Vous êtes sur la page 1sur 27

Transportation Research Part B 45 (2011) 10351061

Contents lists available at ScienceDirect

Transportation Research Part B


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

Dynamic user equilibrium with side constraints for a trafc network:


Theoretical development and numerical solution algorithm
R.X. Zhong a, A. Sumalee a,, T.L. Friesz b, William H.K. Lam a
a
Department of Civil and Structural Engineering, The Hong Kong Polytechnic University, Hong Kong SAR, China
b
Department of Industrial and Manufacturing Engineering, Pennsylvania State University, USA

a r t i c l e i n f o a b s t r a c t

Article history: This paper investigates a trafc volume control scheme for a dynamic trafc network
Received 27 March 2009 model which aims to ensure that trafc volumes on specied links do not exceed preferred
Received in revised form 29 April 2011 levels. The problem is formulated as a dynamic user equilibrium problem with side con-
Accepted 6 May 2011
straints (DUE-SC) in which the side constraints represent the restrictions on the trafc vol-
umes. Travelers choose their departure times and routes to minimize their generalized
travel costs, which include early/late arrival penalties. An innite-dimensional variational
Keywords:
inequality (VI) is formulated to model the DUE-SC. Based on this VI formulation, we estab-
Dynamic user equilibrium (DUE)
Trafc volume control
lish an existence result for the DUE-SC by showing that the VI admits at least one solution.
Trafc management and control To analyze the necessary condition for the DUE-SC, we restate the VI as an equivalent opti-
Trafc assignment with side constraint mal control problem. The Lagrange multipliers associated with the side constraints as
Pontryagin minimum principle derived from the optimality condition of the DUE-SC provide the trafc volume control
Lagrange multiplier scheme. The control scheme can be interpreted as additional travel delays (either tolls or
access delays) imposed upon drivers for using the controlled links. This additional delay
term derived from the Lagrange multiplier is compared with its counterpart in a static user
equilibrium assignment model. If the side constraint is chosen as the storage capacity of a
link, the additional delay can be viewed as the effort needed to prevent the link from spill-
back. Under this circumstance, it is found that the ow is incompressible when the link
trafc volume is equal to its storage capacity. An algorithm based on Eulers discretization
scheme and nonlinear programming is proposed to solve the DUE-SC. Numerical examples
are presented to illustrate the mechanism of the proposed trafc volume control scheme.
 2011 Elsevier Ltd. All rights reserved.

1. Introduction

Various forms of trafc control scheme have been introduced to maintain or control the level of service of a trafc net-
work, such as trafc signals, ramp metering, and trafc volume (or queue) control. These schemes mainly mitigate conges-
tion or maximize capacity. However, there is a need in some cases to ensure the safety or desired environmental condition of
an urban area or road section that may require a restriction on trafc volume. This paper focuses on that type of trafc vol-
ume control problem, wherein one aims to derive an appropriate control scheme from a dynamic trafc network model.
Similar studies have been conducted in the context of side constrained static trafc assignment problem (SC-TAP) or capac-
itated static trafc assignment problem (Larsson and Patriksson, 1995, 1999; Larsson et al., 2004; Yang and Huang, 2005).
Typically, the side constraints impose restrictions on trafc volumes or maximum delays of certain links. There are two types
of side constraint in trafc assignment models: prescriptive (hard) and descriptive (soft) side constraints. Prescriptive side

Corresponding author. Tel.: +852 3400 3963; fax: +852 2334 6389.
E-mail addresses: cezhong@gmail.com (R.X. Zhong), ceasumal@polyu.edu.hk (A. Sumalee), tfriesz@psu.edu (T.L. Friesz), cehklam@polyu.edu.hk
(W.H.K. Lam).

0191-2615/$ - see front matter  2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trb.2011.05.004
1036 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

constraints typically arise from trafc management and control policies (e.g. link capacity constraint, trafc signal, access con-
trol, and trafc volume control, etc.). With this type of constraint, the SC-TAP can be used to model and evaluate necessary
optimal control parameters (implemented as access delay times or tolls). The validity of such controls depends on the
assumption that their effects are transferable to travelers perceptions as additional disutilities of travel. Descriptive side con-
straints, in contrast, can be introduced to better represent the physical restriction of highways or junctions (e.g. joint capac-
ities of roundabouts), which may not be fully represented by the standard ow-delay relationship (e.g. the BPR function). In
this paper, we concentrate on the prescriptive side constraints for the purposes of access control and trafc volume control.
Under the SC-TAP, Larsson and Patriksson (1995, 1999) show that the Lagrange multiplier as derived from the optimality
condition of the SC-TAP can be used as the control parameter (equivalent delay or adjusted travel cost function) on that link.
However, the basic assumption of the static model is that the trafc system operates at its steady state with constant
demand and link trafc volume. This may not be suitable or plausible for the analysis of urban trafc management, which
involves temporal ow and congestion. We thus aim to extend the SC-TAP to the dynamic case and derive the dynamic con-
trol parameter similar to the Lagrange multiplier of the SC-TAP, which can be adopted in the dynamic trafc volume control
scheme.
The dynamic user equilibrium (DUE) condition is a type of route/departure time choice principle for dynamic trafc assign-
ment (DTA) problems. Under the DUE, each traveler chooses a route and/or departure time to minimize his generalized travel
time/cost (including early/late arrival penalties) (Carey, 2008). Several approaches have been proposed to formulate and solve
the DTA problem, e.g. Friesz et al. (1993, 2001, 2008, 2011), Huang and Lam (2002), Wie et al. (2002), Lindsey (2004), Friesz and
Mookherjee (2006), Carey (2008), Chow (2009), and Friesz (2010). The dynamical framework allows a representation of tem-
poral travel demand and trafc condition in which a time-varying trafc control scheme can also be derived from the model.
The need for the dynamic trafc volume control scheme can be found in several practical cases. For example, in Hong
Kong there are several cross harbor tunnels, with the central tunnel the most congested due to its lower toll level. It is thus
necessary to control the number of vehicles inside the tunnel to maintain sufcient reserve capacity/space for handling any
possible incident (e.g. car accident or disruption due to disaster). We provide a sound theoretical analysis of the optimal de-
sign of dynamic trafc volume control scheme by exposing the DUE-SC condition and the additional travel cost induced by
side constraints. The introduction of the DUE-SC also provides a potential approach to determining the DUE condition under
a pre-specied level of service (LOS) for a trafc network (Yang et al., 2000).
In this paper, the DUE-SC is formulated as an innite-dimensional variational inequality following Friesz et al. (1993);
Friesz (2010). We adopt the whole-link linear travel time model as proposed by Friesz et al. (1993) that guarantees the
rst-in-rst-out (FIFO) discipline of dynamic ow when appropriate arc exit/entrance ow rates are positive. The existence
of equilibrium to the DUE/DUE-SC is veried by showing the VIP admits at least one solution. The Pontryagin minimum prin-
ciple is then applied to analyze the necessary condition of the DUE-SC for a general trafc network by restating the VI as an
equivalent optimal control problem. We then interpret the physical meaning of the Lagrange multipliers, as in the static case.
Based on the VI formulation, a xed point algorithm is developed for the DUE-SC. An algorithm based on Eulers discretiza-
tion scheme and nonlinear programming is then proposed to solve the DUE-SC. Numerical tests are conducted to illustrate
the applicability and mechanism of the proposed trafc volume control scheme.
The remainder of this paper is organized as follows. In Section 2, the SC-TAP is briey reviewed. The DUE-SC is formulated
and analyzed in Section 3. The solution algorithm for solving the DUE-SC problem is proposed in Section 4. Numerical exam-
ples are then provided to demonstrate the proposed method in Section 5. Conclusions and directions for future work are
highlighted in Section 6. Companion materials are given in the Appendix.

2. Static side-constrained trafc assignment problem

Before addressing the DUE-SC, it is necessary to review the formulation and properties of the SC-TAP. An urban road
transportation system can be described as a strongly connected directed network G(N, A) where N and A denote the sets
of nodes and links, respectively. We assume that the network has n links in total. Let Ca be the capacity of link a 2 A. Let
W be the set of OD pairs and Qw be the number of trips made between OD pair w 2 W. A path is dened as a connected se-
quence of links, Pw denotes the set of all non-cyclic paths connecting OD pair w 2 W and P denotes the set of all non-cyclic
paths. fp denotes the ow on path p 2 P and ta represents the trafc volume on link a 2 A. We presume that the SC-TAP has at
least one feasible solution, which means that link capacities are sufciently large to allow all travel demands to traverse the
network for at least one assignment. The SC-TAP (Larsson and Patriksson, 1995) is formulated as
XZ ta
min Zt t a xdx; 1
a2A 0
X
s:t: fp Q w ; w 2 W; 2
p2P w
X
fp dpa ta ; a 2 A; 3
p2P

ta 6 C a ; a 2 A; 4
fp P 0; p 2 P; 5
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1037

where t = (ta:"a 2 A), (4) denes the capacity constraints, ta(ta) is the link travel time on link a 2 A when the trafc volume
on link a is ta, and

1; if a 2 p;
dpa
0; otherwise;
is the Kronecker Delta function. We have the following Lemma for the SC-TAP.

Lemma 2.1. (Larsson and Patriksson, 1995) Let lw and ka be the Lagrange multipliers associated with constraints (2) and (4),
respectively. The rst-order optimality condition for UE with side constraints is given as follows:
fp cp  lw 0; w 2 W; 6
cp  lw P 0; p 2 Pw ; w 2 W; 7
X
fp  Q w 0; w 2 W; 8
p2Pw

fp P 0; p 2 P; 9
ta 6 C a ; a 2 A; 10
ka ta  C a 0; a 2 A; 11
ka P 0; a 2 A; 12
where cp, the generalized path travel time, is given by
X
cp ^ta ta dp ; p 2 P; 13
a
a2A

and ^t a ta , the generalized link travel time, is dened as


^ta ta t a ta ka ; a 2 A: 14
The generalized link travel time comprises two components: ta(ta) is the normal travel time on link a 2 A, and ka is the Lagrange
multiplier associated with constraint (4). ka P 0 only if ta = Ca, and ka = 0 if ta < Ca. We can regard this multiplier, ka, as an addi-
tional time or cost penalty (besides the normal travel time) that users traveling on this saturated link are willing to wait or pay for
using the link.

3. Dynamic user equilibrium with side constraints

3.1. Preliminaries and dynamic user equilibrium

The dynamic user equilibrium formulation adopted here can be traced back to Friesz et al. (1993), and is also discussed in
detail by Friesz et al. (2001). According to Friesz and Mookherjee (2006); Friesz (2010), DUE models tend to be comprised of
four essential sub-models: (i) a model of path delay; (ii) ow dynamics; (iii) ow propagation constraints; and (iv) a route/
departure-time choice model. We will consider a nite time planning horizon T > 0 and regard time t 2 0; T  R1 as a con-
tinuous variable. Let P denote the set of all paths and jPj denote the number of paths in a network. An arbitrary path p 2 P of
:
the network of interest is dened by a sequence of the links used by that path which is denoted by p {a1, a2, . . ., am(p)},
where m(p) is the number of links used by path p.  jPj
We denote the vector of path ows as h = (hp : p 2 P) and assume that h 2 L L2 0; T , where L denotes the non-
negative cone of the jPj-fold product of the Hilbert space L2 0; T of square-integrable functions on [0, T]. As argued in Friesz
et al. (2011), the most essential component of a dynamic user equilibrium model is the path delay operator which is denoted
as Dp(t, h), "p 2 P. The path delay operator Dp(t, h) provides the travel time needed to traverse path p for per unit of ow
departing from the origin at time t. When departure time choice is enabled, the schedule delay cost function (or early/late
arrival penalty) j(v) is employed, whereby v is dened as the difference between actual and preferred arrival time denoted
by t and t < T:v = t + Dp(t, h)  t. To this end, we dene effective path delay operator Wp(t, h) by adding the schedule delay
to Dp(t, h), i.e.
Wp t; h Dp t; h jv: 15
Let Qw denote the xed total travel demand for origin-destination (OD) pair w 2 W and Pw denote the set of paths connecting
OD pair w. We need to enforce ow conservation, i.e.
X Z T
hp tdt Q w ; 8w 2 W; 16
p2P w 0

where (16) is comprised of Lebesgue integrals. Thus, we dene the feasible region for the DUE problem as
1038 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

( )
X Z T

K h : h 2 L ; hp tdt Q w ; 8w 2 W ; 17
p2P w 0

  and any nonnegative vector . . : w 2 W 2 RjWj ,


Denition 3.1. Dynamic user equilibrium (Friesz, 2010). For any h 2 K w

the pair (h , .) is a simultaneous departure-time-and-path-choice dynamic user equilibrium if and only if the following two
conditions are satised for all p 2 Pw and for all w 2 W:
 
hp t > 0 ) Wp t; h .w ;
  18
Wp t; h > .w ) hp t 0;
where .w is the smallest travel time for the OD pair w, given by

.w minf.p : p 2 Pw g P 0; and .p ess inf fWp t; h : t 2 0 Tg P 0; 19

where ess inf is essential inmum which denes the largest essential lower bound for a given function f in which all in-
ff 6 ess inf f.
As shown by Friesz et al. (1993), any solution of the following variational inequality is a solution of the DUE problem with
  such that
simultaneous departure-time-and-path-choice: nd h 2 K

 
XZ T




hWt; h ; h  h i Wp t; h hp t  hp t dt P 0; 20
p2P 0

 , where W(t, h) = (Wp(t, h):p 2 P) is the effective network delay operator.


for all h 2 K
In the above formulation of DUE, only two of the four essential sub-models of DUE are considered, i.e. a path delay model
and a route/departure-time choice model. Since the path delay operator does not explicitly depend on the detailed analytical
trafc dynamics, i.e. ow dynamics and proper ow propagation constraints, this formulation is reasonable

 if the delays are assumed to be exogenous (Friesz, 2010), i.e. the path delay operators are known in advance or repre-
sented by a simulation model, or
 if one uses the link trafc dynamics to eliminate the vector of link trafc volumes and recognizes that link exit ows are
completely determined by path ows (i.e. hp(t), "p 2 P) and link trafc volumes according to the ow propagation con-
straints (Friesz et al., 1993; Friesz and Mookherjee, 2006).

To make the formulation self-contained and to investigate more analytical properties of the DUE problem, we further
introduce the other two sub-models, i.e. (ii) and (iii), which are also known as network loading models. For a path
:
p {a1, a2, . . . , am(p)}, the dynamics of link ai is assumed to be described by the following equations:
p
dxai t
g pai1 t  g pai t; 8p 2 P; i 2 1; mp; 21
dt
X
xai t xpai tdpai ; 8ai 2 A; 22
p2P

where xpai t denotes the trafc volume on path p traversing link ai at time t, g pai t is the ow exiting link ai and g pai1 t is the
ow entering link ai by path p 2 P at time t. The total trafc volume xai t of link ai at time t is dened by (22). For i = 1, that is
the link connecting the origin of path p, we have
p
dxa1 t
hp t  g pa1 t; 8p 2 P; 23
dt
p
where  g a0 t hp t, i.e. the departure rate from P the origin of path p at time t. We assume that
xP xpai : p 2 1; jPj; i 2 1; mp 2 H1 0; Tn1 , where n1 jPj 1
p1 mp and H 0; T denotes the Sobolev space. Under this
1

assumption, we can represent the vector of link trafc volumes dened by (22), i.e. x xai : ai 2 A, as a function of time
and path ows, i.e. xt; h : R1  L ! H1 0; Tn , where n = jAj is the number of links in the network. Under certain conditions,
e.g. strong FIFO (Theorem 3.3 of Zhu and Marcotte(2000)), we can representthe link ow n operator as x(h) = x(t, h) which is
weakly continuous on K  . We assume also that g g pa : p 2 1; jPj; i 2 1; mp 2 L2 0; T 1 .
i
In addition, a link delay function (or link travel time function) Dai xai t, which denes the link travel time as a function of
the link trafc volume at the entry time to the link, is required to characterize travel time required to traverse a link, say link
ai:

1
Suppose that u 2 Lp(X),1 6 p < 1 and there exist weak derivatives oau for any a with jaj < , 2 Z+ (all derivatives up to order ), such that oau 2 Lp(X), jaj < ,
Then we say that u 2 W,p(X). We use the shorthand H1 0; T W 1;2 0; T as the Sobolev space, see e.g. Minoux (1986); Leoni (2009).
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1039

spa1 t t Da1 xa1 t; 8p 2 P;


  24
spai t spai1 t Dai xai spai1 t ; 8p 2 P; i 2 2; mp;
where spai t is known as the exit time for vehicles entering link ai at time spai1 t by path p that departed from the origin at
time t. As noticed the link delay function depends on its link trafc volume only, hence it is known as separable link delay
model. To proceed with the analysis, we assume that all link delay functions are differentiable with respect to their own
arguments. By differentiating (24) and using the link ow conservation, we can obtain the following proper ow propagation
constraints (Friesz et al., 2001):
 
g pa1 t Da1 xa1 t 1 D0a1 xa1 tx_ a1 t hp t; 8p 2 P; 25

 
g pai t Dai xai t 1 D0ai xai tx_ ai t g pai1 t; 8p 2 P; i 2 2; mp; 26

the superscript 0 denotes differentiation with respect to the associated function argument, and the superscript  denotes
differentiation with respect to time. It is clear that we need
1 D0ai xai tx_ ai t P 0 27

to ensure the non-negativeness of trafc ows, i.e. given hp(t) P 0, "p 2 P, we have g pai t
P 0; 8p 2 P; 8ai 2 A, and
xai P 0; 8ai 2 A. The condition (27) ensures the FIFO queue discipline for the DTA problems. Without loss of generality,
we assume xai 0 0; 8ai 2 A, and hp(0) = 0, "p 2 P. When Dai  is linear, Friesz et al. (1993) rst showed that the model sat-
ises the FIFO for all continuous inow proles and later Zhu and Marcotte (2000) showed the strong FIFO property for all
continuous and bounded inow proles. Xu et al. (1999) obtained fairly weakly FIFO conditions for nonlinear Dai . In this
paper, the whole-link linear travel time model proposed by Friesz et al. (1993) is adopted. The exit time of a vehicle entering
the link at time t can be calculated as:
xai t
sai t t bai ; 28
Rai
where bai is a ow-invariant travel time (free-ow travel time) of link ai. A suitable physical interpretation for the parameter
Rai is the maximum feasible constant outow from the link (Carey and McCartney, 2002) and thus we state it here as the
service rate of link ai. To calculate the path delays in terms of link delays, the nested path delay operators were proposed
by Friesz et al. (1993) which are dened as:
mp
X
Dp t; x : dpai Uai t; x; 8p 2 P; 29
i1

where
Ua1 t; x Da1 xa1 t;
!!
X
i1
Uai t; x Dai xai t Ua1    Uai1 Dai xai t Uaj ; 8i 2 2; mp:
j1

The value of path delay operator Dp(t, h) is calculated by identifying Dp(t, h) to the nested path delay operator Dp(t, x), i.e.
Dp(t, h) = Dp(t, x).

3.2. Existence of the DUE with simultaneous departure-time-and-path-choice

The existence of equilibrium to the DUE without departure time choice is proven by Zhu and Marcotte (2000), wherein
Wp(t, h) = Dp(t, h), i.e. no early/late penalty is considered. The existence result is proven by showing that the link trafc
dynamics, link delay sub-model and path delay operators under rather mild regularity conditions are continuous under
the assumption that departure rates are bounded from above, i.e. hp(t) 6 Bp, where Bp is the upper bound of the path ow.2
Under this assumption, the feasible region for the DUE without departure time choice is:
( )
X
^
K h : h 2 L ; hp t Q w t; hp t 6 Bp ; 8p 2 P w ; 8w 2 W; 8t 2 0; T :
p2P w

The existence result has not been shown for the DUE with simultaneous departure-time-and-path-choice, i.e. the variational
inequality dened by (17) and (20) admits a least one solution. To prove the existence of the DUE-SC, we will rst prove the
existence of the DUE with simultaneous route/departure time choice. In the discrete time framework, Wie et al. (2002) show

2
This is a mathematical assumption imposed by Zhu and Marcotte (2000) to ensure the path departure rates and trafc dynamics are well dened so as to
prove the existence result. Physically, a path departure rate is saturated by the service rate and storage capacity of the link connected to its origin.
1040 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

the existence result for the DUE with simultaneous departure-time-and-path-choice and elastic demand under certain
assumptions, e.g. the path delay operator Wp(t, h) is continuous and each integral constraint comprising (17) is a Riemann
integral. As we consider the integrals in (17) are comprised of Lebesgue integrals and the ow conservation constraints
themselves are interpreted as valid almost everywhere, the niteness of each Qw is not enough to assure bounded path ows,
or the path ows may not be well dened. In this case, we need to assume that the path ows are bounded as done by Zhu
and Marcotte (2000), i.e. the feasible region is further restricted as
( )
X Z T
K h : h 2 L ; hp tdt Q w ; hp t 6 Bp ; 8p 2 Pw ; 8w 2 W; 8t 2 0; T ; 30
p2P w 0

where Bp is the upper bound of path ow hp(t). In the continuous time framework, we state the following lemma for the proof
of the existence result for the DUE with simultaneous departure-time-and-path-choice:

Lemma 3.1. Assume that there is a nite interval 0; T  R1 such that

 (i) all path ow departure rates hp(t) are well dened over K;
 (ii) all links satisfy the strong FIFO condition with a uniform constant over [0, T];
 (iii) the link travel time functions are nonnegative, nondecreasing, differentiable and Lipschitz continuous over [0, T];
 (iv) the early/late penalty function j:R ? R is continuous.

Then the effective network delay operator W(h) is weakly continuous over K.

Proof of Lemma 3.1. If each integral constraint comprising (17) is assumed to be a Lebesgue integral and hp(t) is nite, con-
dition (i) holds. Because we apply the linear link travel time function and link dynamics of Zhu and Marcotte (2000), con-
ditions (ii) and (iii) are satised. Since all links satisfy the strong FIFO condition, a similar condition holds for all paths,
i.e. Dp(tj, h) > Dp(ti, h) whenever tj > ti. From Theorem 2.1 of Zhu and Marcotte (2000), Dp(t, h) is measurable and square inte-
grable. Moreover, we can represent the path delay operator as Dp(t, h) = Dp(h). From Theorem 4.1 of Zhu and Marcotte (2000),
D(h) = (Dp(h):"p 2 P) is weakly continuous over K. By denition, i.e. (15), Wp(h) = Dp(h) + j(Dp(h)). Similar to the proof of
Theorem 4.1 of Zhu and Marcotte (2000) we have j(Dp(h)) is weakly continuous over K as Dp() is weakly continuous over
K and j() is continuous over R. Thus Wp(h) is weakly continuous. The effective network delay operator W(h) is weakly con-
tinuous over K as each of its components is weakly continuous.

Proposition 3.1. Suppose all the conditions in Lemma 3.1 hold. Then the solution set of the innite dimensional variational
inequality (17) and (20) is nonempty.

Proof of Proposition 3.1. With Lemma 3.1, the existence of the DUE with simultaneous departure-time-and-path-choice
can be shown by following the proof of Theorem 4.2 of Zhu and Marcotte (2000) for the feasible set K is closed, convex,
bounded and W(h) is weakly continuous.
In fact, condition (iv) in Lemma 3.1 is consistent with the existence result of the DUE with simultaneous departure-time-
and-path-choice for bottleneck model proposed by Lindsey (2004). Proposition 3.1 proves the existence of equilibrium for
the DUE with simultaneous departure-time-and-path-choice problem. This result will be useful for us to prove the existence
result for the DUE-SC.

3.3. Problem formulation of dynamic user equilibrium with side constraint

To begin with, we rst look into an illustrative example of trafc volume control and dene the side constraints for dy-
namic user equilibrium trafc assignment.
An illustrative example of access control by controlling link trafc volume. Consider a two-lanes urban street with a
length of 100 m that admits a fundamental diagram as depicted in Fig. 1. The free-ow travel time of the street is then
0.1 min. The service rate is 60 veh/min. If the linear travel time function is adopted, the travel time of this street is
ht 0:1 xt
60
min. If vertical queue models are adopted to model the trafc dynamics and the link is assumed to be capable
of accommodating arbitrary large number of vehicles, the congestion effect would be captured by the above linear travel
time function. However, there are physical storage capacities for links to accommodate queuing vehicles. In this example,
the street would be completely blocked if the trafc volume on it equals to 24 vehicles, which may render the travel time
of the link be larger than 1 min. However, it is unlikely for us to capture the spillback effect of congestion by applying the
linear travel time function in conjunction with vertical queue models (since the link can accommodate 24 vehicles only,
which implies the maximal travel time for this link is 0.35 min).
In DTA literature, the side constraint is usually selected as the inow rate to a link is less than or equal to its service rate.
By such kind of restriction, one can eliminate the trafc congestion (or queues) (Carey and Ge, 2003). On the other hand, the
effort needed to prevent the congestion happening is captured by the Lagrange multiplier associated with the side
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1041

constraint, which can be viewed as an additional travel cost imposed on the travelers. The travelers are then charged by this
additional travel cost for using the saturated bottleneck (Yang and Meng, 1998). However, this is not practical due to the fact
that congestion is somehow inevitable. A practical method that is widely applied and has been validated by trafc control
engineering eld applications is to control the link trafc volumes to capture the potential spillback effects of trafc on
ramps and queues in store-and-forward based methods (see e.g. Papageorgiou et al. (2003); Aboudolas et al. (2009) and
the references therein). The trafc volume control can capture the potential spillback effect in the sense that the effort
needed to prevent the spillback happening can be simulated by restricting the link trafc volume be less than or equal to
its jam trafc volume (i.e. storage capacity), e.g. xa(t) 6 24 vehicles in the previous example. Therefore, we consider the trafc
volume control as a kind of side constraints in this paper. Another possible way to capture the spillback effect by the WLM is
to apply a special piecewise linear exit-ow function and explicitly restrict the exit-ow according to downstream condi-
tions, which results in a relaxed cell transmission model (CTM) when the restricted WLM is discretized (Nie, 2011). The other
way to capture the spillback effect is to apply the physical queue models (Lo and Szeto, 2002; Gentile et al., 2005; Nie and
Zhang, 2010), e.g. the CTM. A critical drawback of these two approaches is the physical queue models introduce discontin-
uous and non-differentiable trafc dynamics and travel cost functions, e.g. path delay operator. This prevents us from solving
the DTA problems analytically. The non-monotone and non-differentiability properties also lead to difculties in nding
numerical solutions, because the convergence of existing algorithms rely on either monotonicity and/or differentiability
(Szeto and Lo, 2006).
Recent studies have revealed that these macroscopic link models (including vertical queue models, e.g. the WLM and bot-
tleneck model; and physical queue models, e.g. the CTM) would produce almost the same trafc assignment result unless
there is a spillback (shockwave) (Nie and Zhang, 2005; Mun, 2007; Nie and Zhang, 2010). By this result, another way to cap-
ture the spillback effect is to adopt vertical queue models, to be precise the WLM in this paper, meanwhile to restrict the link
trafc volumes to be less than or equal to the link storage capacities on the network such that no spillback would happen.
The spillback effect is captured in terms of the effect needed to prevent the network from spillback. By doing this, we can
avoid the drawbacks of physical queue models while capturing the spillback effect. Similar to the bottleneck case, we can
regard this effect as additional travel cost imposed on travelers for using the controlled links. The travel time function
adopted in DTA literature, i.e. Da(xa(t)) dened by (24), would be more meaningful under this circumstance since if there
is spillback the link travel time function would not be separable, e.g. depends on the downstream trafc conditions like
the case of physical queue model (Szeto and Lo, 2006). Practically, the dynamic trafc volume control can also be used to
alleviate congestion spillback, avoid gridlock, and increase the safety of some critical facilities.
From the previous example and the analysis above, the side constraint considered here is the link trafc volume control or
restriction on link trafc volume, i.e.
xai t 6 C ai t; 8ai 2 A; 8t 2 0; T: 31
Inequality (31) denes the side constraints imposed on the links of the network. Let us dene the following vector-valued
function for the side constraints as C C ai : ai 2 A; where C : 0; T ! R  Rn , with R being a prescribed compact subset of
Rn with known bounds. We assume that C 2 (W1,1[0, T])n, where W1,1[0, T] denotes the space of Lipschitz functions on [0, T]
(Leoni, 2009). We rewrite the side constraints in a compact form as
xt 6 Ct; 8t 2 0; T: 32
By construction, the set Rx , {x(t):0 6 x(t) 6 C(t)} is compact.

The DUE-SC with simultaneous departure-time-and-path-choice can be formulated as: nd h 2 K such that
 
XZ T




hWt; h ; h  h i Wp t; h hp t  hp t dt P 0; 33
p2P 0

for all h 2 K, where


\
KK e;
K 34
e fh : h P 0; x 2 Rx g.
and K

Fig. 1. Fundamental diagram of an urban street.


1042 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

3.4. Existence of the DUE-SC with simultaneous departure-time-and-path-choice

Following the standard assumption of the SC-TAP, we need to presume that the feasible set K of the DUE-SC with simul-
taneous departure-time-and-path-choice is not empty. For the DUE-SC with simultaneous departure-time-and-path-choice,
we have the following proposition:

Proposition 3.2. Suppose all the conditions in Lemma 3.1 hold. Then the innite dimensional variational inequality (33) and (34)
admits a least one solution.

Proof of Proposition 3.2. For the DUE-SC, as the side constraints are introduced, the feasible region K of the original DUE prob-
lem is further restricted to K. To show the variational inequality (33) and (34) admits at least one solution, we need to show that
the new feasible region K restricted by the side constraints is bounded, convex and closed. As shown by Zhu and Marcotte (2000),
the ow operator x(h) as a function of the path ow h is weakly continuous given that conditions (i)(iii) of Lemma 3.1 hold. By
following the weak continuity of x(h) and the fact that the set Rx is constructed to be compact, the set K e is closed (Noiri, 1974;
Rose, 1984). Therefore, K is a closed set. In conjunction with the assumption that h P 0 and is bounded from above, the set K is
bounded and convex. If all the conditions in Lemma 3.1 hold, the weak continuity of the effective network delay operator W(h) can
be shown. Thus, the variational inequality dened by (33) and (34) admits at least one solution. The existence of equilibrium to the
DUE-SC with simultaneous departure-time-and-path-choice is established.

3.5. Necessary condition of the DUE-SC with simultaneous departure-time-and-path-choice

In the previous sections, we have studied the problem of existence of equilibriums to the DUE and the DUE-SC with simul-
taneous departure-time-and-path-choice, wherein the delay operators are exogenous. To analyze the necessary conditions
for the DUE and DUE-SC problems, we consider the circumstance where the delay operators are endogenous. An equivalent
form that facilitates derivation of necessary conditions for variational inequality (33) and (34) is the following differential
variational inequality (DVI) (Friesz et al., 2001; Friesz, 2010): nd (x, h, g) 2 C such that

XZ T 


hWt; x ; h  h i Wp t; x hp t  hp t dt P 0; 35
p2P 0

for all (x, h, g) 2 C, where C is the admissible set


! fx; h; g P 0 : 23; 21; 16; 25; 26; 31; and 37 holdg; 36
which incorporates zero initial conditions
xai 0 0; 8ai 2 A; hp 0 0; 8p 2 P; 37
and the ow vectors x, h, g belong to some appropriate function spaces presumed in Section 3.1.
To facilitate the analysis of the necessary conditions for (35) and (36), it is helpful to restate the DVI as the following opti-
mal control problem:
XZ T
minJ Wp t; x hp tdt; 38
8p2P 0
p
dxa1 t  
s:t hp t  g pa1 t kpa1 8p 2 P; 39
dt
p
dxai t  
g pai1 t  g pai t; kpai 8p 2 P; i 2 2; mp; 40
dt
dEw t X
hp t; lw 8w 2 W; 41
dt p2P w
   
g pa1 t Da1 xa1 t 1 D0a1 xa1 tx_ a1 t hp t; cpa1 8p 2 P; 42
   
g pai t Dai xai t 1 D0ai xai tx_ ai t g pai1 t; cpai 8p 2 P; i 2 2; mp; 43
 
 hp t 6 0; qpa0 8p 2 P; 44
 
 g pai t 6 0; qpai 8p 2 P; 45
 
 xpai t 6 0; fpai 8p 2 P; 46
xai t 6 C ai t; gai 8ai 2 A; 47
Ew T Q w ; /w 8w 2 W; 48
xai 0 0; 8ai 2 A; Ew 0 0; 8w 2 W; hp 0 0; 8p 2 P; 49
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1043

where Eqs. (41) and (48) dene the ow conservation constraints, where Ew(t) is an extended state (Friesz et al., 2001; Friesz,
2010), which are equivalent to (16). (49) species the zero initial conditions, which mean that the network is empty at the
P
beginning. (47) is the side constraint imposed, where xai t p2P xpai tdpai ; 8ai 2 A. The variables in brackets are Lagrange mul-
tipliers associated with the corresponding constraints. In optimal control theory, (46) and (47) are known as pure state var-
iable inequality constraints as they depend on the state variables only while (42) (45) can be categorized as mixed (state
and control) constraints as they explicitly depend on the controls. (48) and (49) are the terminal and initial conditions,
respectively.
The following proposition states the necessary condition for the DUE-SC with simultaneous departure-time-and-path-
choice, which can be recognized as the type of equilibrium given by (18) and (19).

Proposition 3.3. The necessary condition for the DUE-SC with simultaneous departure-time-and-path-choice can be stated as
follows:

> 0 ) Wp t; x lp t /w ;
hp t 8p 2 Pw ; w 2 W; 50
0 ) Wp t; x lp t > /w ;
where /w is the travel cost of OD pair w under the DUE-SC condition, which is determined by the xed total travel demand Qw of the
Pmp1 R spai1 t
OD pair. lp t : i0 sp t
gai1 udpai1 du, where gai is the Lagrange multiplier associated with the side constraint imposed on
ai

link ai which satises the following complementary slackness condition


gai t P 0; xai t  C ai t 6 0; gai txai t  C ai t 0; 8ai 2 A;
and spai t is the entry time to link ai+1 for vehicles traveling by path p that departed from the origin at time t, which is dened by
(24).

Proof of Proposition 3.3. Our intent is to motivate why the DUE condition (50) will be fullled when the postulated VI (33)
and (34) or DVI (35) and (36) is solved. To this end, we prove in the Appendix that constraint qualications for the pure state
constraints and the mixed constraints are satised and, therefore, the desired dual variables (or adjoint functions (equa-
tions)) are available. However, in optimal control theory, the proper foundation for knowing that the dual variables of a large
system of multiple pure state constraints, like those dual variables needed to express the DUE condition (50), is a computa-
tional one. The appropriate dual variables and their trajectories are obtained when our computational approach based on the
xed point algorithm converges. In this proof, we derive the necessary condition for the DUE-SC by applying the Pontryagin
Minimum Principle. To simplify the notation and proceed to the analysis, let us dene
g~pai t g pai t Dai xai t; 8p 2 P; i 2 1; mp: 51

The Hamiltonian function for the optimal control problem is


"
X   X   
Ht Wp t; x hp t kpa1 t hp t  g pa1 t kpai t g pai1 t  g pai t cpa1 t hp t  g~pa1 t
p2P i22;mp
  X    X
1 D0a1 xa1 tx_ a1 t cpai t g pai1 t  g~pai t 1 D0ai xai tx_ ai t gai txai t  C ai t
i22;mp i21;mp
# !!
X X X X
qpa0 thp t  qpai tg pai t fpai txpai t lw t hp t : 52
i21;mp i21;mp w2W p2P w
  
To simplify the notation, let us further dene t pi t pi1 Dai xpai t pi1 , where t pi1 is the time of the ow g pai1 entering link
p p
ai 2 P and ti is the time of the ow g ai exiting the link, which is a shorthand of (24). The partial derivatives of the Hamiltonian
function are given as
@Ht
Wp t; x kpa1 t lw t cpa1 t  qpa0 t; 8w 2 W; p 2 Pw ; 53
@hp t
@Ht
kpai t kpai1 t  qpai t cpai1 t; 8p 2 P; i 2 1; mp  1; 54
@g pai t
@Ht
kpamp t; 8p 2 P; 55
@g pamp t
@Ht  
p cpai t 1 D0ai xai x_ ai ; 8p 2 P; i 2 1; mp; 56
@ g~ai t
@Ht
fpai t gai tdpai ; 8p 2 P; i 2 1; mp: 57
@xpai t
In addition, a set of complementary slackness conditions for the nonnegative ow constraints and the side constraints are:
1044 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

hp t P 0; qpa0 t P 0; qpa0 thp t 0; 8p 2 P; 58


g pai t P 0; qpai t P 0; qpai tg pai t 0; 8p 2 P; 8i 2 1; mp; 59
xpai t P 0; fpai t P 0; fpai xpai t 0; 8p 2 P; 8i 2 1; mp; 60
gai t P 0; xai t  C ai t 6 0; gai txai t  C ai t 0; 8ai 2 A: 61

For an open path p = {a1, a2, . . . , ai1, ai, ai+1, . . . , am(p)} 2 Pw, we have (Friesz et al., 2001)
   
g pa0 t p0 hp tp0 > 0;
 
g pai t pi1 > 0; 8i 2 1; mp;
 
g pai t pi > 0; 8i 2 1; mp;
xpai t > 0; 8t 2 t pi1 ; t pi ; i 2 1; mp:

The Lagrange multipliers of the complementary slackness conditions (58) and (60) are relaxed to
 
qpa0 tp0 0;
   
qpai tpi1 qpai tpi 0; 8i 2 1; mp;
 
fpai t 0; 8t 2 t pi1 ; t pi ; i 2 1; mp:

With this in hand, we are ready to address the rst order necessary condition for the optimal control problem. First, the fol-
lowing stationary conditions hold

@H
0; 8w 2 W; p 2 Pw ; 62
@hp tp
0 " #
@H @H 1
0; 8p 2 P; i 2 1; mp: 63
@g pai p @ g~pai 1 D0ai xai x_ ai p
ti t i1

The adjoint equations3 are given by


dkpai t @Ht
 gai tdpai  fpai t; 8p 2 P; i 2 1; mp; 64
dt @xpai t

dlw t
0; 8w 2 W: 65
dt
At terminal time T, the following boundary conditions hold:
kpai T 0; lw T  /w T 0; 8w 2 W; 8p 2 P w ; i 2 1; mp: 66

Together with (65), the following equality is obtained:


lw t /w T lw ; 8w 2 W: 67
Note that (62) is equal to
     
Wp t p0 ; x kpa1 t p0 lw cpa1 tp0 0; 8w 2 W; p 2 P; 68
 
where we use the fact that lw(t) = lw, which is a constant and for an open path q t p0 0. Now, by incorporating p
a0
 
qpai tpi 0; 8i 2 1; mp, the stationary condition (63) gives
       
kpai1 t pi cpai1 tpi kpai tpi c
~pa tpi ;
i
8p 2 P; i 2 1; mp  1; 69
   
kpamp tpmp  c
~pa
mp
t p
mp 0; 8p 2 P; 70

where we adopt the following notation in line with (Friesz et al., 2001; Friesz, 2010):
 
c~pai tpi : cpai tpi1 ; 8p 2 P; i 2 1; mp: 71

The adjoint Eq. (64) for an open path can then be given as
dkpai t  
gai tdpai ; 8p 2 P; i 2 1; mp; t 2 t pi1 ; t pi : 72
dt

dkpa t @Ht
3
In Friesz et al. (2001); Friesz (2010), the adjoint equations are given as i
dt
 @xp
t
dtd @@Ht
x_ p t
. It is proven by Friesz (2010) that the term d @Ht
dt @ x_ pa t
0.
ai ai i
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1045

Evaluating the adjoint Eq. (72) backward in time and space yields the following. To begin with, let us apply (72) to the last
link of path p, i.e. link apmp :
    Z t pmp
kpamp tpmp1  kpamp t pmp gamp tdpamp dt: 73
p
t mp1

From (70) we have


     
c~pamp tpmp cpamp tpmp1 kpamp tpmp : 74

From (69), (72) and (74), the following relation is obtained


        Z p
tmp
dmp1 kpamp tpmp1 cpamp t pmp1 kpamp t pmp1  kpamp t pmp gamp tdpamp dt
p
t mp1
   
kpamp1 t pmp1 c
~pa
mp1
t pmp1 : 75

Similarly,
       
dmp2 kpamp1 t pmp2 cpamp1 t pmp2 kpamp2 t pmp2 c
~pa
mp2
t pmp2 : 76

From (76) we have


    Z p
tmp1
dmp1  dmp2 kpamp1 t pmp1  kpamp1 t pmp2  gamp1 tdpamp1 dt: 77
p
t mp2

Proceeding the induction, we have


Z p
t i1
di di1 gai1 tdpai1 dt; 8i 2 1; mp  1: 78
tpi

Finally, we have
Z tpi1
    mp1
X
d1 kpa1 t p1 cpa1 tp0 gai1 tdpai1 dt: 79
p
i1 ti

Thus,
Z t1
p mp1
X Z p
t i1
       
lp tp0 kpa1 t p0  kpa1 t p1 kpa1 t p1 cpa1 t p0 ga1 tdpa1 dt d1 gai1 tdpai1 dt: 80
p p
t0 i0 ti

 
The dynamic user equilibrium condition with side constraints is as follows: for an open path p 2 Pw, hp tp0 > 0,
 
Wp t p0 ; x lp tp0 /w T;
and if
     
Wp t p0 ; x lp t p0 > /w T; p 2 Pw ) hp t p0 0: 81

As tp0 is arbitrarily, without loss of generality, we represent (81) as


Wp t; x lp t /w T;
and if

Wp t; x lp t > /w T; p 2 Pw ) hp t 0; 82

which is immediately recognized as the fundamental condition for the dynamic network user equilibrium described by (18)
and (19). We also have the following complementary slackness conditions for the side constraints: (i) if xai t < C ai t, the
Lagrange multiplier gai t 0, (ii) if xai t C ai t, the Lagrange multiplier gai t P 0. The value of /w(T) is determined by
the total amount of trafc Qw. As /w(T) is a constant, we drop the time script T to denote it as /w. Wp(t, x) is interpreted
as the effective path delay at time t on path p under travel condition x.

Remark 3.1. To avoid the spillback effect of congestion, the side constraints C ai t; 8ai 2 A are chosen as jam trafc volumes
(i.e. storage capacities) of the links, e.g. the trafc volume Ka that corresponds to the jam density qJ in Fig. 1. In this case, we
have that xai t 6 K ai ; 8ai 2 A. Clearly, when xai t < K ai ; 8ai 2 A; t 2 0; T, the problem reduces to normal DUE trafc
assignment. For the case when xai t K ai for some links during several time intervals, we have the following interesting
implication that
1046 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

d d
x_ ai xa K a 0:
dt i dt i
Under this circumstance, the ow propagation becomes
g pai t Dai xai t  g pai1 t 0;

which implies that trafc behaves like an incompressible uid when its density (or volume) approaches certain maximum,
i.e. the ow is incompressible when trafc density (or volume) is equal to the jam density (or volume). This phenomenon is
consistent with the ndings in Ross (1988) where the LWR model was applied to describe the trafc dynamics.

3.6. Interpretation of DUE with side constraints

For a path p with hp(t) > 0, we have qpa0 t 0. Furthermore, if there p is no side constraint, i.e. xai t <
dka t
C ai t; 8ai 2 A; 8t 2 0; T, we have gai t 0. The adjoint Eq. (72), implies that, dti 0, i.e. kpai t is a constant with respect
to time. This, in conjunction with the boundary condition kpai T 0, implies that Wp(t,x) = /w(T), which is consistent with
the normal DUE condition (Friesz et al., 2001).
We can also show that the DUE-SC can reduce to the SC-TAP. The system is operating in a steady state in the static case,
dxa t
i.e. hp(t) = hp, dti 0 ) xai t xai ; 8ai 2 A, and all of the side constraints are constant. By dening the performance index
(38) to be (1), the necessary condition for the DUE-SC reduce to the necessary condition for the SC-TAP. From Lemma 2.1, the
SC-TAP can be written as

> 0 ) c p lw ;
fp 8p 2 Pw ; w 2 W: 83
0 ) c p P lw ;
Rs p
ai t
The necessary conditions for the DUE-SC and the SC-TAP are similar. The term s gai udpai du, can be regarded as an addi-
p
ai1 t
tional time penalty, or toll (access price) to be imposed on travelers during their presence on the controlled link ai along their
travel path p, which is similar to the interpretation of the term kai in the SC-TAP. Note that the additional time penalty kai in
the static case is constant over time, whereas the additional time penalty in the DUE-SC is time-varying and a cumulative
effect of the associated Lagrange multiplier over time.

4. Solution algorithm

4.1. The xed point problem and its optimal control formulation

Friesz et al. (2001) stated that the optimal control problem formulation cannot be used for the computation of the DUE/

DUE-SC because its articulation presumes the knowledge of departure rates hp t; 8p 2 P, under the DUE/DUE-SC condition.
The optimal control problem formulation is a mathematical convenience for analyzing the necessary conditions of the DUE/
DUE-SC. There are several solution algorithms for the DUE without departure time choice, such as a quasi-VI based approach
(Ban et al., 2008) or an equivalent gap function-based algorithm (Lu et al., 2009). However, only a few solution algorithms,
such as those proposed by Huang and Lam (2002), Friesz and Mookherjee (2006), Friesz et al. (2011), have been proposed for
the DUE with simultaneous departure-time-and-path-choice. The DUE with simultaneous departure-time-and-path-choice
has been shown to be equivalent to a xed point problem. As explained in the previous section, the side constraints can be
viewed as a restriction on the feasible region of the VI formulation of the DUE-SC, i.e. (33) and (34). Based on this VI formu-
lation, the DUE-SC is equivalent to the following xed point problem:

Lemma 4.1. (Huang and Lam, 2002) (Fixed point problem) When the conditions in Section 3.4, which guarantee the existence of
the DUE-SC, hold, any solution of the xed point problem

h PK h  aWh; 84
is also a solution of the DUE-SC, where PK x argminz2K kx  zk, i.e. the minimum norm projection onto K and a 2 R1 .
In line with Friesz and Mookherjee (2006), the xed point problem (84) requires that


1
h arg min kh  aWh  v k2 : v 2 K ; 85
v 2
which is equivalent to seeking the solution of the following optimal control (innite dimensional mathematical program-
ming) problem
Z T
1
min Jv h  aWh  v T h  aWh  v dt: 86
v 2K 2 0

The advantage of the VI formulation of the DUE/DUE-SC is that it includes almost all DUE models regardless of the link trafc
dynamics, ow propagation, and link travel time function employed. However, in this formulation, the effective path delay
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1047

Fig. 2. Fixed point algorithm for DUE-SC.

operators Wp(t, h), "p 2 P, generally cannot be expressed in a closed form whereas Wp(t, x), "p 2 P are well dened in a closed
form by employing the link trafc dynamics (39) and (40), ow propagation constraints (42) and (43), and certain travel time
function. By Lemma 4.1 and (86), the xed point algorithm can be formulated and solved as an iterative optimal control
problem as depicted in Fig. 2. In the gure, the effective path delay operators Wp(t, x), "p 2 P, in conjunction with (39),
(43) and (28) are applied to replace W(h) in (86), and to express the side constraints and the feasible region K explicitly
rather than the abstract embedded formulation used in Section 3.3. In Fig. 2, x_ ~f  denotes the link trafc dynamics with
the ow propagation constraints (39) and (43). We write x_ explicitly because the link trafc volumes are the state variables
of the optimal control problem. The ow propagation constraints (42) and (43) with travel time function dened by (28) are
substituted into (39) and (40) to update the state variables. The equality constraint M1() denotes the ow conservation Eqs.
(41) and (48) with zero initial condition. The inequality constraint M2() denotes the pure state constraints. v P 0 is a vector
representation of (44).4 Detailed analysis of the convergence of such a xed point algorithm is given by Friesz and Mookherjee
(2006) for an abstract feasible region with certain properties. However, some conditions required to guarantee the convergence
of the algorithm, such as the strong monotone of F(xk, hk, v), are unlikely to be veried for general trafc networks. Thus, the
convergence of this xed point algorithm is generally heuristic.
As Fig. 2 indicates, the optimal control subproblem is what we need to solve. To the best of our knowledge, no effective
algorithm has been reported to solve the optimal control subproblem with state dependent time lags and state constraints.
However, several effective algorithms have been proposed to solve the optimal control problem with state and/or control
constraints using nonlinear programming algorithms, such as sequential quadratic programming (SQP) (Buskens and Maur-
er, 2000). To solve this optimal control problem by nonlinear programming algorithms, it is necessary to approximate the
functional differential equations (FDEs), which govern the trafc dynamics (i.e. x_ ~f ), by ordinary differential equations
(ODEs). We use a typical method for handling transportation delays in control engineering to deal with these state depen-
dent time lags. The idea is to use polynomial approximation of the time lags. A similar idea has been proposed by Astarita
(1996), Friesz and Mookherjee (2006). We use Pad approximant to approximate the state dependent time lags, which can be
easily implemented using Matlab and Simulink.
After approximating the FDEs by ODEs, we denote the trafc dynamics as xt _ f0 x; v ; t. It is convenient for us to rewrite
the optimal control subproblem as

Z T
k
minJ k v Fxk ; h ; v dt; 87
v 0
_
s:t: xt f0 x; v ; t; M 1 v ; t 0; M 2 x; t 6 0;
 v 6 0; 8t 2 0; T; x0 0: 88

4
Other non-negative ow constraints are not explicitly stated since they are implicitly included in this formulation and can be guaranteed by the strong
FIFO.
1048 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

4.2. Solution algorithm for the DUE-SC

To apply an off-the-shelf nonlinear optimization algorithm to the optimal control problem (87) and (88), it is necessary to
apply the time-discretization scheme to the problem. Here we present the recursive discretization approach based on Eulers
method, which is commonly used in the literature (Buskens and Maurer, 2000). The planning time interval is divided into
T
N  1 segments uniformly, i.e. a xed time step Dt is dened in the discretization as Dt N1 ; tl
l  1Dt; l 1; 2; . . . ; N. Applying Eulers method to the differential equation in (88) yields
xl1 xl Dt  f0 xl ; v l ; tl ; l 1; 2; . . . ; N  1: 89
T
Dene the optimization variable z = [v1, . . . , vN] , and compute the state variables from (89) recursively as
xl = xl(z, t) = xl(v1, v2, . . . , vl1, tl1), l = 2, . . . , N, which are functions of the control variables with initial condition
x1 = x(1) = 0. The following NLP is dened
X
N1
k
minyz, Dt  Fxk ; h ; v l ; 90
z
l1

s:t: xl xl v 1 ; v 2 ; . . . ; v l1 ; tl1 ; l 2; . . . ; N; x1 0;


M 1 v l ; t l 0; M 2 xl ; t l 6 0; v l 6 0; l 1; 2; . . . ; N: 91

After the optimal control subproblem is reformulated as an NLP problem, we can apply nonlinear optimization algorithms to
solve it. The overall solution algorithm is summarized in Fig. 3.

Step 1. Approximation of FDEs: Apply polynomial approximants (such as the Pad approximant) to approximate the state
dependent time delays, which converts the original FDEs into ODEs.
Step 2. NLP formulation of the optimal control subproblem: Formulate a nonlinear programming problem of the optimal
control problem (87) and (88) using the Eulers discretization approach.
1
Step 3. Initialization: Set initial feasible solutions of state and inow x1 ; h 2 K, iteration counter k = 1.
Step 4. Solution of the optimal control subproblem: Apply any off-the-shelf nonlinear optimization algorithm to the equiv-
alent NLP problem (e.g. SQP or GAUSS Pseudo Spectral Method).
Step 5. Check the convergence of the optimal control subproblem: If the preset tolerance e1 or the maximum iteration num-
ber Iterm of the NLP is achieved, declare the solution to be hk+1 and go to Step 6. Otherwise declare the solution to be
the initial condition of the NLP and go to Step 4.
Step 6. Stopping test: If khk+1  hkk 6 e,where e 2 R+ is a preset tolerance, stop and declare h hk+1. Otherwise, set k = k + 1
and go to Step 3.

5. Numerical examples

We have imposed upper bounds on path ows so as to prove the existence of equilibriums to the DUE and the DUE-SC. In
this section, the numerical examples are solved with the upper bounds on path ows set to +1.

5.1. A simple network case

Consider a network with a single OD pair connected by two parallel links as shown in Fig. 4. The link delay functions are
given by D(x1) = 0.27 + x1/70 unit-times and D(x2) = 0.3 + x2/140 unit-times, respectively. The overall travel demand is Jod = 75
units. The planning time horizon is T = 5 unit-times. The desired arrival time is set as tda = 3 unit-time. We consider the fol-
lowing symmetric early/late arrival penalty function:
8 2
>
< 0:1t Dx  t de ; t < tde ;
jv 0; t de 6 t 6 t dl ; 92
>
:
0:1t Dx  t dl 2 ; t > tdl ;
where tde = 2 unit-time and tdl = 4 unit-time. The effective delay is given by (15). Without loss of generality, we assume zero
initial conditions. a = 0.01 is chosen in the xed-point algorithm as represented in Fig. 2, and the xed point stopping criteria
is e = 0.0001.
We rst solve the DUE with simultaneous departure-time-and-path-choice of this network. Fig. 5 shows that at the very
beginning of the departure stage, the travel cost of link 1 is less than that of link 2 and not greater than the minimum travel
cost. Thus, travelers all choose to use link 1. After some time, the costs of the two links tend to the same level, i.e. the DUE
cost. Travelers then select their departure times to maintain the equilibrium state. Note that the inows become zero when-
ever the travel times on the links are greater than the minimum travel time, which is consistent with the DUE condition.
Fig. 5b demonstrates the link trafc volumes of both links under the DUE condition.
Next, we restrict the trafc volume on link 1 to be less than or equals to 3 units during the whole planning horizon, i.e.
x1 6 3. By applying the DUE-SC condition, we obtain the results that are depicted in Fig. 6. When there is no trafc volume
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1049

Fig. 3. Solution algorithm for DUE-SC.

control, the trafc volume on link 1 exceeds the desired saturation, i.e. x1 = 3, after one unit-time as demonstrated in Fig. 5b.
Fig. 6b illustrates that the control parameter (or additional travel cost induced by the side constraint) for link 1 becomes
positive at around one unit-time, which corresponds with the time that the side constraint is rst violated. Note that the
control parameter or additional travel cost imposed on link 1 is calculated following Eq. (80). Fig. 6a shows the travelers
responses to the additional travel cost imposed on link 1. The departure rate of link 1 suddenly decreases at around one
unit-time (when the additional travel cost is rst imposed) and then remains at the saturation level of around 9 units/
time-steps (which is lower than the saturation level of around 12 of the uncontrolled case as shown in Fig. 5a). Fig. 6a also
shows an increase of the departure rate of link 2 as compared to the uncontrolled case (from the maximum ow rate of 12 to
more than 14 units). This illustrates the diversion of trafc from link 1 to 2 to maintain the trafc volume of link 1 and the
DUE-SC condition. We can also observe a slightly larger travel time window for those traveling by link 2 as compared to the
uncontrolled case, which highlights the shift of travelers departure times to maintain the DUE-SC condition. Fig. 6b shows
that the trafc volume of link 1 under DUE-SC condition satises the side constraint. The value of additional travel cost in-
duced by the side constraint varies from 2.5% to 8.1% of the minimum effective travel time. Compared to the DUE case, the
number of travelers using link 1 declines by about 23% constant trafc volume control scheme in this example will yield a
steady state to the departure rate of link 1 if the departure time window is long enough. As demonstrated in Fig. 6a, the
departure rate of link 1 gradually approaches a steady state controlled by the trafc volume control scheme. As we do
not impose a side constraint on link 2, the additional travel cost of link 2 is zero throughout the planning horizon.
For a congested roadway, different trafc volume control schemes can be implemented to alleviate the congestion. For
instance, the control may vary by time, or different restrictions on trafc volumes may be adopted during different peak
 
p
hours. In the second test, we impose the trafc volume control function x1 t 6 3 sin t20100
on link 1. This constraint rep-
resents a kind of control that adjusts its amplitude responding to the peak hour trafc volume. In this example, the control
has a small value at the beginning, and gradually increases as the time approaches the peak hour. After the peak period, the
control then decreases. By applying the DUE-SC condition, we obtain the results depicted in Fig. 7. Compared with the pre-
vious constant trafc volume control scheme, the travelers using link 1 are indirectly controlled to depart earlier to satisfy

Fig. 4. Network connected with parallel links.


1050 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

(a) Comparison of path flow and associated (b) The traffic volumes of both links
travel cost
18 2 6
Path flow of path 1 Traffic volume of link 1
Path flow of path 2 Traffic volume of link 2
16
Travel cost of path 1
5
Travel cost of path 2
14
1.5

12 4

Traffic volume
Travel cost
Path flow

10
1 3
8

6 2

0.5
4
1
2

0 0 0
0 1 2 3 4 5 0 1 2 3 4 5
Time Time

Fig. 5. Inow proles, link trafc volumes and travel costs of both links under the DUE with symmetric penalty.

(a) Comparison of path flow and associated (b) Comparison of link traffic volume and
generalized travel cost additional travel cost

20 3 8 0.1
Path flow of path 1 Traffic volume of link 1
18 Path flow of path 2 Traffic volume of link 2
Generalized travel cost of path 1 7
Additional travel cost of link 1
2.5
16 Generalized travel cost of path 2 Additional travel cost of link 2
6
14

Additional travel cost


2
5
Traffic volume

12
Travel cost
Path flow

10 1.5 4

8
3
1
6
2
4
0.5
1
2

0 0 0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time Time

Fig. 6. Inow proles, link trafc volumes and travel costs of both links under constant trafc volume control.

the new trafc volume control scheme. The inow of link 2 does not change signicantly. The inow of the controlled link is
shaped by the trafc volume control with some oscillations. This veries that the control adjusts the departure rate of the
controlled link to maintain its trafc volume and satisfy the time-varying trafc volume control scheme.
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1051

In the third test, a trafc volume control scheme with two step functions is tested. In our continuous time formulation,
the trafc volume controls (or side constraints) are assumed to be continuously differentiable functions. This kind of discon-
tinuous dynamic trafc volume control cannot be tested under the continuous time DUE-SC framework. However, due to the
discretization scheme adopted in our solution algorithm, we can construct the following trafc volume control scheme, i.e.
x1 k 6 2; 8k 1;    ; 19; and x1 k 6 3; 8k 20;    ; 50: 93
The simulation results are plotted in Fig. 8. By comparing Figs. 8 and 6, we can see that, some travelers who originally
planned to use link 1 (see Fig. 6) are now diverted to link 2. Some of the travelers on link 1 also depart earlier to avoid
the additional travel cost. These effects arise because more stringent trafc volume control is imposed on link 1 during
the rst stage. The additional travel cost will be higher if the travelers persist in traveling during the same time period
and on the same link. In the second stage, the trafc volume control forces the departure rate of link 1 to the steady state,
which is determined by the side constraint. As demonstrated in Fig. 8, the control forces the trafc volume on link 1 to meet
the restriction in a smooth manner rather than a sudden switch, which is consistent with our assumption of continuity in the
trafc volume control. The additional travel cost increases during the switch of these two trafc volume control strategies to
depress the travel demand attracted by the relaxation of the side constraint. After the switch of the trafc volume control
strategies the cost decreases, which in turn increases the travel demand on link 1. The control is then changed to adjust
the departure rate of the controlled link so that the side constraint is satised.

5.2. The Braess Network

Next, we will test the Braess network depicted in Fig. 9. The forward star array and link travel time functions of this net-
work are summarized in Table 1. The planning horizon is T = 6 unit-times. The time incremental step is 0.1 unit-times. The
desired arrival time is tda = 3 unit-time. The overall travel demand is J14 = 200 units. The symmetric early/late arrival penalty
function (92) is applied in this example with tde = 2 unit-time and tdl = 5 unit-time. We dene the path set as P14 , {p1, p2, p3}
with p1 , {a1, a4}, p2 , {a1, a3, a5} and p3 , {a2, a5}. Without loss of generality, we assume zero initial conditions. We choose
a = 0.01 for the xed point algorithm. The xed point stopping criteria is e 2 [0.00001, 0.0001].
First, the normal DUE with simultaneous departure-time-and-path-choice is solved. Fig. 10 illustrates the path departure
rates and the path travel costs under the DUE condition for comparison. From Fig. 10, at the very beginning none of the path
is used because the travel costs are higher than the minimum travel cost. Path p2 becomes active rst due to its relatively
lower free-ow travel cost. After a short time, the costs of the other two paths tend to the DUE cost. All three paths then
become active, and these departure rates are adjusted to maintain the DUE condition. The path departure rates become zero

(a) Comparison of path flow and associated (b) Comparison of link traffic volume and
generalized travel cost additional travel cost

20 3 8 0.1
Path flow of path 1 Traffic volume of link 1
18 Path flow of path 2 Traffic volume of link 2
Generalized travel cost of path 1 7 Additional travel cost of link 1
2.5
Generalized travel cost of path 2 Additional travel cost of link 2
16
6
14
2
Additional travel cost

5
Traffic volume

12
Travel cost
Path flow

10 1.5 4

8
3
1
6
2
4
0.5
1
2

0 0 0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time Time

Fig. 7. Inow proles, link trafc volumes and travel costs of both links under time-varying trafc volume control.
1052 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

(a) Comparison of path flow and associated (b) Comparison of link traffic volume and
generalized travel cost additional travel cost

20 Path flow of path 1 3 8 0.1


Traffic volume of link 1
Path flow of path 2
Traffic volume of link 2
18 Generalized travel cost of path 1 7 Additional travel cost of link 1
Generalized travel cost of path 2
2.5 Additional travel cost of link 2
16
6
14

Additional travel cost


2
5

Traffic volume
12

Travel cost
Path flow

10 1.5 4

8 3
1
6
2
4
0.5
1
2

0 0 0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time Time

Fig. 8. Inow proles, link trafc volumes and travel costs of both links under step trafc volume control.

Fig. 9. The Braess network.

whenever the travel costs of the paths are greater than the minimum travel cost, which is consistent with the DUE condition.
Fig. 11 demonstrates the link trafc volumes of the ve links under the DUE condition.
Next, the trafc volume on link a3 are controlled because it has a relatively smaller capacity. In this example, we impose
the side constraint xa3 6 2 on link a3. By solving the DUE-SC, we obtain the results depicted in Fig. 12. Compared with the
previous DUE test, we note that the departure rate of the controlled path, i.e. path p2, decreases throughout the departure
time window. As demonstrated in Fig. 12, this amount of trafc is mainly diverted to path p1, whereas the departure rate
of path p3 does not change signicantly. Travelers on all of these paths depart earlier to avoid the additional travel cost.
Fig. 13 illustrates that the additional travel cost for link a35 becomes positive at around one unit-time, which corresponds
to the time that the side constraint is rst violated. The departure rate of path p2 decreases sharply at the same time, compared
with the DUE case in Fig. 10. This veries that the control (i.e. the additional travel cost) adjusts the departure rate to maintain
the trafc volume of link a3 to satisfy the side constraint. The amount of time-varying additional travel cost caused by the side
constraint varies from 0.5% to 12.5% of the travel time of link a3. The path ow of p2 under the DUE-SC condition is about 64% of
that under the DUE condition. As in the previous example, the departure rate of path p2 is shaped by the side constraint with
some oscillations.
Finally, we test the DUE-SC for this network with time-varying side constraints on different links. To be more specic, we
impose the following side constraints on links a2 and a3:

5
Note from (47) and the proof of Proposition 3.3 that the side constraints and the associated link additional travel costs (i.e. integrals of Lagrange multipliers
associated with the side constraints over time) are dened for the links on the network, i.e. they are link based. They are then converted to path additional
travel costs by Eq. (80).
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1053

Table 1
Link congurations.

Link name From node To node Link delay function, Da(xa(t))


xa
a1 1 2 0:5 2000
1

xa
a2 1 3 0:71 2000
2

xa3
a3 2 3 0:18 100
xa4
a4 2 4 0:71 2000
xa5
a5 3 4 0:5 2000

Comparison of path flow and associated travel cost


28 4
Path flow of p1
26
Path flow of p
2 3.5
24
Path flow of p3
22 Travel cost of p1
3
20 Travel cost of p2
18 Travel cost of p
3 2.5

Travel cost
Path flow

16
14 2
12
10 1.5

8
1
6
4 0.5
2
0 0
0 1 2 3 4 5 6
Time

Fig. 10. Path departure rates and travel costs of the Braess network under the DUE condition.


t 20p t 20p
xa2 t 6 15 sin units; xa3 t 6 2 sin units: 94
100 100

Fig. 14 depicts the path inow rates of the three paths and the associated travel costs. Because we impose the side constraint
on link a2, paths p1, and p3 are no longer symmetric. Consequently, travelers on path p3 depart earlier than in the case with-
out side constraint to avoid the additional travel cost. Meanwhile, the amplitude of the path ow curve of p3 decreases to
satisfy the side constraint imposed on link a2. The departure rate of path p2 further deceases because the new side constraint
imposed on link a3 has relatively smaller values. Travelers switch to path p1 to avoid the penalty caused by the side con-
straints as illustrated in Fig. 14. Fig. 15 shows the trafc volumes of the ve links. The comparison of link trafc volume,
time-varying side constraint and the associated additional travel cost is depicted in Fig. 16. The additional travel cost of link
a3 has larger value when xa3 rst reach the level specied by the side constraint. As a result, the trafc is diverted to other
paths, which in turn decreases the trafc volume on link a3. As indicated by Fig. 16, the additional travel cost of link a3 is
generally larger than that of link a2, perhaps because path p3 has a larger free-ow time than path p2. Travelers are more
sensitive to the additional travel cost on that path. The path ows of paths p2 and p3 in this case are about 56% and 97%
of those under the DUE condition, respectively. This veries that the smaller the additional travel cost, the smaller the con-
trol effect it yielded.
The convergence errors for these three test with respect to iterations are depicted in Fig. 17, where the error is dened as
k1 k
ek1 kh kh k. As shown in the gure, the proposed numerical solution algorithm converges sharply for this example, and
kh k
the proposed numerical solution method converges faster with the DUE-SC. This occurs because the DUE-SC has a relatively
smaller search region (or feasible region) for the algorithm if it admits a solution.
One should note that the cumulative trafc volume on each link does not tend to zero as time tends to the end of the
planning horizon. This phenomenon is also observed by Friesz and Mookherjee (2006), Friesz et al. (2008). Numerical errors
of the network loading may be one of the reasons. However, this phenomenon occurs because of the so-called double-count-
ing-effect of the WLM (Nie and Zhang, 2002; Nie and Zhang, 2005).
1054 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

Traffic volumes of the five links


25
Traffic volume of a
1
Traffic volume of a
2
Traffic volume of a
20 3
Traffic volume of a4
Traffic volume of a5
Traffic volume

15

10

0
0 1 2 3 4 5 6
Time

Fig. 11. Trafc volumes of the ve links under the DUE condition.

Comparison of path flow and associated generalized travel cost


28 5
Path flow of p1
26
Path flow of p 4.5
2
24 Path flow of p3
22 Travel cost of p1 4
Travel cost of p2
20 3.5
Travel cost of p3
18

Travel cost
3
Path flow

16
14 2.5
12
2
10
8 1.5

6 1
4
0.5
2
0 0
0 1 2 3 4 5 6
Time

Fig. 12. Path departure rates and travel costs of the Braess network under constant trafc volume control.

6. Conclusions

The trafc assignment with side constraint problem is extended to the dynamic case in this paper to allow the study of a
trafc volume control scheme. The side constraints are related to the desired temporal trafc volumes on certain links, which
can be set according to the safety or environmental requirements. The dynamic user equilibrium problem with side con-
straints and simultaneous departure-time-and-path-choice is formulated as an innite-dimensional variational inequality.
We show the existence of equilibrium to the DUE-SC based on the VI formulation under certain assumptions. To analyze
the necessary condition, we restate the problem as an equivalent optimal control problem. The optimality condition of
the DUE-SC is obtained by applying the Pontryagin minimum principle to the optimal control problem. The equilibrium dy-
namic travel cost under the DUE-SC condition is shown to be the effective path delay function plus a term of additional travel
cost induced by the side constraints. This additional travel cost term is governed by the accumulation of the Lagrange
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1055

Comparison of traffic volume and additional travel cost


25 0.1
Traffic volume of a
1
Traffic volume of a2
Traffic volume of a3
Traffic volume of a4
20
Traffic volume of a5
Additional travel cost of a3

Additional travel cost


Traffic volume

15

10

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time

Fig. 13. Trafc volumes of the ve links and the additional travel cost under constant trafc volume control.

Comparison of path flow and associated generalized travel cost


30 5
28 Path flow of p1
Path flow of p2 4.5
26
Path flow of p3
24 4
Travel cost of p
1
22 Travel cost of p
2 3.5
20 Travel cost of p3 Travel cost
Path flow

18 3
16
2.5
14
12 2
10
1.5
8
6 1
4
0.5
2
0 0
0 1 2 3 4 5 6
Time

Fig. 14. Path departure rates and travel costs of the Braess network under time-varying trafc volume control.

multipliers associated with the side constraints over time (unlike the static case). This additional travel cost term also rep-
resents the control parameter that allows the trafc volume control scheme to achieve the required link trafc volumes. If
the side constraints are chosen as the link storage capacities, the additional cost can be viewed as the effect needed to pre-
vent the network from spillback. In this sense, the paper proposes a novel analytical approach to access the DTA considering
the spillback effect while avoiding the drawbacks of physical queue models. Another meaningful implication of imposing the
arc storage capacity constraint is that the ow is incompressible when the link trafc volume is equal to its storage capacity.
This is consistent with the ndings in Ross (1988) where the LWR model was applied to describe the trafc dynamics. The
paper has also highlighted the similarity between the additional delay terms from the static and dynamic cases.
1056 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

Traffic volumes of the five links


25
Traffic volume of a1
Traffic volume of a2
Traffic volume of a3
20 Traffic volume of a
4
Traffic volume of a5
Traffic volume

15

10

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time

Fig. 15. Trafc volumes of the ve links under time-varying trafc volume control.

(a) Comparison of traffic volume of link a and (b) Comparison of traffic volume of link a and
2 3
additional travel cost additional travel cost
0.01 5 0.1
24
Traffic volume of a Traffic volume of a3
2
22 Side constraint of a Side constraint of a3
2
Additional travel cost of a Additional travel cost of a
2
20 4 3

18
Additional travel cost

Additional travel cost


16
Traffic volume

Traffic volume

3
14

12

10 2

6
1
4

0 0 0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6
Time Time

Fig. 16. Link trafc volumes, time-varying side constraints and additional travel costs of the Braess network under the DUE-SC condition.

We propose a solution algorithm for solving the DUE-SC by using the nonlinear programming approach with Eulers dis-
cretization scheme. Numerical examples are presented to illustrate the application of the theory. The numerical results con-
rm the solution algorithms satisfactory performance against the small test networks. The results obtained in the constant
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1057

0
Convergence error of the Braess network
10

DUE
1 DUESC with timevarying side constraints
10 DUESC with constant side constraint

2
10
Convergence error

3
10

4
10

5
10

6
10

7
10
2 4 6 8 10 12 14 16 18 20
Number of iterations

Fig. 17. Change in the convergence error with iteration for the Braess network.

trafc volume control test verify that given a long enough departure time horizon, the inow prole of the controlled link
will converge to a steady state determined by the trafc volume control. This is consistent with the static case (Yang et al.,
2004). Tests are also provided to verify that the proposed method is applicable to the time-varying trafc volume control
scheme, which is one of the major contributions of this paper and is the key difference between the DUE-SC and SC-TAP.
The dynamic user equilibrium with strict capacity constraints, which has been intensively studied in the simulation-based
DTA models, can also be included as a special case of the current DUE-SC framework by specifying the side constraints to be
the link capacities.
To apply the proposed scheme, we will develop a more efcient numerical algorithm to solve the DUE-SC. However,
developing such an algorithm for the DUE-SC or even for the standard DUE is still an open problem. Our next step is to tackle
this challenge but with a trade-off between the tractability and theoretical property of the algorithm, e.g. extending the heu-
ristic solution algorithms of DUE to solve the DUE-SC (Tong and Wong, 2010).

Acknowledgments

The work described in this paper was jointly supported by two research grants from the Research Grant Council of the
Hong Kong Special Administration Region to the Hong Kong Polytechnic University (Project Numbers: PolyU 5261/07E
and 5195/07E) and an internal research grant J-BB7Q from the Research Committee of the Hong Kong Polytechnic University.
The rst author would like to thank the Hong Kong Polytechnic University for providing his Ph.D studentship. The authors
want to thank the referees for the careful review as well as the very interesting and constructive comments, which provided
insights that allowed a substantial improvement of this paper.

Appendix A. Constraint qualications for the DUE/DUE-SC

In this appendix, we verify that the mixed constraints (42) (45) and the pure state constraints (46) and (47) satisfy some
constraint qualications (CQs). To this end, we rst introduce the following notions and results for the analysis of CQs.

Denition A.1. The mixed constraints Ki x; u; t 6 0 for i = 1, . . . , q satisfy the Arrow constraint qualication if

@Ki x; u; t
rank jItj;
@uj i2It

for every admissible pair (x, u) where


It fi : Ki x; u; t 0g;
x and u denote the state and control variables respectively, and jI(t)j denotes the cardinality of the set I.
This CQ means that the gradients with respect to the control variables of all the active mixed constraints must be linearly
independent.
1058 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

As claimed by Hartl et al. (1995), compared with the mixed constraints, the pure state constraints are, in general, more
difcult to deal with since they do not explicitly depend on the controls. But note that the state variables can be controlled
indirectly via the state space equations, i.e. (39) and (40). It is therefore convenient to differentiate the pure state constraints
with respect to time t as many times as required until it contains a control variable. A pure state constraint is called a con-
straint of the rth order if the rth time derivative of this pure state constraint is the rst time in which a term in control var-
iable(s) appears. It is easy for us to show that the pure state constraints (46) and (47) are of the 1st order. For example, by
dxa t dC a t dxa t
evaluating the derivative of xai t  C ai t 6 0 with respect to time once, we obtain dti  dti . But dti depends on the con-
trol variables explicitly. Therefore, the constraint is of the 1st order. The same reasoning can be applied to other pure state
constraints to conclude that all of the pure state constraints are of the 1st order. In fact, the pure state constraints are of the
1st order for many economic optimal control problems (Adida and Perakis, 2007,Hartl et al., 1995). Let us denote the pure
state constraints as -i(x(t), t), i = 1, . . . , s and their rst order derivatives with respect to time as -1i xt; t; i 1; . . . ; s.

Denition A.2. For the pure state constraints of the 1st order, the CQ can be veried by the following full rank condition on
any boundary interval [11, 12]6 (Adida and Perakis, 2007,Hartl et al., 1995):
2 @ -1 3
1

6 @u 7
6 7 ^
rank6 ... 7 b;
4 5
@ -1^
b
@u

^ 6 s, - x t; t < 0; i b
for t 2 [11, 12], -i x t; t 0; i 1; . . . ; b ^ 1; . . . ; s, where x(t) denotes the optimal state
i
trajectory.
As the above CQs are veried with respect to the control variables, it is essential for us to clarify the control variables
 p 
adopted
 p in the proof of Proposition  3.3. These control variables are h = (hp:p 2 P), g g ai : p 2 1; jPj; i 2 1; mp , and
g~ g~ai : p 2 1; jPj; i 2 1; mp dened by (51). For the mixed constraints (42) and (43), we may isolate the state-depen-
dent and control-dependent terms by writing

dA1 x
g a1 t A1 x 1 hp t; 8p 2 P;
dt

dAi x
g pai t Ai x 1 g pai1 t; 8p 2 P; i 2 2; mp:
dt
Note further that a very important characteristic of the necessary conditions for time-shifted problems is that the variables
g pai and g~pai g pai t Ai x for all relevant subscripts and superscripts are treated independently in expressing the minimum
principle. We assume that all the mixed constraints (42) (45) are binding; if the CQ holds for this case, then it will hold for
any subset of binding constraints. We rearrange the mixed constraints according to control variables specied previously as
 hp 6 0; 8p 2 P; 95
 g pai t 6 0; 8p 2 P; i 2 1; mp; 96
g~pa1  1 A_ 1  hp t 0; 8p 2 P; 97
g~ai  1 A
p _ i  g p t 0; 8p 2 P; i 2 2; mp: 98
ai1

Two matrices are important in the proof of the Jacobian matrix Jm of the binding mixed constraints, which are both diagonal
matrices of appropriate dimensions.
2 _ i;p 3
2
1 0  0 0
3 1A 1
0  0 0
6 _ i;p 7
6
6 1    0 0 7
7
6
6 1A 2
 0 0 7
7
6 .. .. ..7 6 .. .. .. 7
Jm 6
i
6 . . .7;
7 Jb 6
i
6 . . .
7:
7
6 7 6 7
4 1 0 5 6 _ i;p
1A 0 7
4 jPj1 5
1 _ i;p
1A jPj

The subscript of A_ i;p means the ith link of path pj and remind that jPj is the number of paths. By Theorem 5.1 of Zhu and
j

Marcotte (2000), when the linear link travel time function is adopted, we have 1 A _ i;p > 0. These matrices are important,
j
h i
take (95) for example, the nonzero entry of column col(hp) is @h@ p hp 1; 8p 2 P, or J 0m in matrix form if we ignore
jPjjPj
the other zero entries, i.e. the partial derivatives with respect to g and g~. Similar reasoning can be applied to (96). For (97), we
have the nonzero entry for column col(hp) as

6
An interval [11, 12]  [0, T] with 11 < 12 is called a boundary interval if -i(x(t), t) = 0 for t 2 [11, 12].
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1059

@  p   
_ 1  hp 1:
g~a1  1 A
@hp
h i  
Again we can express this in a matrix of the form J 0m . Similarly, we denote the matrix induced by column col g pai as J im .
jPjjPj
 
Note that there is another nonzero entry for column col g~pa1 , i.e.

@  p   
_ 1  hp 1 A
_ 1;p :
p g~a1  1 A
@ g~a1

We represent this in a matrix form as


2 _ 1;p 3
1A 1
0  0 0
6 _ 1;p 7
6
6 1A 2
 0 0 7
7
6 .. .. .. 7
Jb 6
1
6 . . .
7;
7
6 7
6 _ 1;p
1A 0 7
4 jPj1 5
_ 1;p
1A jPj

_ 1;p means the rst link of path pj 2 P. Similarly, we can conduct the analysis on (98). The Jacobian matrix Jm
the subscript of A j

of the binding mixed constraints can be obtained as


0 1
J 0m
B C
B J 1m C
B C
B J 2m C
B C
B C
B .. C
B . C
B C
B C
Jm B J mp
m C:
B C
B J0 J 1b C
B m C
B C
B J 1m J 2b C
B C
B C
B .. .. C
@ . . A
mp
J mp1
m Jb

By simple calculation, we can eliminate the non-zero entries in the lower triangular matrix such that Jm is similar to the fol-
lowing block diagonal matrix, i.e.
0 1
J 0m
B C
B J 1m C
B C
B J 2m C
B C
B C
B .. C
B . C
B C
B C
Jm
B J mp
m C;
B C
B J 1b C
B C
B C
B J 2b C
B C
B C
B .. C
@ . A
mp
Jb

which is in a special Jordan normal form. Therefore, the Jacobian matrix Jm of the binding mixed constraints is of full rank, see
e.g. Roger and Johnson (1985), the Arrow CQ is satised.
Next we move on to the pure state constraints. Note that the constraint C ai t and its time derivative do not depend on the
control variables but the time derivative of xai t does. Regarding to (46), one can proceed with the analysis similar to that of
the mixed constraints. However, there is a simpler way for us to verify the CQ. Note that (46) is imposed to ensure nonneg-
ative link volumes. In line with the side constraint (47), we may represent it as xai t 6 0. We may assume all the pure state
constraints are binding. Again this is unnecessary, as we can easily show that the constraints xai t 6 0 and
xai t  C ai t 6 0 will not be activated simultaneously for any link. To see this, if C ai t > 0, the two constraints cannot hold
simultaneously is obviously. If C ai t 0, i.e. the link is closed, these two constraints simply reduce to xai t 0 if they are
binding. Therefore, we have at most n binding pure state constraints hold simultaneously. Therefore, it is enough for us to
assume there are m 6 n empty links, i.e. xai t 0, and the other n  m links are binding with xai t  C ai t 0. We have only
two basic elements of the partial derivatives of the pure state constraints with respect to the control variables. Denote the
1060 R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061

 
control vector for path p 2 P as up col hp ; g pa1 ; . . . ; g pan , where we have ignored g~pa1 ; . . . ; g
~pan since the rst order time deriv-
atives of the pure state constraints do not depend on them explicitly, we have
X 
x_ ai g pai dpai  g pai1 dpai1 :
p2P

For the empty links, we have


@
x_ ai 0 0    0 1 1 0    0 :
@up
For the saturated links, we have
@
x_ a 0 0    0 1 1 0    0 :
@up i
For path p, in accordance with the binding constraints, we reorder these links as a01 ; . . . ; a0m and a0m1 ; . . . ; a0n , while the control
variables are rearranged accordingly. We can represent the above partial derivatives in a matrix form as
2 3
1 1
6 7
6 1 1 7
6 .. .. 7
6 7
6 . . 7
6 7
6 1 1 7
p 6 7
J 6 7 :
6 1 1 7
6 7
6 1 1 7
6 7
6 .. .. 7
6 7
4 . . 5
1 1 nmp1

p
It is obviously that the rankJ = n. But the Jacobian matrix J of the rst time derivatives of pure state constraints with respect
to the control variables is
J Jp       n2n1 jPj ;

where denotes other matrices that can be evaluated similar to Jp, i.e. the matrices induced by evaluating partial derivatives
of the rst time derivatives of pure state constraints with respect to the path control vectors excluding up, and the zero en-
tries contributed by g~pai ; 8p 2 P; i 2 1; mp. Clearly, rankJ = n, which implies the CQ for the binding pure state constraints,
dened by Denition A.2, is satised. To see this, we write the transpose of J as
0 1
1 0 0   0 0 0   0
B 1 1 0   0 0 0   0 C
B C
B C
B 0 1 1   0 0 0   0 C
B C
B 0 0 1   0 0 0   0 C
B C
B .. .. .. .. .. .. .. .. .. C
B C
B . . . . . . . . . C
B C
B 0 0 0   1 0 0   0 C
B C
B C
B 0 0 0   1 1 0   0 C
B C
B 0 0 0   0 1 1   0 C
B C
JT B
B 0 0 0   0 0 1  
C
0 C,C 1 ; C2 ;C 3 ;  ;C m ; Cm1 ; Cm2 ;  ; Cn ;
B C
B .. .. .. .. .. .. .. .. .. C
B . . . . . . . . . C
B C
B 0 0 0   0 0 0   1 C
B C
B C
B 0 0 0   0 0 0   1 C
B C
Br r r   r mp2;m r mp2;m1 r mp2;m2   r mp2;n C
B mp2;1 mp2;2 mp2;3 C
B C
B r mp3;1 rmp3;2 r mp3;3   r mp3;m r mp3;m1 r mp3;m2   r mp3;n C
B C
B .. .. .. .. .. .. .. .. .. C
B C
@ . . . . . . . . . A
r 2n1 jPj;1 r 2n1 jPj;2 r 2n1 jPj;3   r 2n1 jPj;m r 2n1 jPj;m1 r 2n1 jPj;m2   r 2n1 jPj;n

where ri,j denotes the elements of J in . We have


X
n
ri Ci 0 ) ri 0; 8i 2 1; n;
i1

for scalars ri, "i 2 [1, n]. Thus, we have rank JT = n= rank J.
R.X. Zhong et al. / Transportation Research Part B 45 (2011) 10351061 1061

References

Aboudolas, K., Papageorgiou, M., Kosmatopoulos, E., 2009. Store-and-forward based methods for the signal control problem in large-scale congested urban
road networks. Transportation Research Part C 17 (2), 163174.
Adida, E., Perakis, G., 2007. A nonlinear continuous time optimal control model of dynamic pricing and inventory control with no backorders. Naval
Research Logistics 54 (7), 767795.
Astarita, V., 1996. A continuous time link model for dynamic network loading based on travel time function. In: Proceedings of 13th International
Symposium on Transportation and Trafc Theory, pp. 79102.
Ban, X., Liu, X., Ferris, M., Ran, B., 2008. A link-node complementarity model and solution algorithm for dynamic user equilibria with exact ow
propagations. Transportation Research Part B 42 (9), 823842.
Buskens, C., Maurer, H., 2000. SQP-methods for solving optimal control problems with control and state constraints: adjoint variables, sensitivity analysis
and real-time control. Journal of Computational and Applied Mathematics 120 (1), 85108.
Carey, M., McCartney, M., 2002. Behaviour of a whole-link travel time model used in dynamic trafc assignment. Transportation Research Part B 36 (1), 83
95.
Carey, M., Ge, Y., 2003. Comparing whole-link travel time models used in DTA. Transportation Research Part B 37 (10), 905926.
Carey, M., 2008. A framework for user equilibrium dynamic trafc assignment. Journal of the Operational Research Society 60 (3), 395410.
Chow, A., 2009. Dynamic system optimal trafc assignment-a state-dependent control theoretic approach. Transportmetrica 5 (2), 85106.
Friesz, T., Bernstein, D., Smith, T., Tobin, R., Wie, B., 1993. A variational inequality formulation of the dynamic networks user equilibrium problem.
Operations Research 41 (2), 179191.
Friesz, T., Bernstein, D., Suo, Z., Tobin, R., 2001. Dynamic network user equilibrium with state-dependent time lags. Networks and Spatial Economics 1 (34),
319347.
Friesz, T., Mookherjee, R., 2006. Solving the dynamic network user equilibrium problem with state-dependent time shifts. Transportation Research Part B 40
(3), 207229.
Friesz, T., Mookherjee, R., Yao, T., 2008. Securitizing congestion: the congestion call option. Transportation Research Part B 42 (5), 407437.
Friesz, T., 2010. Dynamic Optimization and Differential Games. Springer, New York.
Friesz, T., Kim, T., Kwon, C., Rigdon, M., 2011. Approximate network loading and dual-time-scale dynamic user equilibrium. Transportation Research Part B
45 (1), 176207.
Gentile, G., Meschini, L., Papola, N., 2005. Macroscopic ARC performance models with capacity constraints for within-day dynamic trafc assignment.
Transportation Research Part B 39 (10), 319338.
Hartl, R., Sethi, S., Vickson, R., 1995. A survey of the maximum principles for optimal control problems with state constraints. SIAM Review 37 (2), 181218.
Huang, H., Lam, W., 2002. Modeling and solving the dynamic user equilibrium route and departure time choice problem in network with queues.
Transportation Research Part B 36 (3), 253273.
Larsson, T., Patriksson, M., 1995. An augmented lagrangian dual algorithm for link capacity side constrained trafc assignment problems. Transportation
Research Part B 29 (6), 433455.
Larsson, T., Patriksson, M., 1999. Side constrained trafc equilibrium models-analysis, computation and applications. Transportation Research Part B 33 (4),
233264.
Larsson, T., Patriksson, M., Rydergren, C., 2004. A column generation procedure for the side constrained trafc equilibrium problem. Transportation Research
Part B 38 (1), 1738.
Leoni, G., 2009. A First Course in Sobolev Spaces. American Mathematical Society, Providence.
Lindsey, R., 2004. Existence, uniqueness, and trip cost function properties of user equilibrium in the bottleneck model with multiple user classes.
Transportation Science 38 (3), 293314.
Lo, H.P., Szeto, W.Y., 2002. A cell-based variational inequality formulation of the dynamic user optimal assignment problem. Transportation Research Part B
36 (5), 421443.
Lu, C., Mahmassani, H., Zhou, X., 2009. Equivalent gap function-based reformulation and solution algorithm for the dynamic user equilibrium problem.
Transportation Research Part B 43 (3), 345364.
Minoux, M., 1986. Mathematical Programming: Theory and Algorithms. John Wiley, New York.
Mun, J., 2007. Trafc performance models for dynamic trafc assignment: an assessment of existing models. Transport Reviews 27 (2), 231249.
Nie, X., Zhang, H.M., 2002. Delay-function-based link models: their properties and computational issues. Transportation Research Part B 39 (8), 729751.
Nie, X., Zhang, H.M., 2005. A comparative study of some macroscopic link models used in dynamic trafc assignment. Networks and Spatial Economics 5 (1),
89115.
Nie, Y., Zhang, H.M., 2010. Solving the dynamic user optimal assignment problem considering queue spillback. Networks and Spatial Economics 10 (1), 49
71.
Nie, Y., 2011. A cell-based MerchantNemhauser model for the system optimum dynamic trafc assignment problem. Transportation Research Part B 45 (2),
329342.
Noiri, T., 1974. On weakly continuous mappings. Proceedings of the American Mathematical Society 46 (1), 120124.
Papageorgiou, M., Diakaki, C., Dinopoulou, V., Kotsialos, A., Wang, Y., 2003. Review of road trafc control strategies. Proceedings of the IEEE 96 (12), 2043
2067.
Roger, A., Johnson, C., 1985. Matrix Analysis. Cambridge University Press, New York.
Rose, D., 1984. Weak continuity and strongly closed sets. International Journal of Mathematics and Mathematical Science 7 (4), 809816.
Ross, P., 1988. Trafc dynamics. Transportation Research Part B 22 (6), 421435.
Szeto, W.Y., Lo, H., 2006. Dynamic trafc assignment: properties and extensions. Transportmetrica 2 (1), 3152.
Tong, C., Wong, S., 2010. Heuristic algorithms for simulation-based dynamic trafc assignment. Transportmetrica 6 (2), 97120.
Wie, B., Tobin, R., Carey, M., 2002. The existence, uniqueness and computation of an ARC-based dynamic network user equilibrium formulation.
Transportation Research Part B 36 (10), 897918.
Xu, Y.W., Wu, J.H., Florian, M., Marcotte, P., Zhu, D.L., 1999. Advances in the continuous dynamic network loading problem. Transportation Science 33 (4),
341353.
Yang, H., Meng, Q., 1998. Departure time, route choice and congestion toll in a queuing network with elastic demand. Transportation Research Part B 32 (4),
247260.
Yang, H., Bell, M., Meng, Q., 2000. Modeling the capacity and level of service of urban transportation networks. Transportation Research Part B 34 (4), 255
275.
Yang, H., Meng, Q., Lee, D., 2004. Trial-and-error implementation of marginal-cost pricing on networks in the absence of demand functions. Transportation
Research Part B 38 (6), 477493.
Yang, H., Huang, H., 2005. Mathematical and Economic Theory of Road Pricing. Elsevier, Amsterdam.
Zhu, D., Marcotte, P., 2000. On the existence of solutions to the dynamic user equilibrium problem. Transportation Science 34 (4), 402414.