Deux problèmes sur les réseaux
1er octobre 2008
Ecole des Ponts, France
Affectation dynamique du trafic : il y a un équilibre ! avec Nicolas Wagner.
Traffic assignment
Traffic assignment
Traffic assignment
Traffic assignment
Static traffic assignment
Supply A directed graphG = (V,A), acost function ca :R+ →R+for eacha∈A.Assumptions : ca isnon-decreasingandcontinuousandGis strongly connected.
Demand Ademand functionb :V ×V →R+.
Realization Aflowx : R →R+ whereR = (Ro,d)(o,d)∈V×V, such that X
r∈Ro,d
x(r) = b(o,d), withRo,d being the set of allo–d route ofG.
Equilibrium Wheneverr,r0 ∈Ro,d, we have
x(r)>0⇒X
a∈r
ca(xa)≤ X
a∈r0
ca(xa),
Static traffic assignment
Supply A directed graphG = (V,A), acost function ca :R+ →R+for eacha∈A.Assumptions : ca isnon-decreasingandcontinuousandGis strongly connected.
Demand Ademand functionb :V ×V →R+.
Realization Aflowx : R →R+ whereR = (Ro,d)(o,d)∈V×V, such that X
r∈Ro,d
x(r) = b(o,d), withRo,d being the set of allo–d route ofG.
Equilibrium Wheneverr,r0 ∈Ro,d, we have x(r)>0⇒X
a∈r
ca(xa)≤ X
a∈r0
ca(xa),
wherexa :=P
r∈R:a∈rx(r)(Wardrop ).
Static traffic assignment
Supply A directed graphG = (V,A), acost function ca :R+ →R+for eacha∈A.Assumptions : ca isnon-decreasingandcontinuousandGis strongly connected.
Demand Ademand functionb :V ×V →R+.
Realization Aflowx : R →R+ whereR = (Ro,d)(o,d)∈V×V, such that X
r∈Ro,d
x(r) = b(o,d), withRo,d being the set of allo–d route ofG.
Equilibrium Wheneverr,r0 ∈Ro,d, we have x(r)>0⇒X
a∈r
ca(xa)≤ X
a∈r0
ca(xa),
Static traffic assignment
Supply A directed graphG = (V,A), acost function ca :R+ →R+for eacha∈A.Assumptions : ca isnon-decreasingandcontinuousandGis strongly connected.
Demand Ademand functionb :V ×V →R+.
Realization Aflowx : R →R+ whereR = (Ro,d)(o,d)∈V×V, such that X
r∈Ro,d
x(r) = b(o,d), withRo,d being the set of allo–d route ofG.
Equilibrium Wheneverr,r0 ∈Ro,d, we have x(r)>0⇒X
a∈r
ca(xa)≤ X
a∈r0
ca(xa),
wherexa :=P
r∈R:a∈rx(r)(Wardrop ).
Static traffic assignment : results
Theorem
There is always an equilibrium. Moreover, the values ca P
r∈R:a∈rx(r)
at equilibrium are unique.
Formulation as a convex program→computable.
Static traffic assignment : limitation
This model has alimited significance:
the time needed to travel along the route is not modelled.Each user occupies the whole route continuously.
→need ofdynamic traffic assignment model.
Static traffic assignment : limitation
This model has alimited significance:
the time needed to travel along the route is not modelled.Each user occupies the whole route continuously.
→need ofdynamic traffic assignment model.
Dynamic traffic assignment
Since the 70’s, several models has been proposed, for instance :
Vickrey (1969)
Merchant and Nemhauser (1978) Friez and al. (1989)
90’s : Leurent (LADTA), Bellei, Gentile and Papola, Akamatsu and Kuwahara
Roughly speaking, the models are such that
Dynamic traffic assignment model
Time intervalI = [0,H].
Users aredynamic flows:x :R×I. The quantityx(r,h)is the number ofusers choosing the router at timeh.
ya :I →R+. The quantityya(h)is the number ofusers entering the arcaat timeh.
Useful notation : Xr(h) :=
Z h h0=0
x(r,h0) and Ya(h) :=
Z h h0=0
ya(h0)
(cumulated flow).
Dynamic traffic assignment model
Supply A strongly connected directed graphG = (V,A), aarc travel time functiontafor eacha∈Asuch thatta(Ya)is aI →R+ continuous map,
depending continuously of the cumulated flowYa. Demand Ademand functionb :V ×V ×I →R+.
Realization X
r∈Ro,d
x(r,h) =b(o,d,h)for all(o,d)∈V ×V andh∈I.
Equilibrium Wheneverr,r0 ∈Ro,d, we have
x(r,h)>0⇒tr(X~)(h)≤tr0(X~)(h),
where forr =a1. . .an, we havetr(X~)(h) := Pn
i=1tai(Yai)(hi) withYai :=φai(X)and
h1:=handhi+1:=hi +tai(Yai)(hi)fori =1, . . . ,n−1.
Dynamic traffic assignment model
Supply A strongly connected directed graphG = (V,A), aarc travel time functiontafor eacha∈Asuch thatta(Ya)is aI →R+ continuous map,
depending continuously of the cumulated flowYa. Demand Ademand functionb :V ×V ×I →R+.
Realization X
r∈Ro,d
x(r,h) =b(o,d,h)for all(o,d)∈V ×V andh∈I.
Equilibrium Wheneverr,r0 ∈Ro,d, we have
x(r,h)>0⇒tr(X~)(h)≤tr0(X~)(h),
where forr =a . . .a , we havet (X~)(h) := Pn
t (Y )(h)
Dynamic traffic assignment model
Supply A strongly connected directed graphG = (V,A), aarc travel time functiontafor eacha∈Asuch thatta(Ya)is aI →R+ continuous map,
depending continuously of the cumulated flowYa. Demand Ademand functionb :V ×V ×I →R+.
Realization X
r∈Ro,d
x(r,h) =b(o,d,h)for all(o,d)∈V ×V andh∈I.
Equilibrium Wheneverr,r0 ∈Ro,d, we have
x(r,h)>0⇒tr(X~)(h)≤tr0(X~)(h),
where forr =a1. . .an, we havetr(X~)(h) := Pn
i=1tai(Yai)(hi) withYai :=φai(X)and
h1:=handhi+1:=hi +tai(Yai)(hi)fori =1, . . . ,n−1.
Dynamic traffic assignment model
Supply A strongly connected directed graphG = (V,A), aarc travel time functiontafor eacha∈Asuch thatta(Ya)is aI →R+ continuous map,
depending continuously of the cumulated flowYa. Demand Ademand functionb :V ×V ×I →R+.
Realization X
r∈Ro,d
x(r,h) =b(o,d,h)for all(o,d)∈V ×V andh∈I.
Equilibrium Wheneverr,r0 ∈Ro,d, we have
x(r,h)>0⇒tr(X~)(h)≤tr0(X~)(h),
where forr =a . . .a , we havet (X~)(h) := Pn
t (Y )(h)
Dynamic traffic assignment
TheYacan be computed from thetaknowing allXr (under some assumptions).
Existence of an equilibrium ? In general, it is an open question.
Zhu and Marcotte prove an equilibrium for Friez’s model (with a stronger assumptions on travel time).
Ladta : Unknown.
Lindsey + Leurent, Wagner studied time departure choice on one arc (Vickrey model) : their is an equilibrium.
Dynamic traffic assignment
TheYacan be computed from thetaknowing allXr (under some assumptions).
Existence of an equilibrium ? In general, it is an open question.
Zhu and Marcotte prove an equilibrium for Friez’s model (with a stronger assumptions on travel time).
Ladta : Unknown.
Lindsey + Leurent, Wagner studied time departure choice on one arc (Vickrey model) : their is an equilibrium.
Results
Our result
• a general model with minimal assumption :continuity, fifoness (in a weak form), causality, no infinite speed.
• existence of an equilibrium that containsallprevious results.
Traffic assignement as a continuous game
R: set ofroutes,I := [0,H]time interval.
M(R×I): set of measures onR×I (choice of time departure allowed).
Tr : set of continuous mapsM(R×I)→ C(R,R)(tr ∈ Tr is such thattr(~X)(h)gives the time needed to traverse the route r).
user: characterized by a functionu : S
r∈RTr
×R×I →R¯ (that is upper semicontinuous inr andh, and continuous in the travel times, (theutility-function) :u(~t,r,h): a choice(r,h)7→
the payoff, when route travel times are~t := tr1,tr2, . . . ,.
U: space of usersu.
Atraffic gameis a measureU onU.Number of users:
Traffic equilibrium
Arealizationof the traffic game is a measure onU ×R×I.
Denotetr1,tr2, . . .by~t.
Given a traffic gameU, a realizationDis aCournot-Nash equilibriumif we haveDU =U and
D {(u,r,h) : for each(r0,h0)∈R×I, u(~t(DR×I),r,h)≥u(~t(DR×I),r0,h0) o
=N
DR×I is exactly the cumulated flowsX. Set Xr(J) := DR×I({r} ×J).
Khan-Mas-Colell’s theorem ?
Khan-Mas-Colell’s theorem tells us thatthere is a Nash equilibrium, provided that the utility function are continuous in the realization of the game (=here the travel times).
For traffic assignement :DR×I 7→u(~t(DR×I),·,·)mustbe continuous.
Reformulation : Our purpose : prove that~X 7→~t(~X)is continuous.
Khan-Mas-Colell’s theorem ?
Khan-Mas-Colell’s theorem tells us thatthere is a Nash equilibrium, provided that the utility function are continuous in the realization of the game (=here the travel times).
For traffic assignement :DR×I 7→u(~t(DR×I),·,·)mustbe continuous.
Reformulation : Our purpose : prove that~X 7→~t(~X)is continuous.
Khan-Mas-Colell’s theorem ?
Khan-Mas-Colell’s theorem tells us thatthere is a Nash equilibrium, provided that the utility function are continuous in the realization of the game (=here the travel times).
For traffic assignement :DR×I 7→u(~t(DR×I),·,·)mustbe continuous.
Reformulation : Our purpose : prove that~X 7→~t(~X)is continuous.
Arc exit time
Giventa, thearc exit timefunction is
Ha(Y)(h) := h+ta(Y)(h) forY ∈ M(R)andh∈R
Assumptions on travel time
Continuity Ha :M(R)→ C(R)is continuous.
Limited speed there existstmin >0such that for all Y ∈ M(R)and allh∈R, we have Ha(Y)(h)>h+tmin.
Fifo Leth1 <h2inRand letY ∈ M(R). Whenever Y[h1,h2]6=0, we haveHa(Y)(h1)<Ha(Y)(h2).
Causality For allh ∈RandY ∈ M(R), we have Ha(Y|h)(h) = Ha(Y)(h).
Assumptions on travel time
Continuity Ha :M(R)→ C(R)is continuous.
Limited speed there existstmin >0such that for all Y ∈ M(R)and allh∈R, we have Ha(Y)(h)>h+tmin.
Fifo Leth1 <h2inRand letY ∈ M(R). Whenever Y[h1,h2]6=0, we haveHa(Y)(h1)<Ha(Y)(h2).
Causality For allh ∈RandY ∈ M(R), we have Ha(Y|h)(h) = Ha(Y)(h).
Assumptions on travel time
Continuity Ha :M(R)→ C(R)is continuous.
Limited speed there existstmin >0such that for all Y ∈ M(R)and allh∈R, we have Ha(Y)(h)>h+tmin.
Fifo Leth1 <h2inRand letY ∈ M(R). Whenever Y[h1,h2]6=0, we haveHa(Y)(h1)<Ha(Y)(h2).
Causality For allh ∈RandY ∈ M(R), we have Ha(Y|h)(h) = Ha(Y)(h).
Assumptions on travel time
Continuity Ha :M(R)→ C(R)is continuous.
Limited speed there existstmin >0such that for all Y ∈ M(R)and allh∈R, we have Ha(Y)(h)>h+tmin.
Fifo Leth1 <h2inRand letY ∈ M(R). Whenever Y[h1,h2]6=0, we haveHa(Y)(h1)<Ha(Y)(h2).
Causality For allh ∈RandY ∈ M(R), we have Ha(Y|h)(h) = Ha(Y)(h).
Arc flowing function
Let~Yabe inM(R×R); the flow of users following router and leaving the arcaon the time subsetJis :
ψar(~Ya)(J) :=ψa(Y~a)({r}×J) :=
Yar(Ha(Ya)−1(J)) ifa ∈r
0 if not.
(1)
Flows on routes → flow on arcs
AnoutflowofX~ is a measureYa :=P
rYar onRsuch that for everyr =a1a2..an:
1 Yar
1 =Xr
2 Yar
i =ψrai−1(Y~ai−1)fori =2..n
3 Yar =0ifa ∈/ r
Uniqueness and continuity of the outflow
Proposition
GivenX~, there exists a unique outflowYa. Moreover, the map φa :X~ 7→Ya is continuous.
Continuity of t
ris proved !
Continuity oftr is proved.
Indeed
tr(X~)(h) :=
n
X
i=1
tai(Yai)(hi)
withh1 :=handhi+1 :=hi +tai(Yai)(hi)fori = 1, . . . ,n−1 can be rewritten :
tr(~X)(h) = Han
φan(~X)
◦. . .◦Ha1
φa1(~X)
(h)−h for allh ∈I.
Continuity of t
ris proved !
Continuity oftr is proved.
Indeed
tr(X~)(h) :=
n
X
i=1
tai(Yai)(hi)
withh1 :=handhi+1 :=hi +tai(Yai)(hi)fori = 1, . . . ,n−1 can be rewritten :
tr(~X)(h) = Han
φan(~X)
◦. . .◦Ha1
φa1(~X)
(h)−h for allh ∈I.
There is an equilibrium
Continuity of thetr ⇒continuity of the utility functionu(tr,·,·)
⇒we can apply the theorem.
Theorem
Given a directed graphG= (V,A)with arc travel time
functions(ta)a∈Asatisfying assumptions of causality, fifoness, limited speed and continuity and given a measureUon the set of possible users (identified with their utility function), there is a Nash equilibrium.
There is an equilibrium
Continuity of thetr ⇒continuity of the utility functionu(tr,·,·)
⇒we can apply the theorem.
Theorem
Given a directed graphG= (V,A)with arc travel time
functions(ta)a∈Asatisfying assumptions of causality, fifoness, limited speed and continuity and given a measureUon the set of possible users (identified with their utility function), there is a Nash equilibrium.
A versatile model
The model is reallyversatileand contains previous models.
Playing with the mapu(~t(X~),r,h),
1 we cover also the case when there is no possible choice of time departure.
2 we can put arbitrarily taxes on some routes or time departure.
They will always be a Nash equilibrium : contains previous results of equilibrium.
A versatile model
The model is reallyversatileand contains previous models.
Playing with the mapu(~t(X~),r,h),
1 we cover also the case when there is no possible choice of time departure.
2 we can put arbitrarily taxes on some routes or time departure.
They will always be a Nash equilibrium : contains previous results of equilibrium.