Documente Academic
Documente Profesional
Documente Cultură
Wiley Encyclopedia of Operations Research and Management Science, edited by James J. Cochran
Copyright © 2010 John Wiley & Sons, Inc.
1
2 WARDROP EQUILIBRIA
a route r is cr (f ) := a∈r ta (fa ). Let X be the a sensitivity analysis of Wardrop equilibria,
set of feasible flows (f , h) and Xf its projec- which can be useful when testing the
tion into the space of arc flows. Note that this robustness of the model [12].
set is a polytope given simply by the flow- Yet another important characterization of
conservation constraints for each commodity traffic equilibrium problems, due to Smith
on every node (see Multicommodity Flows [13], consists of reformulating Equation
and Ahuja et al. [11]). (2) as a variational inequality problem,
Interpreting Wardrop’s first principle as see also Dafermos [14] and Variational
requiring that all flow travels along shortest Inequalities. Accordingly, a flow f is a user
paths, a flow h is called a Wardrop equilib- equilibrium if and only if
rium if and only if for all k ∈ C, we have that
ta (fa )fa ≤ ta (fa )xa , for all flows x ∈ Xf .
cr (h) = min cq (h), (1) a∈A a∈A
q∈Rk
(4)
for all r ∈ Rk such that hr > 0. Beckmann
et al. [7] proved that such a flow always Note that this inequality is a direct con-
exists by considering the following min-cost sequence of the fact that in equilibrium,
multicommodity flow problem with separable users travel on shortest paths with respect
objective function: to arc costs ta (fa ). Another reformulation that
is useful to characterize equilibria in very
fa general settings is given by nonlinear com-
min ta (z) dz : f ∈ Xf . (2) plementary problems [15] (see Complemen-
a∈A 0
tarity Problems).
Charnes and Cooper [16] were the first to
The previous problem is convex because the
notice that the concepts of Nash and Wardrop
objective is the integral of a nondecreasing
equilibria are related. Haurie and Marcotte
function and, since its domain is a com-
[17] prove that a Nash equilibrium in a net-
pact set, attains its optimum (for some back-
work game with a finite number of players
ground on convex optimization. Actually, it
converges to a Wardrop equilibrium when the
can be proved that its first-order optimality
number of players increases (for some back-
conditions are
ground on game theory, see the section titled
‘‘Noncooperative Strategic-Form Games’’ in
cr (h) ≤ cq (h), for all k ∈ C and all routes
this encyclopedia). For this reason, although
r, q ∈ Rk such that hr > 0, (3) the solution concepts are different, a Wardrop
equilibrium can be viewed as an instance of
which is equivalent to Equation (1). a Nash equilibrium in a game with a large
If cost functions ta are strictly increasing, number of players. De Palma and Nesterov
f is unique (but there can be different [18] look at generalizations and alternative
flow decompositions h). For the case of definitions of the basic model and established
nondecreasing costs, the vector of costs conditions that guarantee the existence of
(ta (fa ))a∈A is unique under the possibly equilibria. For example, Wardrop equilibria
nonunique equilibria. Computationally, still exist if cost functions are only required
Equation (2) implies that an equilibrium to be lower semicontinuous. Marcotte and
can be computed efficiently using general Patriksson [19] also discuss alternative def-
convex optimization techniques (see the initions of equilibria in network games and
section titled ‘‘Computation of Wardrop the relationships between them.
Equilibria’’)1 . A formulation similar to the
one presented here can be used to perform
though, uniqueness is not guaranteed even as flow
on arcs because Equation (2) is not necessarily a
1 Inthe case of nonmonotone costs, the solutions convex problem and hence, may admit multiple
to Equation (2) are also equilibria. In this case, local optima.
WARDROP EQUILIBRIA 3
Equation (1) and dk = δk (minq∈Rk cq (h)). direction, it computes a shortest path with
Depending on the separability of the costs, respect to the prevailing traffic conditions.
both the optimization and variational In the subsequent line search, the original
inequality problems can be extended to solve nonlinear problem is solved restricted to the
models with elastic demand. segment defined by the feasible direction of
Finally, several authors have looked at descent. The algorithm terminates when a
other ways to relax some of the basic assump- certain precision is achieved. To determine
tions. We mention a few examples here. when this is the case, the convexity of
A classical extension to the basic model incor- the objective function is used to derive a
porates capacity and other side-constraints lower bound on the value of an optimal
to the equilibrium model [25] as a way of solution. Alternatively, one can compute
improving the solution quality and correcting the gap—defined as the deviation from
the link travel cost functions so as to bring the the shortest path—in the current solution
flow pattern into agreement with the antici- and terminate when it is smaller than a
pated results. This can be handled by adding threshold. It is well known that this algo-
the constraints to the formulations and rein- rithm always converges to a global minimum
terpreting their dual variables as queuing because Equation (2) is a convex program.
delay caused by congestion. However, this The standard Frank–Wolfe algorithm
approach is controversial and several authors sometimes shows poor convergence because
have looked at alternative models that work it tends to zigzag around the equilibrium
better in some situations [26]. Also, Gabriel solution [8,30–32]. Because ‘‘. . . the
and Bernstein [27] studied the case of nonad- (Frank–Wolfe) algorithm is considered
ditive models where the cost of a path can be sufficiently good for practical use’’ [30, p.
given by more general expressions than just 100], most of the commercial implementa-
the sum over all arcs in the path. Examples tions use this procedure. Still, there are
of this include taking into account the many algorithms that were developed to
variability of travel times and risk-aversion address the slow convergence times. Below,
[28], tolls, and valuation of time. we summarize a few of the main approaches.
LeBlanc et al. [33] introduced an improved
version of the previous algorithm called Par-
COMPUTATION OF WARDROP EQUILIBRIA tan (parallel tangents), which was further
studied by Florian et al. [34] and Arezki and
In this section, we discuss some compu- Van Vliet [35], among others. This improve-
tational approaches to finding Wardrop ment is based on a more intelligent line
equilibria. Let us start by describing the search. It determines the descent direction
Frank–Wolfe method [29], which is a tra- using the results of two consecutive itera-
ditional algorithm that has been used to tions, thereby diminishing the zigzagging
compute equilibria. It is an iterative descent effect. These two methods belong to a class
method that works with the formulation called partial linearization algorithms in
shown in Equation (2) and eventually which the objective function is simplified to
converges to the equilibrium. The algorithm be able to find a search direction.
keeps a current solution, and solves a lin- The structure of Equation (2) leads to
earized version of Equation (2) at every step decomposition algorithms, which separate
to determine a feasible descent direction. the main problem into subproblems. The
Referring to the objective of that problem by Frank–Wolfe algorithm is an example of this
T(f ) and to the current solution by f i , the general method since it considers OD pairs
linearized objective is T(f i ) + ∇T(f i ) · (f − f i ). separately after fixing the prevailing flows in
The linearization enables the algorithm one iteration. But the separation can be done
to decompose the problem by OD pairs, in other ways, and viewed as a block version
allowing it to find a shortest path in Rk of the Gauss–Seidel and Jacobi algorithms.
for each commodity k ∈ C independently of For example, a common decomposition
each other. To identify the steepest descent separates flows by node of origin, whereby
WARDROP EQUILIBRIA 5
every iteration assigns all destinations for since it admits a convex program formulation.
each origin at the same time. A good example In contrast, the asymmetric case, requires
of this approach is given by Bar-Gera’s algo- the machinery of variational inequalities
rithm [36], which is one of the most efficient, or nonlinear complementarity problems.
in existence, to compute Wardrop equilibria. There exist standard algorithms to solve
The class of column generation algorithms these classes of problems and some of the
deal with a path formulation of the model. variants presented earlier for the separable
Since it is computationally challenging to case can be extended to this setting. Notice
keep track of the flow along all routes as that since the nonseparable case does not
opposed to maintaining a vector of flows per admit a convex programming formulation,
arc, instead of having one variable per route checking convergence must rely on regret
initially, a column generation algorithm or other related measures.
adds routes at the time they are needed. The search for efficient algorithms to
After discovering new routes with the search compute Wardrop equilibria for the various
direction procedure, an algorithm of this classes of models is a very active area
type forms a restricted master problem that of research. Some of the latest efficient
consists of a path formulation of Equation approaches are due to Dial [37], Florian
(2) using only the routes discovered thus et al. [38], Gentile [39], and Bar-Gera [40].
far (see also Column Generation). These To conclude the discussion on computation,
methods are especially important when costs we note that there are some test problems
along routes are not additive or when there available in the Transportation Network Test
are constraints based on paths because an Problems website [41] that are typically used
arc formulation is not powerful enough to to study new algorithms.
represent the problem in that case. The class There exist many commercial software
of simplicial decomposition algorithms finds packages that implement some of the algo-
the next iterate using the restricted master rithms described in this section. These and
problem. Since all the route information some additional packages also implement
previously computed is badly utilized by other variants of traffic assignment problems
algorithms that perform line search, this such as dynamic models that explicitly
class can solve problems more efficiently, incorporate time [42–44], and simulation
albeit doing more work per iteration. models that consider finer behavioral details
The method of successive averages is a that analytical models cannot handle [45].
commonly used heuristic method for comput- A nonexhaustive list of software implemen-
ing Wardrop equilibria. This method starts tations is AIMSUN, CUBE, CONTRAM,
by computing the costs on all arcs for an arbi- DYNAMIT, DYNASMART, EMME/2, PARA-
trary feasible flow. Iteratively, it computes a MICS, TRANSCAD, TRANSIMS, TSIS-COR-
new solution using an auxiliary linear pro- SIM, SATURN, VISUM-VISSIM, VISTA,
gram that keeps costs fixed, and updates and UROAD-UTPS.
the current solution by averaging it with
the new one using a factor that depends
on the iteration. This technique is specially EFFICIENCY OF WARDROP EQUILIBRIA
useful for more complicated models where
exact techniques are not readily available. Since an equilibrium model considers that
Some examples are the dynamic and stochas- users unilaterally choose their routes to
tic traffic assignments, see the section titled minimize their route cost, the solution is not
‘‘More Advanced Models’’. necessarily efficient. A natural question is
The case of elastic demands can be thus to quantify how inefficient a Wardrop
incorporated in the previous discussion since equilibrium may be, where efficiency is
it involves adding another term to the convex measured as the flow’s total travel time
minimization problem. The case of nonsepa- C( f ) := r∈R cr (h) hr = a∈A ta ( fa ) fa . Follo-
rable, symmetric cost functions can be han- wing Wardrop’s second principle [5] that
dled similarly to what was described earlier states that users minimize the total travel
6 WARDROP EQUILIBRIA
time in the system, a system optimum + fa∗ ta (fa ) − ta fa∗
f ∗ is an optimal solution to the min-cost a∈A: fa > fa∗
multicommodity flow problem:
1
≤ ta fa∗ fa∗ + ta (fa )fa ,
min C(f ) : f ∈ Xf . (5) a∈A
4
a∈A
It is not hard to observe that, in general, the implying that C(f ) ≤ (4/3) · C(f ∗ ). The first
total travel time incurred by an equilibrium inequality in the previous derivation holds
can be arbitrarily larger than that of a social because of Equation (4), while the last
optimum. Consider, for instance, a two-node inequality follows since for affine functions
two-link network with unit demand and cost the shaded area in Fig. 2 is at most 25% of
functions given by t1 (f1 ) = 1 and t2 (f2 ) = f2n , the area of the big rectangle. This theorem
for some large value of n (Fig. 1). At equilib- is due to Roughgarden and Tardos [46], and
rium, all flow will use the second link, so that the short proof presented here appeared
the total travel time will be 1. On the other in Correa et al. [49]. The result implies
hand in the system optimum a small fraction that, in the worst case among all possible
of the flow will use the first link, so that its networks, the inefficiency introduced by the
total travel time will be close to 0, making self-minded behavior of an equilibrium is
their ratio grow to infinity. Nevertheless, a never worse than 1/3.
sequence of results initiated by Roughgar- The proof above easily extends to other
den and Tardos [46] and further developed in classes of link travel cost functions by
papers [47–51] states that if we only allow only changing the 25% with the corre-
link travel cost functions belonging to a cer- sponding quantity for a given class of
tain class, then the total travel time of an functions. Probably the most interesting
equilibrium is at most a constant times that aspect of this result is that the efficiency
of the system optimum. loss depends on the allowable functions
Let us illustrate this group of results by rather than on the topology of the network.
considering the case in which for all a ∈ A, Furthermore, the idea behind this proof
ta (·) is an affine function with nonnegative provides another interesting result first
coefficients. Consider an equilibrium flow f derived in Roughgarden and Tardos [46].
and a system optimal flow f ∗ , then we have Indeed, note that if ta (·) are
arbitrary non-
that decreasing functions, then a∈A xa ta (fa ) ≤
C(f ) ≤ ta (fa )fa∗ = ta fa∗ fa∗
a∈A a∈A
∗ ∗
+ fa∗ ta (fa ) − ta fa∗ ≤ t a fa fa
a∈A a∈A
ta(fa)
t1(f1) = 1
ta(.)
ta(f*a)
1 1
ba
t2(f2) = f n2
a∈A max{fa ta (fa ), xa ta (xa )} ≤ a∈A fa ta (fa ) + (players) have a priority and thus, in any
a∈A xa ta (xa ), where the fa ’s and xa ’s are given link a flow particle only experiences
any nonnegative numbers. Consider now an a travel time ta (xa ), where xa is the amount
equilibrium flow f of a given instance, and a of flow using link a, having higher priority.
system optimal flow x of a similar instance Interestingly in this context a system
where demands rates d are doubled. From optimum is given by Equation (2).
Equation (4), and since x/2 ∈ Xf , we have
C(f ) = 2C(f ) − C(f ) ≤ 2 ta (fa ) · (xa /2) MORE ADVANCED MODELS
a∈A
This section discusses some extensions of the
− C(f ) ≤ ta (fa ) · xa − C(f ) ≤ C(x). basic Wardrop equilibrium model that con-
a∈A
sider variations on the structure of offer and
In other words, for arbitrary nondecreas- demand in the network.
ing link travel cost functions, the cost of a In most urban transportation networks,
Wardrop equilibrium is at most the cost of an commuters do not have to pay the cost they
optimal solution with the demand doubled. impose to others by a particular route choice,
For a restricted set of cost functions, one can leading to the bad utilization of the avail-
provide improved results [49]. For example, able capacity alluded to in the section titled
under affine costs one can prove that the same ‘‘Efficiency of Wardrop Equilibria’’. Since con-
statement holds with 25% more demand. gestion increases sharply with road utiliza-
As general equilibria typically do not tion, having relatively few drivers switch
minimize the social cost, Koutsoupias and to other routes may significantly improve
Papadimitriou [52] proposed to analyze the commute times. Starting with the seminal
inefficiency of equilibria from a worst-case idea of Vickrey [55,56], many transportation
perspective; this led to the notion of ‘‘price economists have advocated the use of conges-
of anarchy’’ [53], which is the ratio of the tion pricing to achieve this goal. The scheme
worst social cost of a Nash equilibrium to the forces drivers to pay a toll when entering
cost of an optimal solution. In the context congested areas. The underlying idea is to
of our traffic model this quantity has been charge drivers the externality they impose
analyzed in a series of papers for increasingly to others because when commuters inter-
more general classes of cost functions and nalize these externalities, the corresponding
other model features. The result previously choices maximize the system welfare. Singa-
described implies that this worst-case ratio pore introduced congestion pricing in 1975,
is 4/3. This was extended to more general London in 2003 [57,58], and Stockholm in
link travel cost functions by Roughgarden 2007. Increasingly, many large cities have
[47] and by Correa et al. [48], who basically been debating whether a congestion pricing
proved that the efficiency loss in this setting scheme should be adopted. Nevertheless, it
is independent of the topology of the network. has been very hard to implement congestion
Chau and Sim [50] considered the case of pricing because of technical, economical, and
nonseparable, symmetric cost functions political problems. There is a large body of
with elastic demands. They proved that the research on finding the set of tolls for a given
efficiency loss can be bounded in a similar network that optimizes a given objective (e.g.,
way as what is described here. Perakis social welfare, revenue, number of tollbooths)
[51] considered general nonseparable cost under constraints (such as a budget balance,
functions and proved upper bounds using maximum number of tollbooths, or restric-
variational inequalities as well. Her bounds tions on their location). For more details, see
depend on two parameters that measure the books [59,60].
the asymmetry and the nonlinearity of the Stochastic user equilibrium models date
cost functions considered. Farzad et al. [54] back to the 1970s, when Dial [61] proposed a
provide results of a similar flavor in a closely model where the demand on each OD is dis-
related model. In their setting the flow tributed among routes (with random lengths)
8 WARDROP EQUILIBRIA
according to a logit distribution, in the case of considered the frequency of a bus line as a
uncongested traffic networks. To reduce route function of the flow in the network and called
enumeration he considered that flow is dis- this function ‘effective frequency.’ They proved
tributed only among ‘‘efficient routes.’’ This that an equilibrium in this context exists via
model has been widely studied and extended an equivalent optimization problem. Cepeda
[62,63]. Furthermore, Daganzo and Sheffi et al. [82] obtained a new characterization of
[64] looked at the case of dependent route equilibria in the congested setting. This led to
costs, while Fisk [65] studied the model in an effective algorithm that is currently part
the context of congested networks, obtaining of EMME/2.
an equivalent optimization problem. Meth- In some situations, such as logistics
ods avoiding route enumeration have been networks, it is natural to consider that some
proposed by Bell [66], Larsson et al. [67], players control nonnegligible amounts of
and Maher [68], also leading to equivalent flow that can be split among several routes,
optimization problems in the spirit of Fisk’s. as modeled by an atomic network game [83].
Based on the work of Akamatsu [69], Baillon In this context, the equilibrium becomes
and Cominetti [70] proposed a more general significantly more difficult to characterize
concept called Markovian traffic equilibrium, and compute because, even for the basic
provided an equivalent optimization prob- assumptions, the game is generally not
lem, and established the convergence of the potential. Although the existence of equi-
method of successive averages in that con- libria is still guaranteed [84], an instance
text. There are complementary models that may possess multiple equilibria [85], and no
consider that travel times themselves are equivalent convex optimization problem is
stochastic instead of considering that the per- known even for general separable instances.
ception is stochastic. Some examples are the Furthermore, surprisingly an equilibrium of
papers [28,71–75]. this game can be less efficient than that of
Another important extension is modeling the corresponding nonatomic instance (as in
public transportation systems. The time com- the section titled ‘‘The Basic Model’’), even in
muters need to wait until a bus arrives simple networks with two OD pairs [86,87].
to the stop adds a difficulty that was not Cominetti et al. [87] generalized the results
present in models of privately owned vehi- of the section titled ‘‘Efficiency of Wardrop
cles. Indeed, in order to balance waiting and Equilibria’’ to this setting, getting a bound
travel time, users’ strategies may involve of 3/2 when cost functions are afine. It is an
selecting a subset of bus lines and boarding open question whether this bound is tight.
the first available one. This idea was pio- Swamy [88] proved that tolls that induce
neered by Chiriqui and Robillard [76], who an optimal routing of the game with finite
considered a network with a single OD pair players always exist.
and n bus lines serving it, each characterized Generalizing a classic paper by Rosenthal
by a travel time and a frequency. They pro- [89], Milchtaich [90] considered a generaliza-
vided a simple efficient algorithm to solve the tion of the basic model referred to by con-
problem. Spiess and Florian [77] generalized gestion games. In this model the network is
that model to arbitrary networks introducing abstracted away and each user selects a strat-
the notion of general users’ strategies. This egy that consists of a subset of arcs (instead
was further put in graph theoretic terms of a route), selected among a set of feasible
by Nguyen and Pallotino [78], who called strategies defined in advance. The defining
these strategies hyperpaths and incorporated characteristic of these models is that a link
travel times dependent on congestion. How- travel cost function for one arc just depends
ever, besides increasing travel times, conges- on its demand. Milchtaich himself and oth-
tion also increases waiting times because a ers have looked at existence, uniqueness,
user may not be able to board a selected bus. and computation of equilibria in congestion
This is challenging and Wu et al. [79] and games with and without atoms, and with
Bouzaı̈ene-Ayari et al. [80] attempted to deal homogeneous and heterogeneous cost func-
with this problem. Cominetti and Correa [81] tions. Actually, some of the results presented
WARDROP EQUILIBRIA 9
14. Dafermos SC. Traffic equilibrium and vari- 28. Ordóñez F, Stier-Moses NE. Wardrop equi-
ational inequalities. Transp Sci 1980;14(1): libria with risk-averse users. Transp Sci
42–54. 2010;44(1):63–86.
15. Aashtiani HZ, Magnanti TL. Equilibria on 29. Frank M, Wolfe P. An algorithm for quadratic
a congested transportation network. SIAM J programming. Nav Res Log Q 1956;3:
Algebr Discrete Methods 1981;2(3):213–226. 95–110.
16. Charnes A, Cooper WW. Multicopy traf- 30. Patriksson M. The traffic assignment prob-
fic network models. In: Herman R, editor. lem: models and methods. Utrecht, The
Theory of traffic flow. Amsterdam: Elsevier; Netherlands: VSP; 1994.
1961. pp. 85–96. Proceedings of the Sympo- 31. Florian M, Hearn DW. Network equilib-
sium on the Theory of Traffic Flow held at rium models and algorithms. In: Ball MO,
the General Motors Research Laboratories; Magnanti TL, Monma CL, editors. Volume
Warren (MI); 1959. 8, Network routing, Handbooks in Opera-
17. Haurie A, Marcotte P. On the relationship tions Research and Management Science.
between Nash-Cournot and Wardrop equi- New York: Elsevier; 1995. Chapter 6. pp.
libria. Networks 1985;15(3):295–308. 485–550.
18. de Palma A, Nesterov Y. Optimization formu- 32. Patriksson M. Algorithms for computing traf-
lations and static equilibrium in congested fic equilibria. Netw Spat Econ 2004;4:23–38.
transportation networks. CORE Discussion 33. LeBlanc LJ, Helgason RV, Boyce DE.
Paper 9861. Belgium: Université Catholique Improved efficiency of the Frank-Wolfe algo-
de Louvain, Louvain-la-Neuve; 1998. rithm for convex network programs. Transp
Sci 1985;19:445–462.
19. Marcotte P, Patriksson M. Traffic equilib-
rium. In: Barnhart C, Laporte G, editors. 34. Florian M, Guélat J, Spiess H. An effi-
Volume 14, Transportation, Handbooks in cient implementation of the ‘‘Partan’’ vari-
Operations Research and Management Sci- ant of the linear approximation method for
ence. New York: Elsevier; 2007. Chapter 10. the network equilibrium problem. Networks
pp. 623–713. 1987;17:319–339.
35. Arezki Y, Van Vliet D. A full analytical imple-
20. Transportation Research Board. Highway
mentation of the Partan/Frank-Wolfe algo-
capacity manual 2000; 2000.
rithm for equilibrium assignment. Transp
21. Bureau of Public Roads. Traffic assignment Sci 1990;24(1):58–62.
manual. Washington (DC): U.S. Department
36. Bar-Gera H. Origin-based algorithm for
of Commerce, Urban Planning Division;
the traffic assignment problem. Transp Sci
1964.
2002;36(4):398–417.
22. Sandholm WH. Potential games with con-
37. Dial RB. A path-based user-equilibrium
tinuous player sets. J Econ Theory 2001;97:
traffic assignment algorithm that obviates
81–108. path storage and enumeration. Transp Res
23. Monderer D, Shapley LS. Potential games. 2006;B40(10):917–936.
Games Econ Behav 1996;14:124–143. 38. Florian M, Constantin I, Florian D. A new
24. Hearn DW, Lawphongpanich S, Nguyen look at the projected gradient method for
S. Convex programming formulations of equilibrium assignment. Transp Res Rec
the asymmetric traffic assignment problem. 2009;2090:10–16.
Transp Res 1984;18B(4-5):357–365. 39. Gentile G. Linear user cost equilibrium: a
25. Hearn DW. Bounding flows in traffic new algorithm for traffic assignment. Work-
assignment models. Technical Report 80- ing paper. 2010.
4. Gainesville (FL): Department of Indus- 40. Bar-Gera H. Traffic assignment by paired
trial and Systems Engineering, University of alternative segments. Transp Res Part B
Florida; 1980. 2010;44(8):1022–1046.
26. Marcotte P, Nguyen S, Schoeb A. A strategic 41. Bar-Gera H. Transportation network test
flow model of traffic assignment in static problems; 2002. Available at http://www.bgu.
capacitated networks. Oper Res 2004;52(2): ac.il/∼bargera/tntp/. Accessed 2010 Aug 28.
191–212. 42. Merchant DK, Nemhauser GL. A model
27. Gabriel S, Bernstein D. The traffic equilib- and an algorithm for the dynamic traffic
rium problem with nonadditive path costs. assignment problems. Transp Sci 1978;12:
Transp Sci 1997;31:337–348. 183–199.
WARDROP EQUILIBRIA 11
43. Ben-Akiva ME. Dynamic network equilib- 58. Santos G, Fraser G. Road pricing: lessons
rium research. Transp Res 1985;19A(5/6): from London. Econ Policy 2006;21(46):
429–431. 263–310.
44. Peeta S, Ziliaskopoulos AK. Foundations of 59. Yang H, Huang H-J. Mathematical and eco-
dynamic traffic assignment: the past, the nomic theory of road pricing. Oxford: Else-
present and the future. Netw Spat Econ vier; 2005.
2001;1:233–265.
60. Lawphongpanich S, Smith MJ, Hearn DW,
45. Kitamura R, Kuwahara M, editors. Volume editors. Volume 101, Mathematical and
31, Simulation approaches in transportation computational models for congestion charg-
analysis, Operations Research/Computer ing, Applied optimization. Berlin: Springer;
Science Interfaces Series. New York: 2006.
Springer; 2005.
61. Dial RB. A probabilistic multi-path traffic
46. Roughgarden T, Tardos E. How bad is selfish
assignment algorithm which obviates path
routing? J ACM 2002;49(2):236–259.
enumeration. Transp Res 1971;5(2):83–111.
47. Roughgarden T. The price of anarchy is inde-
62. Trahan M. Probabilistic assignment: an algo-
pendent of the network topology. J Comput
rithm. Transp Sci 1974;8(4):311–320.
Syst Sci 2003;67(2):341–364.
63. Sheffi Y, Powell W. An algorithm for the
48. Correa JR, Schulz AS, Stier-Moses NE. Self-
traffic assignment problem with random link
ish routing in capacitated networks. Math
Oper Res 2004;29(4):961–976. costs. Networks 1982;12:191–207.
49. Correa JR, Schulz AS, Stier-Moses NE. 64. Daganzo C, Sheffi Y. On stochastic models
A geometric approach to the price of anar- of traffic assignment. Transp Sci 1977;11(3):
chy in nonatomic congestion games. Games 253–274.
Econ Behav 2008;64:457–469. 65. Fisk C. Some developments in equilibrium
50. Chau CK, Sim KM. The price of anarchy for traffic assignment. Transp Res 1980;14B(3):
non-atomic congestion games with symmet- 243–255.
ric cost maps and elastic demands. Oper Res 66. Bell M. Alternatives to Dial’s logit assign-
Lett 2003;31(5):327–334. ment algorithm. Transp Res 1995;29B:
51. Perakis G. The ‘‘price of anarchy’’ under non- 287–295.
linear and asymmetric costs. Math Oper Res 67. Larsson T, Liu Z, Patriksson M. A dual
2007;32(3):614–628. scheme for traffic assignment problems.
52. Koutsoupias E, Papadimitriou CH. Worst- Optimization 1997;42:323–358.
case equilibria. In: Meinel C, Tison S, edi- 68. Maher M. Algorithms for logit-based stochas-
tors. Volume 1563, Proceedings of the 16th tic user equilibrium assignment. Transp Res
Annual Symposium on Theoretical Aspects of 1998;32B(8):539–549.
Computer Science (STACS), Lecture Notes in
69. Akamatsu T. Cyclic flows, Markov processes
Computer Science; Trier, Germany. Heidel-
berg: Springer; 1999. pp. 404–413. and stochastic traffic assignment. Transp
Res 1996;30B(5):369–386.
53. Papadimitriou CH. Algorithms, games, and
the Internet. Proceedings of the 33rd Annual 70. Baillon J-B, Cominetti R. Markovian traf-
ACM Symposium on Theory of Computing fic equilibrium. Math Program 2008;111B:
(STOC)., Hersonissos, Greece. New York: 33–56.
ACM Press; 2001. pp. 749–753. 71. Liu H, Ban X, Ran B, et al. An ana-
54. Farzad B, Olver N, Vetta A. A priority-based lytical dynamic traffic assignment model
model of routing. Chic J Theor Comput Sci with stochastic network and travelers’
2008;1–29. perceptions. Transp Res Rec 2002;1783:
55. Vickrey WS. A proposal for revising New 125–133.
York’s subway fare structure. J Oper Res Soc 72. Mirchandani PB, Soroush H. Generalized
Am 1955;3(1):38–68. traffic equilibrium with probabilistic travel
56. Vickrey WS. Congestion theory and trans- times and perceptions. Transp Sci 1987;21:
port investment. Am Econ Rev 1969;59: 133–152.
251–261. 73. Uchida T, Iida Y. Risk assignment: a new
57. Santos G. Urban congestion charging: a traffic assignment model considering risk of
comparison between London and Singapore. travel time variation. In: Daganzo CF, edi-
Transp Rev 2005;25(5):511–534. tor. Proceedings of the 12th International
12 WARDROP EQUILIBRIA