities are rather intricate, in a first approach, we can neglect these aspects and give some results in an heuristic point of

Constraint Games for Modeling and Solving Pure Nash Equi- libria, Price of Anarchy and Pareto Efficient Equilibria... Constraint Games for Modeling and Solving Pure Nash

We conclude that credible incen- tive equilibria with dierentiable incentive functions do not exist without strong conditions on the payo functions of the players.. On the other