Transportation Research Part C 18 (2010) 157175

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

Road pricing for congestion control with unknown

demand and cost functions
Hai Yang a,*, Wei Xu b, Bing-sheng He c, Qiang Meng d
a

Department of Civil Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong, PR China
School of Management and Engineering, Nanjing University, Nanjing, PR China
Department of Mathematics, Nanjing University, Nanjing, PR China
d
Department of Civil Engineering, The National University of Singapore, Singapore
b
c

a r t i c l e

i n f o

Article history:
Received 29 January 2008
Received in revised form 14 May 2009
Accepted 17 May 2009

Keywords:
Road pricing
Congestion control
Trafc equilibrium
Unknown demand

a b s t r a c t
It is widely recognized that precise estimation of road tolls for various pricing schemes
requires a few pieces of information such as origindestination demand functions, link travel time functions and users valuations of travel time savings, which are, however, not all
readily available in practice. To circumvent this difculty, we develop a convergent trialand-error implementation method for a particular pricing scheme for effective congestion
control when both the link travel time functions and demand functions are unknown. The
congestion control problem of interest is also known as the trafc restraint and road pricing problem, which aims at nding a set of effective link toll patterns to reduce link ows
to below a desirable target level. For the generalized trafc equilibrium problem formulated as variational inequalities, we propose an iterative two-stage approach with a selfadaptive step size to update the link toll pattern based on the observed link ows and given
ow restraint levels. Link travel time and demand functions and users value of time are not
needed. The convergence of the iterative toll adjustment algorithm is established theoretically and demonstrated on a set of numerical examples.
2009 Elsevier Ltd. All rights reserved.

1. Introduction
Road pricing has been recognized as an effective approach to trafc demand management and control. By charging a suitable toll, trafc demand and ow distribution on the road network can be inuenced to achieve more efcient utilization of
existing limited road capacity and generation of toll revenues for road maintenance and construction. From a theoretical perspective, road pricing has been a subject of substantial research for a few decades by transportation economists and scientists (Lindsey, 2006). The initial idea of road pricing was suggested by Pigou (1920), who used the example of a congested
road to make points about externalities and optimal congestion charges. Seminal works on both intellectual and practical
developments after Pigous idea include Walters (1961), Beckmann (1965) and Vickrey (1969). A notable application of economic theory to road network pricing is the well-known rst-best or marginal-cost pricing. According to this theory, the
additional cost or congestion externality that a road user imposes on other users can be internalized by charging tolls, thereby driving a user equilibrium (UE) ow pattern toward the system optimum (SO). In a congested network, the optimal toll to
be levied on each link is equal to the difference between the marginal social cost and the marginal private cost and it can be
determined by solving a socially optimal trafc equilibrium problem. Various extensions of marginal-cost pricing can be
made to networks with link ow interactions, multiple vehicle types, such as trucks and cars, and with ow capacity
* Corresponding author. Tel.: +852 2358 7178; fax: +852 2358 1534.
E-mail address: cehyang@ust.hk (H. Yang).
0968-090X/$ - see front matter 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trc.2009.05.009

158

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

constraints and queues (Dafermos and Sparrow, 1971; Dafermos, 1973; Smith, 1979a; Yang and Huang, 1998). In spite of its
perfect theoretical basis, the rst-best pricing scheme is of little practical interest. The second-best charging schemes are
more practically relevant and indeed have received ample attention recently. A wide variety of second-best pricing schemes
in general networks have been developed to determine the optimal tolls for system performance optimization under given
physical and economic pricing constraints. Readers may refer to Yang and Huang (2005) and Lawphongpanich et al. (2006)
for recent comprehensive treatments of the rst-best and second-best pricing problems in general networks.
A congestion pricing scheme is often introduced in practice to avoid trafc jams by eliminating queuing delays during
peak periods and to ensure that all the link ows do not exceed their physical capacities. Nevertheless, with the emerging
problem of sustainable development and transportation faced by most major cities, it is often necessary to go beyond simple
physical capacity and queuing control by achieving more strict ow restrictions on certain road sections. For instance, when
the link ow reaches its saturation level, the road network reaches the maximum acceptable risk; the excessive noise generated and the waste of fuel indicate the many serious environmental problems caused by congested roads. It is necessary to
seek a set of link tolls to maintain trafc demand within a desirable level such as the environmental capacity, which can be
regarded as the maximum trafc volume that can be accommodated along the streets without exceeding a certain threshold
of environmental damage. Indeed, for this type of trafc restraint and road pricing problem, Ferrari (1995) proposed a heuristic method to determine tolls through the modication of link cost functions. Larsson and Patriksson (1995) discussed link
capacity-constrained trafc equilibrium models using an augmented Lagrange dual scheme. Yang and Bell (1997) formulated
the problem as an elastic-demand network equilibrium model with queuing and then solved it by an inner penalty function method (Inouye, 1987). All these methods are intimately associated with the extension of the notion of the Wardropian
trafc equilibrium (Wardrop, 1952) to networks with capacity constraints, which was in fact considered in the 1960s and
1970s to improve the modeling of congestion effects (Charnes and Cooper, 1961; Jorgensen, 1963; Thompson and Payne,
1975; Daganzo, 1977a,b). Apart from deriving link tolls for trafc restraints in congested networks, capacity constraints
are also used in many trafc control and management schemes (Yang and Yagar, 1994; Smith and van Vuren, 1993) and system-optimal route guidance (Correa et al., 2004; Jahn et al., 2005).
A fundamental question in the actual design and implementation of road pricing schemes is how to choose the optimal
charge level of congestion tolls in a simple yet practical manner. Because the demand curve is usually unknown and difcult
to estimate in practice, Vickrey (1993) and Downs (1993) rst argued that congestion pricing could be determined on a trialand-error basis without demand functions, but neither proposed an actual mechanism to implement such a novel idea. Only
recently did Li (2002) propose an iterative toll adjustment procedure based on the observed link ow on a single road link,
which was soon extended to a general road network by Yang et al. (2004) for the perfect marginal-cost pricing problem. In
their developed iterative procedures, the marginal-cost-based link tolls are adjusted with changes in the observed link ow
patterns, which are in reaction to different trial tolls by individual road users. Although the proposed iterative trial-and-error
procedure does not require knowledge of demand functions, the link travel time functions and the users value of time (VOT)
are required for each trial, and they are not readily available in actual implementations.
The trial-and-error toll adjustment procedures developed by Li (2002) and Yang et al. (2004) are restricted to the perfect
marginal-cost pricing problem. One naturally wonders whether an efcient toll adjustment procedure can be developed for
the practical trafc restraint and road pricing problem mentioned above, i.e., to seek a set of link tolls to reduce trafc demand
to a desirable threshold level with unknown demand and link travel time functions. In this inquiry, we have to remark that a
straightforward way to remove trafc queues is to set a toll equivalent to the observable queuing delay at each bottleneck link,
given that the value of time of all users is identical and known. By doing so, we can obtain an non-queued equilibrium network
ow pattern that satises the physical capacity constraint of the network. The reason is that the toll charge produces no losses
for road users, as long as it is not beyond queuing delay, because it simply substitutes a charge for wasted time, between
which road users are indifferent (Yang and Bell, 1997). Nonetheless, in most trafc congestion management problems with
link toll charges, link ows need to be restrained to satisfy the environmental capacity constraints that are more severe than
the physical capacity, as mentioned earlier. There should be no queue visible at all at the stationary point of the trafc equilibrium, and hence the previous simple method of substituting queuing delay with a toll charge is not available. In this situation, Meng et al. (2005) developed a practical trial-and-error method for seeking the desirable link tolls for effective
congestion management without demand and link travel time functions. The basic method is based on the dual theory of nonlinear programming (Bazaraa et al., 1993) and the convergence result of the Lagrangian dual algorithm established by Larsson
et al. (1996). Nonetheless, the analyses and convergence proofs in Meng et al. (2005) are based on an optimization model that
requires that the travel time functions are separable and thus the ow interactions among different links are ignored.
In this study, we develop an iterative toll adjustment method for the trafc restraint and road pricing problem with unknown demand and link travel time functions. The method is developed within the framework of the generalized trafc
equilibrium problem with elastic demand and asymmetric link ow interactions. Variational inequality (VI) formulation
(Smith, 1979b; Dafermos, 1980) is adapted for the generalized elastic-demand trafc equilibrium problem with capacity
constraints. By introducing the Lagrange multipliers associated with capacity constraints, a kind of alternating direction
method can be applied for solving the resulting VI formulation with linear capacity constraints, which in turn motivates
us to develop a novel two-stage updating scheme for the iterative trial-and-error adjustment of link toll charges. Unlike
the trial-and-error method for the marginal-cost pricing problem, the whole procedure proposed here updates link tolls
based on the observed link ows after each trial and the given ow restraint levels only; link travel time and demand functions and users value of time are not needed.

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

159

This paper is organized as follows: The next section presents the VI formulation for the generalized, capacity-constrained
trafc equilibrium problem. In Section 3, we develop a two-stage iterative pricing scheme with the general VI based trafc
equilibrium model and prove its convergence. A set of numerical examples are provided in Section 4 to illustrate the potential application of the proposed method. A general discussion of the deciency and potential further extensions of the method are provided in Section 5 and conclusions are given in Section 6.
2. Trafc equilibrium problem with capacity constraints
Consider a general road network with a set of directed links denoted by A, the ow on link a is denoted by v a and
a 2 AgT is a column vector of all link ows. Each link, a 2 A, is associated with a travel time, ta , and let
t fta ; a 2 AgT . Suppose that the travel time on a link is in general dependent upon the ow on every link in the network;
that is, t a t a v . This function contains the separable link travel time function (travel time on a link as a function of the ow
on that link only) as a special case. Let W denote the set of origindestination (OD) pairs in the network and Rw the set of all
simple routes between OD pair w 2 W. The demand associated with traveling between each OD pair, w 2 W, is denoted as
dw and all the OD demands are grouped into a vector, d fdw ; w 2 WgT . We consider the general non-separable travel demand function dened in terms of the travel costs for that OD pair as well as the other OD pairs; that is, dw Dw c, where
c fcw ; w 2 WgT is a column vector of all OD travel costs and cw is the generalized travel cost between OD pair w 2 W.
The separable demand function is included as a special case.
The following standard assumptions are required in the model formulation:

v fv a ;

Assumption 1. The route choice behavior of users on the network under a given link toll pattern follows the deterministic
UE principle.
Assumption 2. All network users have a uniform VOT denoted by

s (homogeneous users).

Assumption 3. The link travel time function, tv ft a v ; a 2 AgT , is non-negative and strongly monotone with respect to
link ow vector v.
Assumption 4. The OD demand function, Dc fDw c; w 2 WgT , is non-negative and Dc is monotone with respect to
OD travel
n cost vector c;
oT furthermore, Dc is bounded and invertible. The inverse demand function is denoted by
.
D1 d D1
w d; w 2 W
Let C a be the desirable upper bound or acceptable threshold level of link ow v a ; a 2 A. As mentioned before, our


objective is to determine a set of link tolls, denoted by ua ; a 2 A , to satisfy the following environmental capacity
constraints:

v a 6 Ca ;


a 2 A:

T
Let u
a 2 A be the column vector of all these valid link tolls. After implementing such a given set of toll charges, by

Assumption 1 we have an elastic-demand network equilibrium ow pattern v  ; d , where v  fv a u ; a 2 AgT is the UE

link ow vector and d fdw u ; w 2 WgT is the OD demand vector. The following equilibrium conditions are thus satised when UE is achieved:


ua ;

8P
P ua  w
 w

>
if frw > 0;
>
< t a v dar
s dar cw ;
a2A
a2A
r 2 Rw ; w 2 W;
P
P ua  w
>
 w

w
>
: t a v dar
s dar P cw ; if fr 0;
a2A
a2A
(



D1
d

c
;
if dw > 0;
w
w
w 2 W;


1
Dw d P cw ; if dw 0;

where cw is the shortest generalized travel cost between OD pair w 2 W and frw is the ow on route r 2 Rw between OD pair
w 2 W; dw
ar is 1 if route r between OD pair w 2 W uses link a, and 0 otherwise.
Without reducing trafc demand excessively by the considered trafc restraint and road pricing scheme, we require that
the chosen link toll pattern should be meaningful in the sense that a link is free of charge if trafc ow on this link is less than
its environmental capacity; otherwise, it would be subject to a toll charge to restrain its ow actively to its prescribed restraint level. That is,

ua 0;
ua

if

P 0; if

v a < C a ;
v a C a ;

a 2 A:

Clearly, when C a ; a 2 A represent physical link capacities, the above conditions ensure that the trafc queue is eliminated
completely by the link toll pattern at links with binding capacity constraints (Hearn, 1980; Larsson and Patriksson, 1994;
Yang and Bell, 1997; Yildirim and Hearn, 2005).

160

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

Without difculty it can be shown that the equilibrium demand and link ow vector satisfying conditions (2)(4) is the
solution of the VI given in (5) below, where S is the feasible demand and the link ow set satisfying link capacity constraints
as dened by (6), and X represents the feasible demand and the link ow set dened by (7).

Find v  ; d 2 S such that


tv  T v  v   D1 d T d  d P 0;

8v ; d 2 S;

where

S fv ; d 2 Xjv a 6 C a ;

a 2 Ag;

and X represents the feasible set dened by

v ; djv a

X X

frw dw
ar ;

a 2 A;

w2W r2Rw

)
frw

dw ;

frw

P 0; r 2 Rw :

r2Rw



Note that the link toll set ua ; a 2 A in (4), after being divided by the uniform VOT, s, is implicitly given by the non-negative Lagrange multipliers associated with the capacity constraints in (6). For a comprehensive exposition of VI, readers may
consult the monographs by Nagurney (1993) and Facchinei and Pang (2003).
Alternatively, by explicitly introducing the non-negative Lagrange multiplier vector, k, to the capacity constraints set, S,
the problem in (5)(7) can be reformulated as the following equivalent VI problem:

tv  k T v  v   D1 d T d  d P 0;
 T

k  k C  v  P 0;

8v ; d; k 2 X  RjAj
;

jAj
where C fC a ; a 2 AgT is the vector of all link capacity constraints, RjAj
and jAj is the
denotes the non-negative orthant of R
cardinality of link set A. Moreover, since k u=s with s > 0, we can substitute it into (8) and obtain the following equivalent
problem:

(
 T


tv  us v  v   D1 d T d  d P 0;
u  u T C  v  P 0;

8v ; d; u 2 X  RjAj
:

The VI problem (9) can be solved by the alternating direction method that seeks the solution by replacing the original
problem with a sequence of sub-VI problems. The alternating direction method, referred to as the augmented Lagrangian
method in the literature, can be found in He and Yang (1998) and He et al. (2004).
It is well-known that, in the special case where the link travel time and OD demand functions are separable, the UE ow

pattern v  ; d in the above VI formulation can be reduced to the following constrained convex optimization problem:

min

v ;d2S

X Z va
a2A

t a xdx 

X Z
w2W

dw

D1
w xdx:

10

Various solution methods for this type of trafc equilibrium problem with link capacity constraints have been proposed
by, for example, Hearn (1980), Hearn and Lawphongpanich (1990), Larsson and Patriksson (1994, 1995, 1999) and Nie et al.
(2004). A comprehensive review of the model formulations and algorithmic developments for the trafc equilibrium problem with general side constraints can be found in Patriksson (1994).
3. The trial-and-error pricing scheme and its convergence
Given the general capacitated, non-separable trafc equilibrium model characterized by the original VI formulation, (5)
(7), or the alternative VI formulation, (8) or (9), the alternating direction method or the augmented Lagrangian methods
developed in He and Yang (1998) and He et al. (2004) work only when full network demand and supply information is available. For the implementation of the trial-and-error pricing scheme in the absence of both the link travel time and the inverse
demand functions, we have to explore a new method tailored for the general VI-based formulation to nd the desired link
toll solution in terms of the Lagrange multipliers associated with the link capacity constraints using observed link ows.
3.1. The trial-and-error pricing scheme
For the general VI model, we propose a new, two-stage, iterative trial-and-error scheme to perform toll-updating. The
scheme is rst given below:
3.1.1. The iterative trial-and-error scheme
Step 0. (Initialization) Let

o
0
ua ; a 2 A be an initial link toll pattern. Set k 0.

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

Step 1. n
(Observe link
o ow) Observe the revealed link ow denoted by
k
ua ; a 2 A .

161

o
v k
after imposition of the obtained tolls
a ; a 2 A

Step 2. (Update toll charge) Set an intermediate toll charge for each link by:

 k
 k
 
 k
u
a max ua bk v a  C a ; 0 ;

a 2 A;

and again observe the revealed link ow denoted by

newly updated toll for the next iteration is given by


 k
 

uk1
max uk
a
a ak bk v a  C a ; 0 ;

11

o
n
o
v ak ; a 2 A after levying link tolls uk
a ; a 2 A . Then, the

a 2 A:

12

Step 3. (Check convergence) If

kuk1  uk k
6 e;
kuk k

13

then stop; otherwise set k : k 1 and go to Step 1.

n
oT
k
In the above procedure, uk ua ; a 2 A and k  k denote the Euclidean norm; e is a small positive number of convergence tolerance; ak ; ak > 0, is a step size parameter and fbk g; bk > 0, is a sequence of parameters that must satisfy the following condition to guarantee the convergence of the proposed iterative scheme:

 k k;
bk kv k  v k k 6 ckuk  u

c 2 0; 1;

14

where c 2 0; 1 is a xed constant.

To execute this scheme, we can start with an appropriate initial (such as zero) link toll pattern. The whole iterative
scheme is terminated after the difference between two successive trial tolls is within a predetermined tolerance. At the convergence point, the observed ows are all below or equal to the preset levels of the environmental capacity constraint. The
basic consideration of the scheme is that, after imposition of any link toll pattern, we can observe, rather than calculate
numerically, the revealed link load pattern on the network that actually occurs under the current price. The observed link
load pattern is exactly the solution of the rst inequality in (9) with given uk , but the second inequality in (8) may not
be satised with uk . Therefore, we consider (11) and (12), a two-stage scheme for updating link toll charges based on
the observed link ows. This scheme provides us with a practicable toll-updating scheme, which only needs the revealed
equilibrium link ow pattern that can be very easily available via observations of the road in real applications. Hence, knowledge of the travel time functions, the users VOT or the demand functions is not required. It should be pointed out that the
link toll and the resulting ow patterns that satisfy the capacity constraints are not unique and the toll pattern identied by
the above trial-and-error scheme is effective for trafc restraints but it is not necessarily optimal. One may introduce a secondary objective function to select the best toll pattern among the feasible solutions based on pre-specied criteria using a
bi-level programming approach (Yang and Bell, 1997; Lawphongpanich and Hearn, 2004) but it requires the full information
on the demand and travel time function as well as users VOT.
To facilitate the proof of convergence of the newly proposed trial-and-error implementation scheme, we now elucidate
jAj
, the revealed link ow, v k , is in fact a
and summarize the scheme in a more compact vector form. Firstly, given uk 2 R
k
k
part of v ; d 2 X, which is the solution of the following VI problem:

tv k

uk

T v  v k  D1 d T d  d P 0;

8v ; d 2 X;

15

and then the updating scheme for uk1 is divided into two-stages:
(1) Prediction:

 k maxfuk bk v k  C; 0g;
u

16

(2) Correction:

uk1 maxfuk ak bk v k  C; 0g;

17

 k , again corresponds to the partial solution of the following VI

 k , the revealed link ow after imposing toll charges, u
where v
problem:

T
 k
u
k T d  d
k P 0;
tv k
v  v k  D1 d

and parameter bk satises condition (14).

8v ; d 2 X;

18

n
o
n
o
k
 k
of VI (15) and VI (18)
It is worth noticing once again that the revealed link ow solution v a ; a 2 A and v
a ; a 2 A
can be observed or measured at ease in practice using advanced trafc detectors. In other words, nding the link ow

162

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

solutions for given link toll charges does not require solving VI (15) and VI (18) numerically, and hence it does not require
exact knowledge of the link travel time and travel demand functions. Furthermore, the toll prediction (16) and correction
(17) are made from iteration to iteration based on the observed link ow and prescribed ow restraint level; the users VOTs
are not involved at all. Therefore, the implementation scheme does not require the VOT information either.
Now, we look at the choice of parameter bk , which is required to satisfy condition (14) for ensuring the convergence of the
algorithm. The following proposition indicates that such a parameter, bk , certainly exists.
Proposition 1. Suppose that the link travel cost function, tv , is strongly monotone with modulus
demand function, D1 d, is also monotone. Then,

l and the negative of the

 k T v k  v k P slkv k  v k k2 :
uk  u

19

k 2 X and v k ; dk 2 X, substituting them into VI (15) and VI (18), respectively, leads to

 k ; d
Proof. Since v

T
uk
k
k  dk P 0;
tv k
v k  v k  D1 d T d

20

T
k
u
k T dk  d
k P 0:
v k  v k  D1 d
tv k

21

s
s

Adding (20) and (21) leads to the following inequality:

k  D1 dk T d
k  dk P lkv k  v k k2 ;
 k T v k  v k P tv k  tv k T v k  v k  D1 d
uk  u

where the second inequality follows from the assumptions on the strong monotonicity of tv and the monotonicity of
D1 . Hence, Eq. (19) holds obviously for s > 0. h
Note that Eq. (19) can be further relaxed to the following inequality:

 k k P slkv k  v k k;
kuk  u

22

which shows that condition (14) is satised automatically provided that bk 6 csl, i.e., such a parameter, bk , does exist. Because it is generally difcult to know the modulus, l, and the VOT, s, may not be available as well, we suggest a self-adaptive
procedure to nd such a suitable small bk in practical implementations. For a given uk > 0 and a trial bk > 0, note that v k is
 k according to the proposed updating scheme and then calculate
automatically produced according to uk . We set the trial u

rk :

bk kv k  v k k
;
 k k
kuk  u

23

 k is likewise automatically produced according to u

 k . If r k 6 c 0 < c < 1, then the trial u
 k is accepted as the prewhere v
dictor; otherwise, reduce bk by bk : bk  c  0:9=r k and repeat the procedure until a suitable one is found. The choice of step
size ak is discussed in convergence proof.
3.2. Some properties of projection and variational inequality
Now, we summarize some important properties of projection and VI, which will be used in the subsequent convergence
analysis. Let U  Rn be a nonempty, closed convex set. For any given x 2 Rn , the projection mapping under the Euclidean
norm, denoted by P U x, is dened as follows:

PU x arg minfkx  zk jz 2 Ug:

It is easy to see that the operator, maxf; 0g, amounts to the projection operation, P R , where R represents the nonnegative orthant. A basic property of the projection mapping on a closed convex set is:

x  PU xT PU x  z P 0;

8 x 2 Rn ; z 2 U:

24

Following from (24), we readily have

kPU x  P U zk 6 kx  zk;

8 x; z 2 Rn

25

and

kPU x  zk2 6 kx  zk2  kx  PU zk2 ;

8 x 2 Rn ; 8 z 2 U:

26

For a given continuous monotone function, Fx, denoting the residue function

ex; b : x  PU x  bFx;

b > 0;

and then its zero point is equivalent to the solution of the following VI problem (Bertsekas and Tsitsiklis, 1989):

27

163

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

x  x T Fx P 0;

8x 2 U:

28

Therefore, in the problem of solving VI, kex; bk is often regarded as an error bound, which quantitatively measures how
much x fails to be in the solution set of VI (28).
3.3. The convergence proof
After introducing the above equivalence and notation, the whole trial-and-error scheme is equivalent to the following
jAj
,
sequential expressions: Firstly, given a trial, u 2 R

PX

v
d

tv us

!)

D1 d

29

 u
 , is obtained by:
 ; d;
and then a predicted interim pattern, v

 PRjAj u bv  C;
u

!)

tv us
v
v
:
 PX
  D1 d

d
d

30
31

Finally, the truly updated Lagrange multiplier vector is given by:

ua P RjAj u abv  C:

32

Here,

v k ;

k
 k ; uk1 ; ak and bk are, respectively, replaced by v ; d; u; u
 ; ua; a and b for convenience.
d ; uk ; u

 u, then v ; d; u is a solution of VI (9).

Proposition 2. If u
 u, by Eq. (30), we have
Proof. Since u

u PRjAj u bv  C:

According to the equivalence property of solving VI problems stated in (27) and (28), the proposition holds.

  uk, can be regarded as a convergence criterion of the iterative process.

This proposition indicates that the distance, ku
Note that in the trial-and-error pricing scheme presented in Section 3.1, the relative error measure of link toll charges between two trial experiments as given in (13) is used as the convergence criterion. This measure is valid because one can eas k  uk k converges to
ily show that as kuk1  uk k converges to zero, the residual keuk ; bk k in (27) or equivalently, ku
zero as well.
Theorem 1. For any y and a solution y , we have

 T gv ; u  a2 kgv ; uk2 ;
Ha P 2au  u

33

Ha : ku  u k2  kua  u k2

34

  bv  v :
gv ; u : u  u

35

where

and

The proof of Theorem 1 is given in Appendix A. This theorem is the foundation for choosing the best value of a in (32). Since
the right-hand-side of (33) is a quadratic function of a, it reaches its maximum at

a

 T gv ; u
u  u
kgv ; uk2

36

From the denition of gv ; u in (35) and inequality (14), we have

T gv ; u P 1  cku  u
 k2 :
u  u

37

 k 0, this step size, a , is always positive. Now, let g 2 0; 2 be a relaxation factor and substitute
Therefore, whenever ku  u
a ga in (33). By simple manipulation, we can get

 T gv ; u P g2  g1  ca ku  u
 k2 :
Ha P g2  ga u  u
Finally, the convergence of the proposed method is proved by the following Theorem 2.

38

164

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

Theorem 2. Let fuk g be the sequence generated by the iterative scheme proposed in Section 3.1. Then, fuk g converges to some
solution, u .
The proof of Theorem 2 is given in Appendix A.
4. Numerical experiments
In this section, we present a set of numerical experiments to demonstrate the proposed iterative trial-and-error pricing
scheme with the generalized trafc equilibrium model. For comparative analysis, we also consider the separable trafc equilibrium model presented in (10) and apply both the VI-based pricing scheme proposed in this study and the Mathematical
Programming (MP) based trial-and-error scheme developed by Meng et al. (2005). The trial-and-error scheme that applies
only to the MP-based trafc equilibrium problem (10) uses the following simple toll-updating equation:
k
uak1 maxfuk
a bk v a  C a ; 0g;

a 2 A;

39

where the step size sequence, fbk g, should satisfy the following conditions for convergence:

bk > 0; 8 k;

1
X

bk 1;

k1

1
X

b2k < 1

40

k1

and a typical step size sequence, fbk 1=kg, is used unless otherwise mentioned. The procedure is repeated until two successive trial toll patterns are close enough to each other.
Four examples are presented. The rst one is taken from Yang and Bell (1997) and the second one uses the Sioux Falls
Network. Both examples involve separable link travel time functions and are solved by both the MP- and the VI-based methods for comparison. The third and the forth examples are taken from Nagurney (1984). They involve general asymmetric,
non-separable link travel time functions and are solved by the VI-based method only. It should be mentioned here that
the link travel time and demand functions for all examples are purely used for generating the observed link ows; none
of them is needed for updating link tolls in the implementation of the trial-and-error schemes.
k
In all four examples, un-tolled link ows are used to obtain an initial set of link tolls. For given toll charge fua ; a 2 Ag, the
observed link ows at each trial are simulated by accurately solving the elastic-demand trafc assignment problem with the
link travel time functions:
k

tk
a v t a v

ua

41

where s, as dened before, is the users VOTs that are inherent in their route choice decisions under pricing, and s is assumed
k
k
to be 60 (HK$/h). Namely, given the toll pattern fua ; a 2 Ag at iteration k, the observed link ow pattern, fv a ; a 2 Ag, is
generated by solving VI (15). For parameter bk , we set initial b0 1 and c 0:95 for the self-adaptive algorithm. For both
MP- and VI-based methods, the relative error measure, kuk1  uk k=kuk k, is adopted to check the rate of convergence,
and a convergence tolerance, e 103 , is adopted for all four examples.
4.1. Numerical experiments with separable link travel time functions
Example 1. The road network, as shown in Fig. 1, consists of 11 links, 7 nodes and 4 OD pairs 1 ! 7; 2 ! 7; 3 ! 7 and
6 ! 7. The true, but unknown, demand functions are given as:

3
10

7
4

11
1

7
2

5
9

Fig. 1. The network used for Example 1.

165

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

D1!7 c1!7 600 exp0:04c1!7 ;

D2!7 c2!7 500 exp0:03c2!7 ;

D3!7 c3!7 500 exp0:05c3!7 ;

D6!7 c6!7 400 exp0:05c6!7 :

The following Bureau of Public Roads (BPR) link travel time function is used together with the link free-ow travel time,
t 0a , link physical capacity, C Pa , and link environmental capacity, C Ea , given in Table 1.

t a v a

t 0a @1

va

0:15

C Pa

!4 1
A;

a 2 A:

42

The nal results (the toll and the ratio of ow to environmental capacity for each link) obtained from the implementation
of both the MP- and the VI-based trial-and-error schemes are given in Table 2. Fig. 2 depicts the change in the selected four
link tolls with the number of iterations of the trial experiments and Fig. 3 displays the convergence of the two methods.
Example 2. This example uses a larger network, the Sioux Falls Network, for testing of both methods. The network topology,
as shown in Fig. 4, consists of 24 nodes and 76 links. Again, the standard BPR link travel time function (42) is used, with the
link input data (free-ow travel time, t0a , and physical capacity, C Pa ) given in Table 3. The environmental capacity, C Ea , is simply
set to be 0:8  C Pa for each link a 2 A. On the demand side, the following true but unknown negative exponential demand
function is assumed:


l
0
dw dw exp nw w

1
;
0

lw

w 2 W;

43
0

where nw > 0 is a cost sensitivity parameter, l0w is the minimum free-ow travel time, and dw is the maximum potential de0
mand when lw l0w . Here, a uniform value of nw 0:04 is taken for all OD pairs and the value of dw for each OD pair,
w 2 W, is presented in Table 4.
The nal results (the toll and the ratio of ow to environmental capacity for each link) obtained from the MP- and VIbased trial-and-error schemes are given in Table 5, and the rate of convergence by both methods is displayed in Fig. 5.
4.1.1. Discussion of the results
It can be seen from Table 2 for Example 1 and Table 5 for Example 2 that the ratio of ow to environmental capacity for
each link is less than or equal to one; a link is subject to a toll charge only when its ow reaches its environmental capacity,
Table 1
Input data for the network in Example 1.
Link no. a

Free-ow travel time t0a (min)

Physical capacity C Pa (veh/min)

Environmental capacity C Ea (veh/min)

1
2
3
4
5
6
7
8
9
10
11

6
5
6
7
6
1
5
10
11
11
15

200
200
200
200
100
100
150
150
200
200
200

160
150
200
150
100
100
150
100
160
160
150

Table 2
Estimated link ow ratios and link tolls for network in Example 1.
Link no. a

Method based on MP
Ratio,

1
2
3
4
5
6
7
8
9
10
11

1.00
0.34
1.00
1.00
1.00
1.00
0.27
1.00
1.00
1.00
1.00

v a =C Ea

Method based on VI
Link toll, ka

Ratio,

1.19
0.00
6.38
8.50
2.74
1.37
0.00
1.83
6.68
6.61
5.13

1.00
0.33
1.00
1.00
1.00
1.00
0.27
1.00
1.00
1.00
1.00

v a =C Ea

Link toll, ka
1.19
0.00
6.40
8.53
2.83
1.33
0.00
1.95
6.66
6.63
5.15

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

60.0

9.0

50.0

7.5

Toll on Link 4
40.0

Link Toll Charge

Link Toll Charge

166

Toll on Link 5
Toll on Link 9

30.0

Toll on Link 11

20.0
10.0

6.0
4.5
3.0

1.5

0.0

Toll on Link 4

Toll on Link 5

Toll on Link 9

Toll on Link 11

0.0
1

11

21

31

41

51

61

Number of Iterations

10

13

16

19

Number of Iterations

(a) Method based on MP

(b) Method based on VI

Fig. 2. Change in the link tolls with iteration for Example 1.

1.80

Convergence Error

1.60
1.40

Method based on MP

1.20

Method based on VI

1.00
0.80
0.60
0.40
0.20
0.00
1

13

19

25

31

37

43

49

55

61

Number of Iterations
Fig. 3. Change in the convergence error with iteration for Example 1.

otherwise it is free of charge. The objective of environmental capacity constraints is achieved perfectly through the effective
link toll charge identied from the iterative trial-and-error scheme. It is noted that the resulting ow to environmental
capacity ratios identied by the MP- and VI-based methods is almost identical for each link but there are certain differences
in the link tolls, which are largely attributed to the non-uniqueness of those effective link toll charges.
From the changes in the toll charges with iteration of trials shown in Fig. 2 for the four tolled links selected for Example 1,
we can see that it takes more than 60 iterations for the MP-based method to reach the convergence criterion and the link
tolls exhibit signicant uctuations especially in several initial trials. This is certainly not welcome in real applications as
a dramatic upward and downward adjustment of toll levels will increase public unacceptability of the trial-and-error pricing
scheme. Fortunately, in Fig. 2b, the VI-based method exhibits a gentle change tendency and quick convergence and thus is
much more preferred in practice. Corresponding to the change of individual link tolls is the aggregate rate of convergence
depicted in Fig. 3, which further conrms the superiority of the new VI-based method. In the pricing experiment on the larger Sioux Falls Network in Example 2, we found that the simple step-length sequence, f1=kg, where k is the number of iterations, does not perform well with the MP-based method. After a few trials, the sequence, f0:01=kg, is found to be more
satisfactory and adopted to produce the convergence result depicted in Fig. 5. Even with this preselection of the step-length
sequence for the MP-based method, the new VI-based method converges more sharply as shown in Fig. 5.
We can, without difculty, understand why the VI-based method performs much better than the MP-based method. The
MP-based method, the step-length sequence fbk g is predetermined independently of the specic question, which is somewhat arbitrary and thus greatly affects the rate of convergence. In contrast, with the VI-based method, the sequence, fbk g, is
not given a priori but is chosen during the process of implementation according to the self-adaptive or self-optimizing strategy described in Section 3.1, which is of course more reasonable.

167

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

1
5

4
8

6
7

35

11
9

9
25

12

36

17
8

24

26

19
7

22 47

10

28

29
51 49
30

16

38

14

15

42 71
72

74

24

22
69
66

60

59

21

61

63

65 68

75

39

56

19

45

70

13

58

46 67

23
73 76

18

50

57

44

54

52

53
37

18

17

43

41

20
55

48

32

34 40

16

21

27
11

12

13 23

10 31

33

15

14

62

20

64

Fig. 4. The Sioux Falls Network used for Example 2.

4.2. Numerical experiments with asymmetric link travel time functions

Here, we further consider two examples taken from Nagurney (1984) with asymmetric link travel time functions. In both
examples, the following form of nonlinear and asymmetric link travel time function is used:

ta v g aa v 4a

g ab v b ha ;

8 a 2 A:

44

b2A

The specic functions can be found in Nagurney (1984). The only difference here is that elastic travel demands are considered and appropriate linear inverse demand functions are assumed.
Example 3. This example uses the network shown in Fig. 6, consisting of 20 nodes, 28 links and the following eight OD
pairs: w1 1; 20, w2 1; 19, w3 2; 17, w4 4; 20, w5 6; 19, w6 2; 20, w7 2; 13 and w8 3; 14. The
asymmetric link travel time functions in Table 6 are used. The following inverse demand functions are assumed and used
only for simulating and generating observed link ows for given toll charges rather than for the actual implementation of the
trial-and-error pricing scheme.

kw1 dw1 0:05dw1 20;

kw3 dw3 0:01dw3 100;
kw5 dw5 0:10dw5 100;
kw7 dw7 0:05dw7 20;

kw2 dw2 0:06dw2 40;

kw4 dw4 0:06dw4 20;
kw6 dw6 0:10dw6 40;
kw8 dw8 0:04dw8 40:

The environmental capacity, C Ea , of each link imposed for the capacity constraint is given in Table 7. After implementing
the VI-based trial-and-error scheme, the convergent link toll pattern and the ratio of the equilibrium link ow to the link
environmental capacity are summarized in Table 8.
Example 4. This example uses the network shown in Fig. 7, consisting of 25 nodes, 37 links and the following six OD pairs:
w1 1; 20, w2 1; 25, w3 2; 20, w4 3; 25, w5 1; 24 and w6 11; 25. A complete set of asymmetric link travel
time functions is given in Table 9. Like in Example 3, the following set of true but unknown linear inverse demand functions
is used for link ow simulation.

168

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

Table 3
Input data for the Sioux Falls Network in Example 2.
Link no. a

t0a (min)

C Pa 103 veh/h)

Link no. a

t0a (min)

C Pa 103 veh/h)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

3.60
2.40
3.60
3.00
2.40
2.40
2.40
2.40
1.20
3.60
1.20
2.40
3.00
3.00
2.40
1.20
1.80
1.20
1.20
1.80
2.00
3.00
3.00
2.00
1.80
1.80
3.00
3.60
3.00
4.20
3.60
3.00
3.60
2.40
2.40
3.60
1.80
1.80

25.90
23.40
25.90
4.96
23.40
17.11
23.40
17.11
17.78
4.91
17.78
4.95
10.00
4.96
4.95
4.90
7.84
23.40
4.90
7.84
5.05
5.05
10.00
5.05
13.92
13.92
10.00
13.51
5.13
4.99
4.91
10.00
4.91
4.88
23.40
4.91
25.90
25.90

39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76

2.40
2.40
3.00
2.40
3.60
3.00
2.40
2.40
3.00
3.00
1.20
1.80
4.20
1.20
1.20
1.20
1.80
2.40
2.40
1.20
2.40
2.40
2.40
3.60
3.00
3.60
1.20
1.80
2.40
3.00
1.20
2.40
2.40
2.40
1.20
2.40
1.80
1.20

5.09
4.88
5.13
4.92
13.51
5.13
15.65
10.32
5.05
5.13
5.23
19.68
4.99
5.23
4.82
23.40
19.68
23.40
15.65
4.82
5.00
23.40
5.00
5.06
5.08
5.06
5.23
4.89
10.32
5.08
5.23
5.00
4.92
5.00
5.08
5.08
4.89
5.09

Table 4
Potential OD demands (103 veh/h) for the Sioux Falls Network in Example 2.
O/D
1
2
4
5
10
11
13
14
15
19
20
21
22
24

1
1.20
1.20
1.20
0.98
1.00
1.14
0.90
0.86
0.82
0.54
0.54
0.70
0.67

10

11

13

14

15

19

20

21

22

24

1.20

1.20
1.14

1.20
1.18
1.20

0.98
1.00
0.98
1.03

1.00
1.02
0.97
0.88
1.21

1.14
0.82
0.86
0.83
0.82
0.85

0.90
0.86
0.82
0.80
0.90
1.20
0.79

0.86
0.85
0.76
0.74
1.20
1.01
0.78
1.20

0.82
1.18
0.73
0.66
1.06
0.86
0.62
1.03
1.20

0.54
0.54
0.56
0.73
0.86
0.67
0.54
0.86
1.15
1.20

0.54
0.62
0.58
0.74
0.82
0.55
0.56
0.79
1.04
1.01
1.20

0.70
0.61
0.54
0.85
0.88
1.00
0.61
0.82
1.20
1.00
1.20
1.20

0.67
0.54
0.73
0.54
0.54
0.95
1.20
1.03
0.83
0.73
0.55
1.20
1.03

1.14
1.18
1.00
1.02
0.82
0.86
0.85
1.18
0.54
0.62
0.61
0.54

1.20
0.98
0.97
0.86
0.82
0.76
0.73
0.56
0.58
0.54
0.73

1.03
0.88
0.83
0.80
0.74
0.66
0.73
0.74
0.85
0.54

1.21
0.82
0.90
1.20
1.06
0.86
0.82
0.88
0.54

0.85
1.20
1.01
0.86
0.67
0.55
1.00
0.95

0.79
0.78
0.62
0.54
0.56
0.61
1.20

kw1 dw1 0:01dw1 10;

kw2 dw2 0:06dw2 8;

kw3 dw3 0:10dw3 20;

kw4 dw4 0:05dw4 60;

kw5 dw5 0:07dw5 80;

kw6 dw6 0:09dw6 70:

1.20
1.03
0.89
0.79
0.82
1.03

1.20
1.15
1.04
1.20
0.83

1.20
1.01
1.00
0.73

1.20
1.20
0.55

1.20
1.20

1.03

For simplicity, the environmental capacity, C Ea , is set as follows: C Ea 45 for a = 115, C Ea 55 for a = 1630 and C Ea 50 for
a = 3137. After implementing the VI-based trial-and-error scheme, the convergent link toll pattern and the ratio of the equilibrium link ow to the link environmental capacity are summarized in Table 10.

169

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

Table 5
Estimated link ow ratios and link tolls for the Sioux Falls Network in Example 2.
Link no. a

Method based on MP
Ratio,

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

0.19
0.55
0.19
1.00
0.55
0.64
0.32
0.64
0.67
1.00
0.67
0.46
1.00
1.00
0.46
1.00
0.41
0.14
1.00
0.41
0.47
0.18
1.00
0.47
0.77
0.77
0.74
0.98
0.55
0.52
0.99
0.74
0.99
1.00
0.32
0.99
0.36
0.36

v a =C Ea

Method based on VI

Toll, ka

Ratio,

0.00
0.00
0.00
2.44
0.00
0.00
0.00
0.00
0.00
0.20
0.00
0.00
0.83
2.44
0.00
2.34
0.00
0.00
2.34
0.00
0.00
0.00
0.83
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.20
0.00
0.22
4.95
0.00
0.22
0.00
0.00

0.19
0.56
0.19
1.00
0.56
0.64
0.32
0.64
0.67
1.00
0.67
0.46
1.00
1.00
0.46
1.00
0.41
0.14
1.00
0.41
0.47
0.18
1.00
0.46
0.77
0.77
0.74
0.98
0.55
0.52
1.00
0.74
1.00
1.00
0.32
1.00
0.36
0.36

v a =C Ea

Link no. a

Toll, ka

Method based on MP
Ratio,

0.00
0.00
0.00
2.44
0.00
0.00
0.00
0.00
0.00
0.13
0.00
0.00
0.77
2.44
0.00
2.28
0.00
0.00
2.29
0.00
0.00
0.00
0.78
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.15
0.00
0.17
4.95
0.00
0.16
0.00
0.00

39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76

v a =C Ea

1.00
1.00
0.79
0.89
0.98
0.79
0.35
1.00
0.18
0.55
0.43
0.08
0.52
0.43
1.00
0.14
0.08
0.20
0.35
1.00
0.95
0.20
0.95
0.98
0.49
0.98
0.99
1.00
1.00
0.48
1.00
0.88
0.90
0.88
1.00
1.00
1.00
1.00

Method based on VI

Toll, ka

Ratio,

5.75
4.95
0.00
0.00
0.00
0.00
0.00
0.47
0.00
0.00
0.00
0.00
0.00
0.00
0.71
0.00
0.00
0.00
0.00
0.71
0.00
0.00
0.00
0.00
0.00
0.00
1.26
0.57
0.49
0.00
1.28
0.00
0.00
0.00
1.27
5.75
0.59
1.27

1.00
1.00
0.79
0.89
0.98
0.79
0.35
1.00
0.18
0.54
0.43
0.08
0.52
0.42
1.00
0.14
0.08
0.20
0.35
0.99
0.95
0.20
0.95
0.98
0.49
0.98
1.00
1.00
1.00
0.48
1.00
0.88
0.90
0.88
1.00
1.00
1.00
1.00

v a =C Ea

Toll, ka
5.71
4.95
0.00
0.00
0.00
0.00
0.00
0.49
0.00
0.00
0.00
0.00
0.00
0.00
0.71
0.00
0.00
0.00
0.00
0.72
0.00
0.00
0.00
0.00
0.00
0.00
1.26
0.58
0.49
0.00
1.26
0.00
0.00
0.00
1.27
5.71
0.58
1.26

4.2.1. Discussion of the results

Fig. 8 shows the changes in link toll charges with the number of iterations of the trial experiments for the two examples.
Again, for sake of clarity, only four tolled links for each example are selected. Like the symmetric cases in Examples 1 and 2, a

1.20

Convergence Error

1.00

0.80
Method based on MP
0.60

Method based on VI

0.40

0.20

0.00
1

10

19

28

37

46

55

64

73

82

Number of Iterations
Fig. 5. Change in the convergence error with iteration for Example 2.

170

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

Fig. 6. The network (source: Nagurney, 1984) used for Example 3.

Table 6
Asymmetric link travel time functions used in Example 3.
t1 v 5  107 v 41 0:05v 1 0:02v 2 5
t3 v 5  107 v 43 0:03v 3 0:01v 4 3:5
t5 v 6  107 v 45 0:06v 5 0:04v 6 6
t7 v 8  107 v 47 0:08v 7 0:02v 8 4
t9 v 107 v 49 0:06v 9 0:02v 10 7
t11 v 7  107 v 411 0:07v 11 0:04v 12 6:5
t13 v 107 v 413 0:07v 13 0:03v 18 6
t15 v 3  107 v 415 0:09v 15 0:02v 14 2
t17 v 3  107 v 417 0:07v 17 0:02v 15 4:5
t19 v 0:08v 19 0:03v 17 6
t21 v 4  107 v 421 0:04v 21 0:01v 22 4
t23 v 3  107 v 423 0:09v 23 0:02v 24 3:5
t25 v 3  107 v 425 0:09v 25 0:03v 26 4:5
t27 v 3  107 v 427 0:08v 27 0:03v 28 5

t2 v 3  107 v 42 0:04v 2 0:04v 1 2

t4 v 3  107 v 44 0:06v 4 0:03v 5 4
t6 v 0:07v 6 0:03v 7 5
t8 v 4  107 v 48 0:05v 8 0:02v 9 6:5
t10 v 0:04v 10 0:01v 12 8
t12 v 0:08v 12 0:02v 13 7
t14 v 0:08v 14 0:03v 15 5
t16 v 0:08v 16 0:05v 12 3
t18 v 0:05v 18 0:01v 16 3
t20 v 2  107 v 420 0:06v 20 0:01v 21 3
t22 v 2  107 v 422 0:06v 22 0:01v 23 5
t24 v 2  107 v 424 0:08v 24 0:01v 25 4
t26 v 6  107 v 426 0:07v 26 0:08v 27 3
t28 v 3  107 v 428 0:07v 28 6:5

Table 7
Environmental capacity for capacity constraint for the network in Example 3.
Link no.
a

Environmental capacity,
C Ea

Link no.
a

Environmental capacity,
C Ea

Link no.
a

Environmental capacity,
C Ea

Link no.
a

Environmental capacity,
C Ea

1
2
3
4
5
6
7

60
70
75
60
60
150
70

8
9
10
11
12
13
14

60
60
60
60
60
60
60

15
16
17
18
19
20
21

60
80
70
70
60
70
70

22
23
24
25
26
27
28

80
60
60
75
80
75
60

Table 8
Estimated link ow ratios and link tolls for the network in Example 3.
Link no. a

Ratio,

1
2
3
4
5
6
7

0.00
0.85
1.00
0.89
0.54
1.00
1.00

v a =C Ea

Toll, ka

Link no. a

Ratio,

0.00
0.00
5.96
0.00
0.00
5.57
3.08

8
9
10
11
12
13
14

1.00
0.00
0.00
0.63
0.83
0.36
0.35

v a =C Ea

Toll, ka

Link no. a

Ratio,

8.77
0.00
0.00
0.00
0.00
0.00
0.00

15
16
17
18
19
20
21

0.29
1.00
0.14
0.86
0.00
0.00
0.54

v a =C Ea

Toll, ka

Link no. a

Ratio,

0.00
2.49
0.00
0.00
0.00
0.00
0.00

22
23
24
25
26
27
28

1.00
0.61
0.96
1.00
0.81
1.00
0.00

v a =C Ea

Toll, ka
7.63
0.00
0.00
10.82
0.00
4.20
0.00

smooth change in link tolls is found with the proposed iterative trial-and-error pricing scheme. In addition, Fig. 9 shows the
nice convergence of the aggregate convergence error measures for both examples. The VI-based method performs well with
the asymmetric, non-separable link travel time function.
5. Outstanding political and behavioral issues
To ensure that the basic theory and implementation method are valid, we have made some simplistic and strong assumptions about the transportation system, such as static networks and OD demands, deterministic travel behavior and observations. These assumptions cannot be guaranteed in practice, particularly under the proposed sequential toll adjustment
scheme. This section is devoted to a thorough discussion of how these assumptions could be relaxed and the implementation
procedure enhanced.

171

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

Fig. 7. The network (source: Nagurney, 1984) used for Example 4.

Table 9
Asymmetric link travel time functions used in Example 4.
t1 v 5  105 v 41 5v 1 2v 2 500
t2 v 3  105 v 42 4v 2 4v 1 200
t3 v 5  105 v 41 3v 3 v 4 350
t4 v 3  105 v 44 6v 4 3v 5 400
t5 v 6  105 v 45 6v 5 4v 6 600
t6 v 7v 6 3v 7 500
t7 v 8  105 v 47 8v 7 2v 8 400
t8 v 4  105 v 41 5v 8 2v 9 650
t9 v 105 v 49 6v 9 2v 10 700
t10 v 4v 10 v 12 800
t11 v 7  105 v 411 7v 11 4v 12 650
t12 v 8v 12 2v 13 700
t13 v 105 v 413 7v 13 3v 18 600
t14 v 8v 14 3v 15 500
t15 v 3  105 v 415 9v 15 2v 14 200
t16 v 8v 16 5v 12 300
t17 v 3  105 v 417 7v 17 2v 15 450
t18 v 5v 18 v 16 300
t19 v 8v 19 3v 17 600

t20 v 3  105 v 420 6v 20 v 21 300

t21 v 4  105 v 421 4v 21 v 22 400
t22 v 2  105 v 422 6v 22 v 23 500
t23 v 3  105 v 423 9v 23 2v 24 350
t24 v 2  105 v 424 8v 24 v 25 400
t25 v 3  105 v 425 9v 25 3v 26 450
t26 v 6  105 v 426 7v 26 8v 27 300
t27 v 3  105 v 427 8v 27 3v 28 500
t28 v 3  105 v 428 7v 28 650
t29 v 3  105 v 429 3v 29 v 30 450
t30 v 4  105 v 430 7v 30 2v 31 600
t31 v 3  105 v 431 8v 31 v 32 750
t32 v 6  105 v 432 8v 32 3v 33 650
t33 v 4  105 v 433 9v 33 2v 31 750
t34 v 6  105 v 434 7v 34 3v 30 550
t35 v 3  105 v 435 8v 35 3v 32 600
t36 v 2  105 v 436 8v 36 4v 31 750
t37 v 6  105 v 437 5v 37 v 36 350

5.1. Short-run and long-run demand changes

The trial-and-error procedure to update link tolls for congestion control was built upon the assumption that changes in
equilibrium demand and link ow patterns take place by short-run changes in the price along the same demand curve. The
ow revelation mechanism thus requires a series of price changes to derive the desirable link tolls properly. Politically speaking, it is clear that a series of quick changes to a toll system are an unlikely social event, as it would make the newly introduced system unsustainable. A plausible toll adjustment by a willing authority would be at annual, semi-annual or at most
quarterly intervals, like the situation in Singapore where toll rate adjustments are made quarterly and more frequently for
holidays. Nevertheless, such a socially acceptable long inter-trial period could create a behavioral dilemma for the analysis
envisioned here. On the one hand, it would allow travelers to have sufcient time to learn and adjust their trip-making decisions in a new pricing environment, although this might not be a critical issue, as advanced information technology is nowadays available and travelers can be quickly informed of altered toll charges. On the other hand, the background variables,

Table 10
Estimated link ow ratios and link tolls for the network in Example 4.
Link no. a

Ratio,

1
2
3
4
5
6
7
8
9
10

1.00
1.00
1.00
0.57
0.00
0.00
0.00
0.00
0.00
1.00

v a =C Ea

Toll, ka

Link no. a

Ratio,

1.93
0.00
13.13
0.00
0.00
0.00
0.00
0.00
0.00
1.42

11
12
13
14
15
16
17
18
19
20

0.35
0.22
0.43
0.57
0.00
0.00
0.00
0.00
0.00
0.86

v a =C Ea

Toll, ka

Link no. a

Ratio,

0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00

21
22
23
24
25
26
27
28
29
30

0.82
1.00
0.45
0.00
0.00
0.00
0.00
0.00
0.59
0.04

v a =C Ea

Toll, ka

Link no. a

Ratio,

0.00
9.12
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00

31
32
33
34
35
36
37

0.31
1.00
1.00
0.64
0.69
1.00
0.20

v a =C Ea

Toll, ka
0.00
11.31
10.94
0.00
0.00
24.07
0.00

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

12

30.0

10

25.0

Link Toll Charge

Link Toll Charge

172

8
6
4

Toll on Link 6

Toll on Link 8

Toll on Link 22

Toll on Link 25

20.0
15.0
10.0
5.0

Toll on Link 3

Toll on Link 22

Toll on Link 32

Toll on Link 36

0.0

10

13

16

Number of Iterations

Number of Iterations

(a) For Example 3

(b) For Example 4

10

Fig. 8. Change in the link toll with iteration with asymmetric link travel time functions for Examples 3 and 4.

0.30

Convergence Error

0.25
0.20
Example 3
0.15

Example 4

0.10
0.05
0.00
1

11

13

15

Number of Iterations
Fig. 9. Change in the convergence error with iteration for Examples 3 and 4.

such as ination and real income growth, change over such a length of time. Thus, it is imperative to address the issues of the
short-run and long-run differences in behaviors as well as the time lag of user responses in formulating a more practical pricing mechanism. In this respect, the current method makes efcient utilization of the trafc counts that are relatively inexpensive to collect with automatic devices, even on a daily basis with good accuracy. These readily and frequently available
trafc counts effectively capture the long-run demand shift prior to the next price trial. It is thus interesting to see how the
proposed tolling scheme would work under such a changing environment and what modications should be made to make
the procedure more adaptive to long-run demand shifts.
5.2. Plausible toll development trajectory
The proposed tolling scheme aims at minimizing the convergence process towards the nal solution, and the sequential
experimental process involves both upward and downward adjustment of the starting and intermediate toll levels. Such a
process should be designed such that it is politically acceptable in the rst place. To make the experimental procedure practically feasible, it is thus necessary to optimally consider some plausible bounds for the trajectory of realistic toll evolution
over time, including the starting toll levels, the trajectory towards the nal optimum, and the number and size of intervals in
between.
5.3. Incorporation of stochastic variations and observations
To generate useful, valid, unbiased and converging estimates of the toll levels, it is critical to extend the deterministic
behavioral and optimization models to stochastic conditions. First, choice behavior can be treated through a stochastic equilibration approach instead of through the deterministic route choice model considered so far. Second, instead of assuming a
single or deterministic observation value for link ows, which is inconsistent with the proposed experimental approach that

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

173

is intended to have an update interval of multiple months, we can incorporate repeated ow observations that are always
correlated stochastic variates, the within-day stochastic route choices and their day-to-day evolution process in a new pricing experiment. Thus, it is meaningful to look into a stochastic trial-and-error tolling scheme to handle explicitly the random
variations that are immanent in the observations.
5.4. Multiple vehicle types and heterogeneous users
The proposed tolling scheme allows for non-separable, asymmetric link cost functions. It should be relatively easy to take
into account multiple types of vehicles on the network, such as heavy versus light vehicles, although the shared link capacity
constraint could cause difculty. In this case, we should recognize the fact that different types of vehicles make different
contributions to trafc congestion and road capacity consumption and they should be subjected to different toll charges
according to their observationally distinguishable characteristics. In addition, the tolling scheme makes sequential toll
adjustment without requiring the knowledge of the users VOTs. This is true under the current assumption of homogeneous
users. Nonetheless, with heterogeneous users having different VOTs, it is unclear whether the proposed tolling scheme
would continue to work. In this case, the tolling scheme must be anonymous in the sense that the same amount of toll is
levied on each link for all user classes, because users differ from one another in VOT only, which is observationally indistinguishable, and thus toll differentiation across user classes is unrealistic in reality.
6. Conclusions
Recognizing the widespread practical difculty in determining appropriate toll charges for mitigating trafc congestion
and reducing environmental damage, we developed an attractive and convergent trial-and-error implementation scheme for
emerging trafc restraint and road pricing problems. Distinguished from the traditional modeling approach, the proposed
scheme allows trafc planners to estimate link tolls easily from observed link ows only, without resorting to both link travel time and demand functions and users VOTs. As demonstrated through extensive numerical experiments, the proposed
method applies to the generalized trafc equilibrium problem with asymmetric link ow interactions, and thus offers a powerful practical method for the network toll design problem for effective congestion management.
In spite of the appealing features of the proposed tolling scheme, by no means is the problem well solved. A number of
concerns for practical implementation have been identied and have to be addressed in future research.
Acknowledgements
The research described in this paper was substantially supported by a grant from the Research Grants Council of the Hong
Kong Special Administrative Region, China (Project No. HKUST6215/06E). The second author was partially supported by the
NSF of China (Project Nos. 70571033 and 70831002) and the third author was partially supported by the Cultivation Fund of
the Key Scientic and Technical Innovation Project, Ministry of Education of China (No. 708044). The authors wish to express
their thanks to four anonymous reviewers for their very constructive comments on an earlier version of the paper.
Appendix A. Proofs of Theorems
  C and z : u in (26), we get
Proof of Theorem 1. First setting x : u abv

kua  u k2 6 ku abv  C  u k2  ku abv  C  uak2 :

Substitute this into (34) leads to

Ha P ku  uak2  2abu  u T v  C  2abua  uT v  C:

45

In view of the similarly to the proof of Eq. (19), we have

  u T v   v P 0;
u
which implies

  u T bC  v P u
  u T bC  v  P 0;
u
and, consequently, it yields

 T bC  v :
u  u T bC  v P u  u

46

Substituting (46) into (45), we obtain

 T v  C
Ha P ku  uak2  2abua  u

47

174

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

Now, we observe the second term in the right-hand-side of the inequality above. By setting x : u bv  C and z : ua
in (24), we have:

 T u bv  C  u
  6 0;
ua  u
that is,

 T v  C 6 u
  uaT gv ; u:
bua  u
From (47), we thus arrive at:

 T gv ; u ku  ua  agv ; uk2 2au  u

T gv ; u  a2 kgv ; uk2
Ha P ku  uak2 2aua  u
 T gv ; u  a2 kgv ; uk2 :
P 2au  u
The theorem thus holds.

Proof of Theorem 2. From (38), we know that

 k k;
kuk1  u k2 6 kuk  u k2  c0 kuk  u

48

where c0 g2  g1  ca . This means that the sequence fu g is Fejer monotone with respect to the solution set. Therefore, fuk g is bounded and


 k k 0:
lim kuk  u

49

k!1

 k g is also bounded. Since we can always nd a b > 0 such that fbk g  b; 1), it then follows
Consequently, the sequence fu
from Lemma 2.2 in He and Liao (2002) that

 k ; bk 6 keu
k ; bk k ku
 k  P RjAjjWj u
 k bk v k  Ck
keu

 k bk v k  Ck 6 kuk  u
 k bk v k  v k k
kPRjAjjWj yk bk v k  C  PRjAjjWj u

 k k
6 1 ckuk  u

50

and thus

 k ; b 0:
lim eu

51

k!1

 k g and the sub-sequence fu

 kj g converges to u . Since eu; b is a continuous function of u, it
Let u be a cluster point of fu
follows from (51) that

 kj ; b 0:
eu ; b lim eu

52

j!1

Therefore, u is a solution point. Note that inequality (48) is true for any solution point. We thus have

kuk1  u k2 6 kuk  u k2 ;
 kj

As fu

g ! u and ku

 kl

ku

 2

8 k P 0:

 k

 u k ! 0, for any given

 u k < e=2 and ku

kl

kl

u

53

e > 0, there is an l > 0 such that

k < e=2:

54

Thus, for any k P kl , it follows from (53) and (54) that

kl k ku
 kl  u k < e
kuk  u k 6 kukl  u k 6 kukl  u
and thereby the sequence fuk g converges to u . h
References
Bazaraa, M.S., Sherali, H.D., Shetty, C.M., 1993. Nonlinear Programming: Theory and Algorithms, 2nd ed. Wiley, New York, NY.
Beckmann, M.J., 1965. On optimal tolls for highways, tunnels and bridges. In: Edie, L., Herman, R., Rothery, R. (Eds.), Vehicular Trafc Science. Elsevier, New
York, pp. 331341.
Bertsekas, D.P., Tsitsiklis, J.N., 1989. Parallel and Distributed Computation: Numerical Methods. Prentice-Hall, Englewood Cliffs, NJ.
Charnes, A., Cooper, W.W., 1961. Multicopy trafc network models. In: Herman, R. (Ed.), Theory of Trafc Flow. Elsevier, Amsterdam, pp. 8596.
Correa, J.R., Schulz, A.S., Stier-Moses, N.E., 2004. Selsh routing in capacitated networks. Mathematics of Operations Research 29, 961976.
Daganzo, C.F., 1977a. On the trafc assignment problem with ow dependent costsI. Transportation Research 11, 433437.
Daganzo, C.F., 1977b. On the trafc assignment problem with ow dependent costsII. Transportation Research 11, 439441.
Dafermos, S.C., 1980. Trafc equilibrium and variational ineqalities. Transportation Science 14, 4254.
Dafermos, S.C., 1973. Toll patterns for multiclass-user transportation networks. Transportation Sciences 7, 211223.
Dafermos, S.C., Sparrow, F.T., 1971. Optimal resource allocation and toll patterns in user-optimized transport network. Journal of Transport Economics and
Policy 5, 198200.
Downs, A., 1993. Point of view: implementing peak-hour road pricing at fullscale: nding solution to practical problems. TR News 167, 79.

H. Yang et al. / Transportation Research Part C 18 (2010) 157175

175

Facchinei, F., Pang, J.S., 2003. Finite-dimensional Variational Inequalities and Complementarity Problems. Springer-Verlag, New York.
Ferrari, P., 1995. Road pricing and network equilibrium. Transportation Research 29B, 357372.
He, B.S., Liao, L.Z., 2002. Improvements of some projection methods for monotone nonlinear varaitioanl inequalities. Journal of Optimization Theory and
Applications 112, 111128.
He, B.S., Yang, H., 1998. Some convergence properties of a method of multipliers for linearly constrained monotone variational inequalities. Operations
Research Letters 23, 151161.
He, B.S., Yang, H., Zhang, C.S., 2004. A modied augmented Lagrangian method for a class of monotone variational inequalities. European Journal of
Operational Research 59, 3551.
Hearn, D.W., 1980. Bounding Flows in Trafc Assignment Models. Technical Report 80-4, Department of Industrial and Systems Engineering, University of
Florida, Gainesville, FL.
Hearn, D.W., Lawphongpanich, S., 1990. A dual ascent algorithm for trafc assignment problems. Transportation Research 24B, 423430.
Inouye, H., 1987. Trafc equilibria and its solution in congested road networks. In: Genser, R. (Ed.), Proceedings of 5th IFAC Conference on Control in
Transportation Systems, Vienna, pp. 267272.
Jahn, O., Mhring, R.H., Schulz, A.S., Stier-Moses, N.E., 2005. System-optimal routing of trafc ows with user constraints in networks with congestion.
Operations Research 53, 600616.
Jorgensen, N.O., 1963. Some Aspects of the Urban Trafc Assignment Problem. Masters Thesis, Institute of Transportation and Trafc Engineering,
University of California at Berkeley, Berkeley, CA.
Larsson, T., Patriksson, M., 1999. Side constrained trafc equilibrium models analysis, computation and applications. Transportation Research 33B, 233
264.
Larsson, T., Patriksson, M., 1995. An augmented Lagrangean dual algorithm for link capacity constrained trafc assignment problems. Transportation
Research 29B, 433455.
Larsson, T., Patriksson, M., 1994. Equilibrium characterizations of solutions to side constrained asymmetric trafc assignment models. Le Matematiche 49,
249280.
Larsson, T., Patriksson, M., Stromberg, A.B., 1996. Conditional subgradient theory and application. European Journal of Operations Research 88, 382403.
Lawphongpanich, S., Hearn, D.W., 2004. An MPEC approach to second-best toll pricing. Mathematical Programming 101, 3355.
Lawphongpanich, S., Hearn, D.W., Smith, M.J. (Eds.), 2006. Mathematical and Computational Models for Congestion Charging. Springer.
Li, M.Z.F., 2002. The role of speedow relationship in congestion pricing implementation with an application to Singapore. Transportation Research 36B,
731754.
Lindsey, R., 2006. Do economists reach a conclusion on road pricing? The intellectual history of an idea. Econ Journal Watch 3, 292379.
Meng, Q., Xu, W., Yang, H., 2005. A trial-and-error procedure for implementing a road-pricing scheme. Transportation Research Record 1923, 103109.
Nagurney, A., 1993. Network Economics: A Variational Inequality Approach. Kluwer Academics Publishers, Dordrecht.
Nagurney, A., 1984. Comparative test of multimodal trafc equilibrium methods. Transportation Research 18B, 469485.
Nie, Y., Zhang, M., Lee, D.H., 2004. Models and algorithm for the trafc assignment problem with link capacity constraints. Transportation Research 38B,
285312.
Patriksson, M., 1994. The Trafc Assignment Problem. The Netherlands, VSP, Utrecht.
Pigou, A.C., 1920. The Economics of Welfare. MacMillan, London.
Smith, M.J., 1979a. The marginal cost taxation of a transportation network. Transportation Research 13B, 237242.
Smith, M.J., 1979b. The existence, uniqueness and stability of trafc equilibrium. Transportation Research 13B, 295304.
Smith, M.J., van Vuren, T., 1993. Trafc equilibrium with responsive trafc control. Transportation Science 27, 118132.
Thompson, W.A., Payne, H.J., 1975. Trafc Assignment on Transportation Networks With Capacity Constraints and Queuing. Paper Presented at the 47th
National ORSA/TIMS North American Meeting, Chicago, IL.
Vickrey, W., 1993. Point of view: principles and applications of congestion pricing. TR News 167, 45.
Vickrey, W., 1969. Congestion theory and transport investment. American Economic Review (Papers and Proceedings) 59, 251261.
Walters, A.A., 1961. The theory and measurement of private and social cost of highway congestion. Econometrica 29, 676699.
Wardrop, J.G., 1952. Some theoretical aspects of road trafc research. Proceedings of Institution of Civil Engineers Part II 1, 325378.
Yang, H., Bell, M.G.H., 1997. Trafc restraint, road pricing and network equilibrium. Transportation Research 31B, 303314.
Yang, H., Huang, H.J., 2005. Mathematical and Economic Theory of Road Pricing. Elsevier, Oxford.
Yang, H., Huang, H.J., 1998. Principle of marginal-cost pricing: How does it work in a general network? Transportation Research 32A, 4554.
Yang, H., Meng, Q., Lee, D.H., 2004. Trial-and-error implementation of marginal-cost pricing on networks in the absence of demand functions.
Transportation Research 38B, 477493.
Yang, H., Yagar, S., 1994. Trafc assignment and trafc control in general freeway-arterial corridor systems. Transportation Research 28B, 463486.
Yildirim, M.B., Hearn, D.W., 2005. A rst best toll pricing framework for variable demand trafc assignment problems. Transportation Research 39B, 659
678.