Transportation Research B: Mike Smith, Wei Huang, Francesco Viti, Chris M.J. Tampère, Hong K. Lo

Transportation Research Part B 122 (2019) 140–166
Contents lists available at ScienceDirect
Transportation Research Part B

journal homepage: www.elsevier.com/locate/trb
Quasi-dynamic traffic assignment with spatial queueing,

control and blocking back
Mike Smith a, Wei Huang b,∗, Francesco Viti c, Chris M.J. Tampère d, Hong K. Lo b
a
Department of Mathematics, University of York, York, United Kingdom
b
Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Hong Kong, China
c
Faculty of Science, Technology and Communication, University of Luxembourg, Luxembourg
d
L-Mob Leuven Mobility Research Center, CIB, KU Leuven, Leuven, Belgium
a r t i c l e i n f o a b s t r a c t
Article history: This paper introduces a steady-state, fixed (or inelastic) demand equilibrium model with
Received 23 March 2018 explicit link-exit capacities, explicit bottleneck or queueing delays and explicit bounds on
Revised 20 January 2019
queue storage capacities. The model is a quasi-dynamic model. The link model at the heart
Accepted 21 January 2019
of this quasi-dynamic equilibrium model is a spatial queueing model, which takes account
of the space taken up by queues both when there is no blocking back and also when
Keywords: there is blocking back. The paper shows that if this quasi-dynamic model is utilised then
Quasi-dynamic traffic assignment for any feasible demand there is an equilibrium solution, provided (i) queue storage ca-
Capacity-constrained equilibrium pacities are large or (ii) prices are used to help impose capacity restrictions; the prices
Queueing either remove queueing delays entirely or just reduce spatial queues sufficiently to ensure
Blocking back that blocking back does not occur at equilibrium. Similar results, but now involving the P0
Merge control
control policy (introduced in Smith (1979a, 1987)) and two new variations of this policy
(i.e., the spatial P0 control policy, and the biased spatial P0 control policy) are obtained. In
these results, the control policies allow green-times to vary in response to prices as well
as spatial queueing delays. These three policies are also tested on a small simple network.
In these tests, the biased spatial version of P0 is much the best in reducing equilibrium
delays (on this simple network). The paper further illustrates how the spatial queueing
model works on simple networks with different merge models; it is demonstrated that
equilibrium may be prevented by certain (fixed ratio) merge models. It is also shown in
this case that equilibrium may be imposed on just the controlled area itself by a variety of
(merge model, gating strategy) combinations. Opportunities for developing such combined
gating and merging control strategies are finally discussed.
© 2019 Elsevier Ltd. All rights reserved.
1. Introduction
Developing dynamic traffic control policies able to respond to, or even anticipate, traffic state variations caused by
changes in control settings is of great importance, especially to manage congestion propagation on the network (Taale, 2008).
Demand responses may be analysed in a day-to-day horizon to simulate e.g. learning from past experiences and from in-
formation collected over the days, or in a within-day framework to account for short-term demand responses during a
day, triggered by e.g. real-time information and guidance systems. An accurate representation of traffic flow and congestion
∗
Corresponding author.
E-mail addresses: 152mikesmith@gmail.com (M. Smith), weihuang@ust.hk (W. Huang), francesco.viti@uni.lu (F. Viti), chris.tampere@kuleuven.be (C.M.J.
Tampère), cehklo@ust.hk (H.K. Lo).
https://doi.org/10.1016/j.trb.2019.01.018
0191-2615/© 2019 Elsevier Ltd. All rights reserved.
M. Smith, W. Huang and F. Viti et al. / Transportation Research Part B 122 (2019) 140–166 141
Fig. 1. A simple two route network; the only bottleneck is at node B. The maximum throughput at node B is to be s vehicles per second. Route 2 has a
very large capacity but is longer than route 1.
dynamics is necessary within such applications. Unfortunately, Dynamic Traffic Assignment (DTA) models are still compu-
tationally time consuming, hence their application in real time may be problematic. A quick approximation to an optimal
control, created by a simpler but fast model, may be more useful than a slower but more accurate approximation.
Dynamic traffic control models should identify, in principle within a short time, a desirable future (flow, control) state,
which may be used to define future control settings and eventually to provide route guidance. This state is often identified
with the steady state equilibrium determined by the “optimal” control settings according to a chosen objective function
(total travel time spent, total delay, network throughput, etc.). This stationary equilibrium state serves in this context as an
“attractor” for control and route guidance, which would be optimal if all conditions would remain in the same stationary
conditions. Under varying conditions, the stationary point serves as an optimization direction rather than an end state, hence
is used as a “target” state for changing signal settings dynamically. The steady state control optimization may for instance
act as an attractor for control implemented sequentially in a rolling-horizon context, such as model predictive control (Hegyi
et al., 2005; Lin, 2011; Rinaldi et al., 2013; Hajiahmadi et al., 2016) or iterative learning control (Huang et al., 2016, 2017).
Dynamic and steady state equilibrium models dealing with congested networks require an explicit consideration of road
capacities and queueing. These factors give rise to complex patterns of interacting demands, flows and queues (and prices
if these are considered); this means that modelling (1) the spatial extent of queues within links and (2) blocking back
propagation of these queues from link to link are both of substantial importance. Unfortunately, these phenomena are not
well represented within the simple link performance functions utilised in steady state equilibrium models. On the other
hand, the current simple steady state cost/delay formulations do presently guarantee the existence, uniqueness and stability
of equilibrium, properties that, for a number of reasons, are desirable in a control optimisation framework.
Attempts to relax some of the current modelling simplifications without increasing modelling complexity have been
made by, e.g., Thompson and Payne (1975), who formulated a rigid demand or inelastic traffic equilibrium model with
capacity constraints and considered queueing delays as independent variables (not given by a cost flow function). Many
have followed, e.g. Smith (1984a,b, 1987), Larsson and Patriksson (1995), Lam and Zhang (20 0 0), Marcotte et al. (2003),
Nie et al. (2004), Gentile et al. (2005) and Bundschuh et al. (2006), and Smith (2013). Regarding queue storage capacities
and queue spillback, examples of capacity-constrained equilibrium models with queueing that explicitly consider the impact
of blocking back are very much less numerous. Notable exceptions are in Daganzo (1998), Gentile et al. (2007), Bliemer
et al. (2012) and Bliemer and Raadsen (2017). Despite these works, what is still absent is a sound comprehensive modelling
methodology for dealing with capacity constraints, queueing delays and blocking back within a traffic assignment model,
explicitly considering control effects and ideally with guaranteed convergence. One of the contributions of this paper is to
make progress towards this unifying approach.
1.1. A motivational example
In this subsection we use the network illustrated in Fig. 1 to motivate the approach taken in this paper.
Imagine that the Fig. 1 network is empty at time zero but that at time zero a steady flow of D vehicles per second
leaves the origin for the destination. The uncongested travel time along route 1 is less than that along route 2 and we
suppose that initially all D vehicles per second use route 1. If D > s, as soon as the steady flow reaches node B a queue
will start to grow on the link approaching node B. Making simple assumptions, the queue volume upstream of node B will
grow at a rate which is somewhat larger than D-s (vehicles per second). So the queueing delay on the approach to node
B will also grow; until it is quicker for some traffic to choose the uncongested but longer route 2. Further, assuming that
travellers are well informed and act in their own best interests, a flow of D-s vehicles per second will now choose route
2 and s vehicles per second will choose route 1. Also the queueing delays will stay constant. This distribution of flows is,
under natural conditions, a stable Wardrop equilibrium with queueing delays; we call this a quasi-dynamic equilibrium since
queues upstream of capacity constraints are explicitly considered. The equilibrium will persist for as long as the outflow D
persists; and may be regarded as a reasonable steady state description of the network performance. In this paper (i) we
focus on this stable Wardrop equilibrium with queueing delays; and (ii) we do not examine closely the initial transient flow
adjustments leading to that equilibrium state.
We show how the approach adopted in this simple example may be applied to general networks, often by building upon
the papers listed above. The paper designs a general approach which includes (a) a simple way of combining queueing delays
and free-flow travel times, allowing for the effects of spatial queueing, (b) traffic control, and (c) road pricing. The effects
142 M. Smith, W. Huang and F. Viti et al. / Transportation Research Part B 122 (2019) 140–166
of blocking back and mechanisms for avoiding blocking back are also considered in the general approach. The approach
adopted allows for new variations of the “capacity-maximising” P0 control policy (without and with prices) to be studied
and shows that variations which take account of the spatial queueing perform better than the original P0 control on a simple
network.
The paper also considers how different merge models may be studied within the overall approach put forward here; it
is demonstrated that certain merge models may be inconsistent with a Wardrop equilibrium with queueing delays.
1.2. Context and contribution
In this paper, we account for the ideal long-term effects of space taken up by traffic queues and spillback onto nodes
within the equilibrium model, which requires only the explicit calculation of steady state conditions, and not its transient
evolution. Our approach is inspired by previous quasi-dynamic equilibrium formulations (e.g. Daganzo, 1998; Bliemer et al.,
2012). Such models may be thought of as between steady state equilibrium and fully dynamic equilibrium models: steady
state models usually have no explicit queues, whereas quasi-dynamic models do, and dynamic models usually have a de-
mand, which varies with time, which quasi-dynamic models do not. Compared to Bliemer et al. (2012), our quasi-dynamic
approach is further simplified, in that these authors reconstruct the queue state at the end of a chosen analysis period
(e.g. 1 h). In our case, we assume stationary conditions where the analysis period does not actually matter. More recently,
Bliemer et al. (2014) presented a model with (vertical) point queues that for the first time includes a generic node model,
originally developed by Tampère et al. (2011), such that it yields more realistic turning capacities. This model is further
extended to take account of storage constraints and hence capture queues that spill back onto upstream links (Bliemer and
Raadsen, 2017). In general, our proposed approach aims to formulate a rigorous mathematical problem with well-known
equilibrium solution properties, while incorporating the more realistic network effects (queueing and blocking back) caused
by traffic congestion. To properly analyse the impact of blocking back, we adopted relatively simple node models, and in par-
ticular merge models where different merging policies can be applied to derive turning fractions. We showcase through a
single destination network that equilibrium may not be guaranteed under specific merging policies as stationary equilibrium
queues cannot be calculated.
The spatial queueing quasi-dynamic model discussed in this paper fits within an equilibrium model very similar to that
in Smith (1987). The paper proves an existence result which uses generalisations of the Beckmann et al. (1956) objective
function, extending Smith (1987). The approach generalises Beckmann et al. (1956) in two aspects (i) there are capacity
constraints which may be exactly attained and also (ii) the allocation of green-times is involved (and not only traffic flows).
(Neither are considered by Beckmann et al. (1956)). We intend to provide rigorous analysis using generalisations of the
Beckmann expression in part because this approach may lead to fast computation of both flows and green-times since
Beckmann-like integrals (a) allow standard optimisation procedures to be applied and (b) have been used for decades to
design fast ways of calculating equilibria; see for example Larsson and Patriksson (1995) and Bar Gera (2002, 2010). In this
paper, however, we focus on investigating theoretical properties of the spatial link model and the equilibrium solutions,
while we leave the efficient solution algorithms and general network applications to future research.
It is worth mentioning that Nesterov and de Palma (2003) considered stationary solutions to the dynamic assignment
problem with the standard Beckmann expression in a similar way as the models developed in this paper. However, they do
not consider the spatial extent of queues within links and blocking back effects to upstream links. In addition, their model
has not been integrated in a traffic assignment and control problem. These modelling features are a main contribution of
this paper.
To summarise, the main contributions of this paper are as follows:
1. We introduce a spatial queueing link model that estimates link delays while considering the space taken up by queues.
2. We develop a novel modelling methodology for dealing with capacity constraints, queueing delays and blocking back
within a traffic assignment model. The modelling framework is further integrated in a traffic assignment and control
problem.
3. We prove existence of equilibrium results that use generalisations of the Beckmann objective function. Rigorous analyses
are provided under a variety of scenarios, considering (1) point queues and spatial queues, and (2) the P0 control policy
proposed in Smith (1979a, 1987) and its variations.
4. We showcase in a simple two-route network four equilibrium scenarios to compare different control policies.
5. Moreover, we illustrate the properties of the spatial queueing model and demonstrate the existence and non-existence
of equilibria in a single destination network.
6. Different merge models and their impacts on the equilibrium results are finally analysed, including an idea of combining
control with gating.
The rest of the paper is outlined as follows: in Section 2, we formulate the new spatial queueing link model and prove
the existence of equilibrium solutions under a variety of scenarios considering different control policies and queueing con-
ditions. Four equilibrium performance tests are elaborated in a simple example. The model is then extended to consider
blocking back effects in Section 3. Section 4 illustrates the existence and non-existence of equilibria with a hypothetical
small network. Different merge models and their impacts on the equilibrium results are also analysed. Concluding remarks
are provided in Section 5.
Fig. 2. A single link model representing a single traffic lane.
2. A fixed demand capacitated traffic assignment model with spatial queues
The route-choice principle adopted throughout this paper follows the familiar Wardrop (1952) notion: for each origin-
destination pair, more costly routes are not used. This Wardrop equilibrium traffic assignment principle may also be ex-
pressed using link flows and costs (as in for example Smith (1979b) and more recently He et al. (2010)). In this paper we
utilise a link flow formulation of “Wardrop equilibrium”; this formulation is given in detail later. More realistic probabilistic
approaches may be employed to account for stochastic route cost perception. Having the belief that such approach would
not have an impact on the main findings of this study, we leave this exercise to future research.
All the equilibrium models described in this paper utilise link flows, explicit link-exit capacities, and explicit queues.
The Thompson and Payne (1975) queueing model is utilised to represent queues in steady state assignment models. The
whole equilibrium model may then be expressed as a variational inequality following Smith (1987). A multi-level framework
(with one base network and several upper networks) building upon that suggested by Charnes and Cooper (1961) and
Aashtiani and Magnanti (1983) is also utilised. The base level network represents the real road network and each upper
level contains flow between a single origin-destination pair.
The models described in this paper may be thought of as seeking to model the peak of a peak period where queues are
at their maximum, without modelling the building up or the decay of those queues. An alternative approach could consider
queues to grow linearly over a single time-slice from zero (or a past queue state) and that future decisions are made based
on the average delays over that time slice. This alternative approach is only briefly discussed in this paper and is left for
further studies.
To handle modelling complexity in a gradual way, first we suppose that the queueing capacity of all links is very large
so that blocking back does not occur and later we include storage capacity and blocking back in the model.
2.1. A basic link model with a spatial representation of queues and a finite queue storage capacity
As is usual in traffic modelling, each real-life traffic lane is here represented by

• a node which represents the entry point of the lane,
• a node which represents the exit point or the stop line of the lane, and
• a directed link joining these two nodes, which represents the stretch of lane between the entry and the exit of the lane.
A graphical representation of link i is shown in Fig. 2. The flow along link i is to be vi vehicles per second; the saturation
flow at the exit of link i is si vehicles per second; the queue originating at the exit of link i is Qi vehicles; the maximum
possible value of Qi is MaxQi , which is fixed for a link i; and the cost (e.g. travel time) to traverse the entire length of link i
(when the queue Qi = 0 and the flow is vi ) is ci (vi ). The link i “state” may be thought of as (vi , Qi ) and this link state 2-vector
is to be confined to a set of supply-feasible pairs as follows:
vi ≤ si ,
and
Qi ≤ MaxQi
Throughout the rest of the paper, each link cost function ci (.) is to be positive, continuous, non-decreasing and defined
for all non-negative vi ≤ si . ci (vi ) constant is allowed.
Consider no congestion back-propagating from any downstream link, so that the saturation flow si at the link i exit is the
only constraint on the link flow vi . To calculate the queueing or bottleneck delay bi (expressed in e.g. seconds per vehicle)
at the link i exit, it has often been proposed that (see for example Thompson and Payne, 1975; and Smith, 1987):
Qi
bi = (1)
si
By adopting this formula, we choose to evaluate the delay experienced by the Qi -th vehicle in the queue when the queue
has reached Qi (vehicles). If we consider queues to dynamically move towards the equilibrium state, an average value of the
Q
delay should be used. In the case of queues growing linearly from a zero state, Eq. (1) can be replaced by bi = 1/2 s i . In this
i
case, all following equations should be adapted accordingly.
The whole time of travel along link i has often been written (see again Thompson and Payne, 1975; Smith, 1987):
ci ( vi ) + bi (2)
Suppose now that Qi = 0. In this case (using (1)), bi = 0, and formula (2) is entirely reasonable, i.e. the travel time would
simply consist of the traversing time of the whole link ci (vi ). Consider now the case where 0 < Qi < MaxQi . In this case,
the queue on link i covers part of its length and only the remainder has to be traversed (with no queue). The queue will, in
this case, take up space on link i and so formula (2) in general overestimates the travel time for a link; the overestimation
is greater the larger the queue is and the error is maximum when the queue reaches the upstream node. The time for
traversing link i should thus be the sum of a non-queueing component and a queueing component, denoted as ti . (We
suppose that the head of the queue on link i is always adjacent to the link exit.)
We suppose here that the total travel time ti spent in traversing link i depends only on vi and bi . To reduce or eliminate
the double counting described above, here we choose to specify the sum of the queueing and non-queueing travel-time
components as follows:
ti ( vi , bi ) = hi ci ( vi ) + bi (3)
The “shrinkage factor” hi (≤1) here is to be chosen so as to take some careful but simple account of the fact that as the
queueing delay grows, the unqueued link length (and hence the unqueued traversing time) shrinks.
The hi factor may here naturally be estimated by considering various values of Qi when vi = si in (3). Note that deter-
mining hi is trivial for smaller values of vi , since in this section (where there is by assumption no blocking back) there is no
queueing and no queueing delay when vi < si , so in this case hi = 1.
Consider now the case where vi = si , so that the link flow equals the saturation flow rate. Then (3) becomes:
ti ( si , bi ) = hi ci ( si ) + bi (4)
A natural expression for hi is a linear function of the queue state, i.e. (1 − Qi /MaxQi ) = (MaxQi − Qi )/MaxQi . This expression
is adopted in this study:
MaxQi − Qi
hi =
MaxQi
This is certainly correct if Qi = MaxQi for then hi = 0 and all the time spent traversing link i using (4) becomes entirely
queueing time bi , and is also correct if Qi = 0 for then hi = 1 and all the time spent traversing link i becomes entirely running
time ci (si ).
Q Q
Substituting hi = 1 − Maxi Q into (4) and exploiting again the fact that s i = bi , we obtain
i i
Qi

ti ( si , bi ) = 1 − ci ( si ) + bi
MaxQi
Qi
= ci ( si ) − ci ( si ) + bi
MaxQi
bi si
= ci ( si ) − ci ( si ) + bi
MaxQi

si ci ( si )
= ci ( si ) + 1− bi
MaxQi
Thus, this value of hi in (4) leads to the following equations:
ti ( si , bi ) = ci ( si ) + ki bi (5a)
where
si ci ( si )
ki = ki ( si ) = 1 − . (5b)
MaxQi
In this equivalent formulation, the shrinkage factor ki in (5a) given in (5b) becomes independent of Qi , unlike hi . (This in
turn may help in various ways with the design of solution algorithms.)
We now may naturally generalise (5b) so that the formula (5a) applies also when the (saturated) link has a green-time
proportion gi at the exit (as well as the exit saturation flow si ). The corresponding formula is:
si gi ci ( si gi )
ki = ki ( si gi ) = 1 − . (5c)
MaxQi
Q
The derivation of this utilises s gi = bi . Finally, the formula for ki may be altered so that it applies also when the link is
i i
not saturated (when vi < si gi ) perhaps because blocking back restricts the link flow; this further generalisation is:
vi ci ( vi )
ki = ki ( vi ) = 1 − . (5d)
MaxQi
Q
The derivation of this utilises v i = bi .
i
There are corresponding changes to (5a). These are as follows. Going with (5d) we have
ti ( vi , bi ) = ci ( vi ) + ki ( vi )bi (5e)
and going with (5c) we have

ti ( si gi , b i ) = ci ( si gi ) + ki ( si gi ) b i . (5f)
A feasibility constraint arises here from the form and sense of, for example, (5a) and (5b). This is:
si ci ( si )
0 < ki = 1 − < 1 or (6a)
MaxQi
MaxQi
0 < ci ( si ) < . (6b)
si
Condition (6a) ensures that 0 < ki < 1, which is common sense. Condition (6b), which is equivalent, is also itself reason-
able: for non-decreasing cost functions it says that the time taken to traverse link i when there is no queue must be smaller
than the maximum queueing delay.
The proposed analytical queueing delay formula can be flexible to describe the blocked vehicles (queue) that cannot
proceed in the considered time period. In this paper, to first focus on theoretical properties of the model, the queue length
Qi and hence the queueing delays bi in this paper are assumed to be independent variables; they are not given by a cost-flow
function or by a loading procedure, but are instead determined by equilibrium conditions involving equilibrium routeing just
like in Smith (1987). Moreover, in this study we do not consider transients leading up to or away from that equilibrium. Of
course transient behaviour and stability are important. We leave these extensions to further research, where we expect our
spatial queueing model to have interesting applications.
2.2. Upper multi-level network representation, and connection to the base network
We model the network by considering several upper level networks and one base level network; there is one upper level
network for each (origin-destination pair, vehicle class) = (od, κ ) combination and the single base level network represents
the real road network with lanes, junctions and traffic signals. Each upper network must here be a copy of the single base
network. Total link flows, queues and certain travel times / costs will all be represented on the base network; link costs /
travel times are computed using the equations introduced in Section 2.1 above.
In this paper we suppose that for each upper network the origin destination demand is fixed (inelastic) and is given in
terms of passenger car units or pcus. The flow along any particular base network link is then obtained by simply adding the
flows (in pcu terms) along all upper network links corresponding to that particular base network link. We suppose in this
paper that each pcu has the same impact on queueing space and on the exit flow capacity of each base network link.
Here we also assume that, while upper network flows are added to obtain base network flows, the time needed to
traverse an upper network link is defined to be the time of travel along the corresponding base network link (possibly
adding a constant which depends on the upper network link). Thus, flows on the upper network links add to yield flows
on each base network and link and costs / travel times on each base network link fix the costs / travel times on all upper
network links.
A route is defined to be any sequence of distinct upper level links which are contiguous. The entries Lir in the matrix L
are defined as follows.

1, if upper level route r traverses a link which corresponds to base link i
Lir =
0, otherwise
NOTE: The links in a route are distinct: so a route does not traverse any link more than once. Hence, there are a finite
number of routes (say N) and each route flow vector X belongs to RN+.
2.2.1. Supply side conditions and supply feasibility

Suppose that the base network has m links. Throughout we suppose that there are given positive, non-decreasing cost
functions ci (• ), defined for all link flows vi such that 0 ≤ vi ≤ si . Then the vector
c(v ) = [c1 (v1 ), c2 (v2 ), ..., ci (vi ), ...cm (vm )]
is defined for all v belonging to the set S = [0, s] of supply-feasible base link flows, where
S = [0, s] = {v = [v1 , v2 , ..., vi , ..., vm ]; 0 ≤ vi ≤ si for 1 ≤ i ≤ m} (7a)
We also define Sa, the augmented set of supply feasible base link flows, by:
Sa = {v = [v1 , v2 , ..., vi , ..., vm ]; vi ≤ si for 1 ≤ i ≤ m} = {v; v ≤ s} (7b)
In (7b) there are no lower bounds. In real life, of course, the set of supply-feasible base-link flows will then be the set
of non-negative vectors belonging to Sa. The lower bounds of zero in S given by (7a) are omitted from Sa given by (7b);
they are not actually needed as non-negativity will be assured in any case by the specification of the demand feasible sets
(defined below). Moreover there is an advantage in leaving the lower bounds out; this is that all normals to the set Sa (at
any point) must then have the form n = [n1 , n2 ,…, ni ,…, nm ] where for each link i the co-ordinate ni ≥ 0. This is not true
for S. Normals (and variational inequalities) will be utilized for the analyses on equilibrium, which is to be elaborated in
Section 2.3 below.
Base link flow vector v arises from the upper route flow vector X by putting:

vi = r
Lir Xr for 1 ≤ i ≤ m or v = LX
Given the set Sa of augmented supply feasible base-network link flow vectors there is a corresponding set of augmented
supply feasible upper-level route flow vectors. This is defined as follows.
SaU = {X; LX belongs to Sa = {v; v ≤ s}} = {X; LX ≤ s} (7c)
The ci (vi ) in turn yield, for each X ∈ SaU such that X ≥ 0 and each upper level route r, the route r traversing time:

Cr (X ) = LT ci ( (LX )i )
i ri
For each upper-level route flow vector X, these route travel times Cr (X) may be written as an N-vector C(X) where:
C(X ) = LT c(LX )
Since travellers also experience bottleneck delays bi as discussed in Section 2.1 above, the route r travel times (or costs)
become:

tr ( X, b ) = LTri [ci ( (LX )i ) + ki bi ] = Cr (X ) + LTri [ki bi ] (8a)
i i
where ki is given by (5b).

Eq. (8a) can be formulated in vectorial form. If k and b are m-vectors then the Schur or Hadamard product k ◦ b is the
m-vector whose ith co-ordinate is ki bi . Using this notation for all X ∈ SaU and any bottleneck delay vector b ≥ 0, (8a) may
be written:
t(X, b ) = LT [c(LX ) + k ◦ b] = C(X ) + LT [k ◦ b] (8b)
where k is given by Eq. (5b).
2.2.2. Demand feasibility

κ representing the
We suppose initially here that, for each OD pair od and vehicle class κ , there is a positive number dod
flow of traffic of vehicle type or class κ from upper level node o to upper level node d. A demand feasible route flow vector
X must satisfy: X ≥ 0 and for each upper level origin-destination pair od and each κ the total of the flows of class κ vehicles
κ .
along all routes joining o to d = dod
Given all the demands dodκ the upper level demand set DU is thus defined by:

κ =
DU = X ≥ 0; for each upper level network dω Xr (9a)
r∈Rκω
Here r ∈ Rκod means that route r carries vehicles of only class κ , starts at node o and ends at node d. The demand flows dod
κ
of vehicle class κ from o to d will be measured in passenger car units or pcus.
Given an upper demand set DU defined by (9a) the base network demand set D is then specified (in pcus) by:
D = L(DU ) = {LX; X belongs to DU } (9b)
2.2.3. Feasibility
The set of feasible upper level route flows is then defined to be DU ∩ SaU where DU is given by (9a) and SaU is given by
(7c). The set of feasible base level link flows is then defined to be D ∩ Sa where Sa is given by (7b) and D is given by (9b).
(Throughout this paper all feasible variables will in fact be non-negative. Thus: Xr ≥ 0 for all r, and bi ≥ 0, vi ≥ 0 and Qi ≥ 0
for all i.)
We suppose throughout this paper that DU ∩ SaU is non-empty and then of course D ∩ Sa is non-empty. This means that
the problems or conditions considered in this paper are feasible: there is an X belonging to both the demand-feasible set
DU and the supply feasible set SU or there is an X in DU ∩ SU = DU ∩ SaU .
2.3. Wardrop equilibrium (X, b) with bottleneck delays (without blocking back)
We initially suppose there is no blocking back. Consider an upper network route-flow vector X together with a base
network bottleneck delay vector b ≥ 0 say (X, b). This pair (X, b) will here be called a “Wardrop equilibrium with bottleneck
delays” if
(i) X meets all the given demands dod κ or belongs to DU ,
(ii) X is also supply feasible or belongs to SaU and

(iii) for each single upper level network: for each pair of routes which have different costs (e.g., travel times) the more costly
route carries no flow. Here travel times or costs are determined by (8a, b) so that spatial queueing is represented here in
the base level link travel times (but not blocking back). (10a)
To state the Wardrop equilibrium condition (10a) above in a different way, as a variational inequality (see Smith (1979b),
Dafermos (1980)), we need the definition of normal.
Definition. The vector u is normal at v to the set D if and only if u • (w – v) ≤ 0 for all w ∈ D.
(This condition is illustrated in Fig. 7 in Appendix 1.)

Suppose now that the link cost vector c is a function of v defined for all v in D and consider the above normality
condition when u = −c. This normality condition (with u = −c) leads to the following Variational Inequality (VI):
v ∈ D and − c(v ) · (w − v ) ≤ 0 for all w ∈ D (10b)

In this particular VI (a) link flow vectors v and w are utilized and (b) it is supposed that c(v) is defined for all v in D.
Now suppose that the cost function c is defined only at points in a set of supply-feasible flows S, and that S does not
contain the whole of the demand set D. This often happens in real life because there are capacity constraints on the network.
In order to deal with this case where links have active capacity restrictions it is now natural to “expand” the above VI
(10b) by considering also a vector b of bottleneck delays which arise at capacitated link exits. To fit in with the link travel
time functions in (8a) and (8b) and to fit the definition (10a) of a Wardrop equilibrium with bottleneck delays, we now
choose to consider the following expanded version of the above VI (10b):
v belongs to D ∩ Sa, b ≥ 0 and

−[c(v ) + k ◦ b] · (w − v ) ≤ 0 for all w ∈ D or (10c)
−[c(v ) + k ◦ b] is normal at v to D
The bottom line of (10c) uses the concept of normal (as illustrated in Fig. 7 in Appendix 1), where now u = u(v,
b) = −[c(v) + k◦b]. Here we think of each co-ordinate ki being given by (5b). Suppose now that X belongs to DU and that
LX = v. By following the argument in Smith (1979b) it follows that the above VI (10c), which is in link flow terms, is equiv-
alent to:
X belongs to DU ∩ SaU , b ≥ 0 and
(10d)
−[C(X ) + LT (k ◦ b )] is normal at X to DU
which is in route flow terms.
Arguments in Smith (1979b) complete a circle here by showing that (10d) holds if and only (X, b) is a Wardrop equilib-
rium with queues as defined in (10a). Thus, we have, if v = LX, that (10c) is equivalent to (10d), which in turn is equivalent
to (10a).
Here k is determined by Eq. (5b) and travel times are given by (8a) and (8b); so spatial queueing is represented, but
blocking back is not (yet) represented.
Clearly, not all queueing delay vectors b can realistically arise; it is reasonable to impose the condition that if a link is
unsaturated then there is no queueing delay (in a steady state model; assuming no blocking back). To restrict attention to
only such feasible bottleneck delay vectors we impose the following condition, which may be thought of as a “queueing
equilibrium” condition on vi and bi : for each base network link i, either
[the flow vi = r Lir Xr = si and the bottleneck delay bi ≥ 0] or

[the flow vi = r Lir Xr < si and bi = 0] (so there is no queueing delay at the exit of link i if vi < si ).
This condition may be written as a complementarity condition:
for all links i : [vi = si and bi ≥ 0]or[vi < si andbi = 0]. (11)
Condition (11) may be simply stated as follows: the flow vi is supply-feasible (vi ≤ si ) and the non-negative queueing delay
must be zero if vi < si .
Thus, queues are assumed to only occur on saturated links. It may be argued that some queueing delay occurs also in
steady-state conditions (usually) when a link exit is nearly saturated Viti and Van Zuylen, 2010). This can be allowed for
in our approach by thinking of si as a non-decreasing function of bi (the inverse of a strictly increasing bottleneck delay
function) and then writing si (bi ) in place of the constant si in (11). We do not do this here because we wish to keep things
as uncomplicated as possible.
By virtue of the definition of Sa in (7b) above, condition (11) ensures that bi ≥ 0. Since all ki > 0 condition (11) holds for
all links i if and only if
k ◦ b is normal at v to Sa (12)
where Sa is given in (7b) above. (This is a variational inequality way of saying that a positive queueing delay bi (or a positive
ki bi ) can only happen at equilibrium if the link exit is saturated or vi = si . Comparing (12) and (10b), in (10b) v is pushed
by – c(v) as travellers seek to reduce costs and in (12) v is pushed by any positive k ◦ b so that k ◦ b is more normal at v to
Sa.)
In this section, to ensure that blocking back does not occur, we add the sufficient queue storage condition:
s ◦ b < MaxQ where MaxQ = [MaxQ1 , MaxQ2 , ..., MaxQm ]

Combining this “sufficient queue storage” condition, the feasible Wardrop equilibrium condition (10c) and the “queueing de-
lay occurs only on saturated links” or queue equilibrium condition (11) or (12) we obtain the following condition on (v, b):
v belongs to D ∩ Sa, s ◦ b < MaxQ
−[c(v ) + k ◦ b] is normal at v to D and also (13a)
k ◦ b is normal at v to Sa
Here bi = Qi / si for 1 ≤ i ≤ m, and ki is given by (5b). This condition (13a) is our basic definition of a Wardrop equilibrium
with spatial queueing delays, also in equilibrium, in a steady state capacitated network model. Condition (13a) does not
address blocking back: the assumption here is that MaxQ is a large vector. Here, the only effect of the upper limit on Qi
(= MaxQi ) is in the definition of the constants ki (given in (5b)) in the link cost functions given in (5a). These introduce
a spatial queueing element to the corresponding vertical queueing quasi-dynamic equilibrium condition; which may be
obtained by letting k = 1 (the m vector where each entry = 1) in (13a).
2.4. Proof of existence of an equilibrium with prices (or point queues) at link exits, and then with spatial queueing
In this paper we consider prices as well as travel times. To be as simple as possible we suppose that all travellers have
the same value of time and that prices are in time units. So we essentially assume that time and price are both measured
in (say) seconds.
Suppose given the sets D and Sa defined in (9b) and (7b), and the continuous positive non-decreasing functions ci , so
that c(.) is defined throughout D ∩ Sa. Let v be an m-vector of link flows and p be an m-vector of prices at the exits of all m
links and consider the following condition on (v, p):
v belongs to D ∩ Sa
−[c(v ) + p] is normal at v to D and also (13b)
p is normal at v to Sa
(This condition is exactly condition (13a) if, in (13b), p is replaced by k ◦ b and the condition “s◦b < MaxQ” is added.) In
fact if (13b) has a solution (v∗ , p∗ ) and we put
bi ∗ = pi ∗ /ki for i = 1, 2, 3, . . . , m
then (v∗ , b∗ ) solves (13a) if MaxQ is large enough. We need this “large enough” condition now as we are here extending
the pricing result to a spatial queueing model and not just a pricing, or a point queue, model. (Plainly, showing that (13b)
has a solution does not yet imply that (13a) has a solution.) Theorem 1 below shows that (13b) has a solution under natural
conditions.
Theorem 1. Suppose given the (closed, bounded and convex) set D ∩ Sa defined above in (9b) and (7b). Suppose:
(i) that D ∩ Sa is non-empty and

(ii) that each ci (.) is continuous, positive, non-decreasing and defined throughout [0, si ].
Then there exists a pair (v, p) satisfying (13b).
Proof. The interested reader can find the proof in Appendix 2.
Theorem 1 may be utilised to prove a number of “similar” results. We give several of these results here as corollaries of
Theorem 1 since they are obtained fairly simply by re-interpreting the variables and constraint sets of Theorem 1 appropri-
ately.
Corollary 1.1. (Existence of an equilibrium with spatial queues when MaxQ is large.) Under conditions similar to those given in
Theorem 1, there exists a pair (v, b) satisfying (13a), provided MaxQ is large enough.
Proof. Under the conditions of Theorem 1, (13b) has a solution (v∗ , p∗ ), by Theorem 1. So, if we put
bi ∗ = pi ∗ /ki for i = 1, 2, 3, . . . , m
then (v∗ , b∗ ) satisfies (13a) aside from the constraint s◦b∗ ≤ MaxQ. (If this last constraint is not satisfied then the bottleneck
delay vector b∗ may not be supportable as the implied queue vector s◦b∗ may have some co-ordinates which are too large
for the link storage capacity). To ensure that this is not the case here; we utilise the vector b∗ just defined and impose the
following obvious constraint on the queueing capacity vector MaxQ:
MaxQ > s ◦ b∗ .
For any MaxQ satisfying this inequality it follows that (13a) holds with feasible queue lengths and the corollary is proven.
Corollary 1.2. (Existence of an equilibrium with both spatial queues and prices where the spatial queues do not block back.)
Under conditions similar to those given in Theorem 1, and given any positive vector MaxQ; there exists a triple (v, b, p) with
b > 0 satisfying:
v belongs to D ∩ Sa, s ◦ b < MaxQ
−[c(v ) + p + k ◦ b] is normal at v to D and also
p + k ◦ b is normal at v to Sa
Remark. This corollary shows that a pricing vector p may be utilised to ensure that the equilibrium bottleneck delay vector
b satisfies the queue volume constraint Q =s◦b < MaxQ for any given MaxQ > 0; here prices just reduce equilibrium queues
enough to obviate blocking back, rather than driving these queues to zero as in Theorem 1.
2.5. Equilibrium when there are green-time proportions and a pricing variant of control policy P0 is utilised
In this section, we choose to utilise a control policy called P0 to change green-time proportions in light of traffic flows
or bottleneck delays or queues. This policy, in its exact form, at each junction equalises the pressure on each stage at that
junction, where the pressure on a stage is the sum over the links in that stage of the terms s1 b1 where s1 is the saturation
flow at the exit of link i and b1 is the bottleneck delay felt at the exit of link i.
This policy is introduced in Smith (1979a, b, 1987) in a cost-flow context and a vertical queueing context. It is shown
there that if P0 is utilised then, under natural conditions (which preclude blocking back), there is a Wardrop equilibrium
consistent with the policy. Smith et al. (2015) show (in a cost-flow context) that, if travellers swap to quicker routes and
green-time proportions swap toward more pressurised stages, then the dynamical disequilibrium trajectory of flows and
green-time proportions cannot leave the supply-feasible set and travel times are bounded. It is to be hoped that such good
results obtained in other contexts may also be obtained in this new quasi-dynamic spatial queueing context.
To start, here, we first utilise the pricing version of P0 where prices take over the role of the bottleneck delays in the
above P0 policy specifications. We suppose that there is no queueing but that instead there is a price, and, as previously, a
green-time proportion corresponding to each link exit and that these prices and green-time proportions may be controlled
responsively as flows vary. So now a closed bounded convex set F of feasible green-time proportion vectors g is supposed
given; and we consider the new demand set D × F (where D is the set of demand-feasible link flows defined previously).
A new green-time proportion-expanded set S^ of augmented supply-feasible (link flow, link green-time proportion) vec-
tors is then defined as follows:
S ∧ = { ( v, g ) ; v − s ◦ g ≤ 0 }.
With these changes new versions of (13b) naturally arise, and these involve green-time proportions as well as traffic flows.
One of these uses a new function h (which may be chosen by the network manager) and is:
(v, g ) belongs to (D × F ) ∩ S∧
− [(c(v ), −h(g ) ) + (p, −s ◦ p )] = [−(c(v ) + p ), h(g ) + s ◦ p )] is normal at (v, g ) to D × F and also (14)
(p, −s ◦ p ) is normal at (v, g ) to S∧ .
Comparing (14) with (13b) line by line it may be seen that:
in the first line (v, g) extends v, and now involves green-time proportions as well as flows,
in the second line (p, −s◦p) extends p and − [(c(v), −h(g)) + (p, −s◦p)] extends – [(c(v), p] and
in the third line (p, −s◦p) extends p again.
This is hence a pricing variant of a biased version of the P0 control policy. This biased version we now call Ph . The new
“bias” function h(g) in (14) is here to be constructed from new continuous non-increasing functions hi of link green-time
proportions gi as follows:
h(g ) = [h1 (g1 ), h2 (g2 ), h3 (g3 ), ..., hi (gi ), ..., hm (gm )].
In this new version of (13b) the green-time proportion vector g is pushed, within the green-time proportion feasible set
F, in the direction h(g) + s◦p this is part of the middle line in (14). Such actions are here thought of as being within the
control of a network manager / controller. An exact form of this h-biased pricing variant of the P0 control policy may be
expressed more fully and accurately as follows:
for a given price vector p, the green-time proportion vector g is to be chosen so that the “force”
s◦p + h(g) is normal at g to F.
This condition is part of the second line of condition (14). This h-biased version of the P0 policy is, in more dynamical
terms, as follows:
continually move g in the direction of the “force” h(g) + s◦p projected onto the set F of feasible link green-time propor-
tion vectors.
(Of course, it is also imagined that the network controller will control prices as well; increasing prices where there are
queues, aiming to reduce queues.)
Following this pricing variant of the h-biased P0 signal control policy, a link i with a high saturation flow and a high
price will push up the link i green-time proportion. Eventually this will tend to increase the total of the prices paid (subject
to the condition that prices on unsaturated links must be zero); under suitable conditions the signals become throughput-
maximising and profit-maximising green-time allocators.
To show that there is an equilibrium consistent with this h-biased pricing variant of P0 (embracing green-time propor-
tions, flows and prices) we need to show that (14) has a solution. This is proven in Theorem 2 below.
Theorem 2. Suppose that (D × F) ∩ S∧ is non-empty, that
c(v ) = [c1 (v1 ), c2 (v2 ), c3 (v3 ), . . ., ci (vi ), . . ., cm (vm )]
and
h(g ) = [h1 (g1 ), h2 (g2 ), h3 (g3 ), . . ., hi (gi ), . . ., hm (gm )].

Suppose that each ci (vi ) is continuous and non-decreasing and that each hi (gi ) is continuous and non-increasing, that each ci (•)
is defined for all vi satisfying 0 ≤ vi ≤ si and that each hi (•) is defined for all gi satisfying 0 ≤ gi ≤ 1. Then there exists a triple (v,
g, p) with (v, g) in S^ and p ≥ 0 satisfying (14). (The conditions here may be weakened significantly without losing the existence
result.)
Proof. This follows from Theorem 1 above by considering the new augmented Beckmann objective function

vi
gi
Za(v, g ) = ci (u )du − hi (u )du (15)
0 0
i i
defined for each (v, g) ∈ (D × F) ∩ S∧ . Using exactly the same arguments as those utilised in the proof of Theorem 1 leads to
a proof. An outline proof is given in Appendix 3.
To solve (14) it is natural to approximately minimise (15), over (v, g) belonging to (D × F) ∩ S∧ . This will yield an approx-
imation to (v∗ , g∗ ) where
−grad Z(v∗ , g∗ ) = (−c(v∗ ), h(g∗ ) )is normal at (v∗ , g∗ ) to (D × F ) ∩ S∧ .

Then an nS^ may be estimated by choosing p to approximately minimise the norm of the projection
Proj(D×F )∩S∧ (−c(v∗ ) − p, h(g∗ ) + s ◦ p )

of ( − c(v∗ ) − p, h(g∗ ) + s◦p) onto (D × F) ∩ S∧ . Such a p must have non-negative co-ordinates by virtue of the design of S^.
Finally, using the ki given by (5b), find b∗ such that k◦b∗ = p; using the notation introduced above, b∗ = p/k. If (v∗ , g∗ ) is
an exact minimiser there will be a p such that Proj(D×F )∩S∧ (−c(v∗ ) − p, h(g∗ ) + s ◦ p ) = (0, 0 ). If (v∗ , g∗ ) is not exact then
provided the norm of Proj(D×F )∩S∧ (−c(v∗ ) − p, h(g∗ ) + s ◦ p ) is sufficiently small moving v toward any cheaper routes or
moving g towards any more pressurised signal stages will reduce Za.
2.6. Equilibrium when there are green-time proportions and a spatial queueing variant of biased control policy P0 is utilised
In this section we seek to apply our previous “pricing” existence results to the link models we have defined previously
which describe spatial queueing.
To do this it is natural to now replace the price vector p above by the spatial queueing delay vector k◦b just as we did
above in Corollary 1. Then (14) becomes:
(v, g ) belongs to (D × F ) ∩ S∧
[−(c(v ) + k ◦ b ), h(g ) + s ◦ (k ◦ b )] is normal at (v, g ) to (D × F ) and also (16)
((k ◦ b ), −s ◦ (k ◦ b )) is normal at (v, g ) to S∧
This is a new equilibrium condition with spatial queueing delays consistent with a new spatial queueing variant of the
h-biased P0 policy.
This policy is called Phk and is as follows: for any observed bottleneck delay vector b:

choose feasible green − time proportions so that the “ force s ◦ (k ◦ b ) + h(g ) is normal at g to F . (17)
Condition (17) is part of the second line of (16). A more dynamic statement of this policy is as follows:
continually move g in the direction of the force s ◦ (k ◦ b ) + h(g ) projected onto F . (18)
Here the drivers of vehicles determine k ◦ b, and not the network controller since queueing delays arise naturally on sat-
urated links. Corollary 2 below shows that there is an equilibrium consistent with this augmented spatial queueing variant,
(17), of the h-biased P0 policy.
Corollary 2.1. Suppose that (D × F) ∩ S∧ is non-empty, that

and
is defined for all vi satisfying 0 ≤ vi ≤ si and that each hi (•) is defined for all gi satisfying 0 ≤ gi ≤ 1; as previously. Suppose
also that each ki > 0.
Then, provided MaxQ is sufficiently large, there exists a triple (v, g, b) satisfying (16).
Proof. This is given in Appendix 4.

and
is defined for all vi satisfying 0 ≤ vi ≤ si and that each hi (•) is defined for all gi satisfying 0 ≤ gi ≤ 1; as previously. Suppose also
that MaxQ is a given positive vector and that each ki > 0.
Then there exists a quadruple (v, g, b, p) with b > 0 satisfying:
(v, g ) belongs to (D × F ) ∩ S∧ , s ◦ b < MaxQ,
[−(c(v ) + p + k ◦ b ), h(g ) + s ◦ (p + k ◦ b )] is normal at (v, g ) to D × F and also
(p + k ◦ b, −s ◦ (p + k ◦ b )) is normal at (v, g ) to S∧ .
Remark. This corollary shows that a pricing vector p may be utilised to ensure that the equilibrium bottleneck delay vector
b satisfies the queue volume constraint Q = s◦b < MaxQ for any given MaxQ > 0; here again prices just reduce equilibrium
queues enough to obviate blocking back, rather than driving these queues to zero as in Theorem 2.
Q
Let k(s ◦ g ) ◦ sQ◦g be the vector whose ith co-ordinate is ki (si gi ) s gi . Then k(s ◦ g ) ◦ Q
g = s ◦ k (s ◦ g ) ◦ Q
s◦g = s ◦ k ( s ◦ g ) ◦ b.
i i
Eq. (17) may be written more explicitly as follows: for any given queue size vector Q:
Q
choose a green − time proportion in F so that the force k(s ◦ g ) ◦ + h(g ) is normal at g to F . (19)
g
This condition is still essentially part of the second line of condition (16). A more dynamic statement of this policy (19) is
as follows:
Q
continually move g in F in the direction of the force k(s ◦ g ) ◦ + h(g ) projected onto F . (20)
g
Of course, on a real network, the bottleneck delays arise from drivers’ decisions and (unlike prices) are not directly control-
lable by a network manager; only the green-time proportion vector is directly controllable here.
There are two aspects which merit observation here. First, it is notable that, if we replace s◦g by the flow vector v, then
the saturation flows in both (19) and (20) disappear from the policy statement. This augmented queueing version of the P0
policy then applies even if saturation flows are unknown or changed; a sudden snow fall leaves the policy unchanged, or
certain approaches to a junction being blocked by an accident leaves the policy unchanged. This may allow the signals to
respond quickly to sudden saturation flow changes or flow changes, and so the policy may be thought of as automatically
taking some (fast) account of such changes. A model running in the background might help here to change green-time
proportions and provide route guidance fast. This would add to the responsiveness of the whole system to a sudden change
on the network.
Q
Second, when there are green-time proportions the formula (5b) becomes (5c). For fixed green-times the terms ki (si gi ) g i
i
Q
in k(s ◦ g ) ◦ Qg in (5c) with smaller MaxQi have smaller ki (si gi ) (see (5c)) and so the (positive) force ki (si gi ) g i on gi is also
i
smaller. Thus, the green-times on short narrow links (those with smaller MaxQi and so smaller ki ) will now tend to be
smaller at equilibrium compared to the green-times on long wide links. Thus, now, for given queues, short narrow links
will have higher delays than with the original P0 policy where maximum queue sizes were not involved in the policy
specification. Thus, with the Phk policy, there will now be a greater tendency to discourage vehicles from using shorter but
narrower links.
Fig. 3. A simple two route signal-controlled network; link 2 is wider and longer than link 1.
2.7. A simple example
It is interesting to compare the above control policies in a simple example; to do this we utilise the network in Fig. 3.
Suppose that the network is in a quasi-dynamic equilibrium state with spatial queues. Let the freeflow travel time from the
signal to the destination be zero seconds. Let
si = the saturation flow at the link exit (in vehicles per second: i = 1, 2);
Ci = the free flow cost/time of travel via route i (in seconds; constant: i = 1, 2);
bi = the bottleneck delay felt at the link i exit (in seconds per vehicle: i = 1, 2);
T = the steady total inflow at the origin, heading for the destination (v/s);
MaxQi = the maximum queue possible on link i = 1, 2.
Here we also suppose that

siCi
s1 < s2 , C2 = C1 + , and = θ for i = 1, 2;
MaxQi
where > 0 and 0 < θ < 1. Here, with the above suppositions, (5c) becomes:
si gi ci ( si gi ) si giCi
ki = ki ( si gi ) = 1 − =1− = 1 − θ gi
MaxQi MaxQi
We now compare four spatial equilibrium scenarios corresponding to four control systems for values of T satisfying
0 < T < s2 .
To assess the four control options, we use “the extra travel time E” where
E = [the travel time at equilibrium minus C1 ] (s).
This will be the performance measure for each of the four control/pricing systems. E will depend importantly on the
constant origin-destination flow T (v/s), = C2 − C1 > 0, as well as on the control policies.
2.7.1. The four equilibrium performance tests and the two link models considered
The following four performance tests will be considered. Here ki = 1 − θ gi for i = 1, 2.
• Test 1. Performance of the P0 control policy (choose g so that s1 b1 = s2 b2 ) with vertical queueing;
• Test 2. Performance of spatial P0 , P0k (choose g so that k1 s1 b1 = k2 s2 b2 ) with spatial queueing;
• Test 3. Performance of P0 (choose g so that s1 b1 = s2 b2 ) with spatial queueing;
• Test 4. Performance of biased spatial P0 , Phk (choose g so that k1 s1 b1 + h1 = k2 s2 b2 + h2 ); with spatial queueing.
Here we will consider a general k and also will especially consider the vector k given by (5c), with constant ci ; and general
and special vectors h will be considered too. The special vectors depend on the network in Fig. 3.
The following two link performance models will be considered:
Vertical queueing: In this case we suppose that the total time of traversing link i, ti (bi ) = Ci + bi ; and
Spatial queueing: In this case we suppose that the total time of traversing link i, ti (bi ) = Ci + ki bi , where ki = 1 − θ gi . (This
is an application to this example of (5a) and (5b), including green-time proportions, but with constant ci )
We now consider the four control tests 1 – 4 individually. In all these tests we only consider T < s2 . We also utilise the
variables r and a, where r = s1 /s2 and a = 1 − θ .
2.7.1.1. Equilibrium performance of the P0 control policy, with vertical queues.

At equilibrium with vertical queueing: C1 + b1 = C2 + b2 = C1 + + b2 .
Thus at equilibrium with vertical queueing: b1 = b2 + .
The P0 policy ensures that: s1 b1 = s2 b2 .
Therefore: s1 b2 + s1 = s1 b1 = s2 b2
and so: s1 = (s2 - s1 )b2
or: = [(s2 /s1 ) – 1]b2
and so: b2 = /[(s2 /s1 ) – 1].
Fig. 4. Graph showing how E(T) varies with T in tests 2, 3 and 4 when T increases from 0 nearly to s2 . In all three tests 2, 3 and 4 the link model is the
“spatial queueing” model (with ki = 1 − θ gi and s1 < (1−θ )s2 ). For all three policies (P0 , P0k and Phk ) E = 0 when T < s1 and E jumps upwards at T = s1 ; the
jump is greatest when the policy is P0 and is minimised by Phk , the biased spatial version of P0 . The dotted increasing curve gives /(1−r/a) as a function
of s1 and is asymptotic to the vertical line originating at (1−θ )s2 ; this dotted curve gives the height of the jump in the equilibrium delay when P0 is
utilised (and the spatial queueing model is utilised).
So, in this case, if we put r = s1 /s2 :
E = [the travel time at equilibrium minus C1 ] = C2 + b2 – C1 = b2 + = r/[1 – r] + = /(1 − r).
This formula only applies when r = s1 /s2 < 1. The question of what happens if r = 1 will be discussed briefly at the end of
this section.
2.7.1.2. Equilibrium performance of spatial P0 , P0k with spatial queueing. Here ki = 1 − θ gi for i = 1, 2.
Similar arguments (more details in Appendix 5) show that, in this case
E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 – C1 = k2 b2 + = /(1-r).
Again, this formula only applies when r < 1.
2.7.1.3. Equilibrium performance of P0 with spatial queueing. Here ki = 1 − θ gi for i = 1, 2.

In this case a similar argument (given in the Appendix 5) shows that now
E = [the travel time at equilibrium minus C1 ] = + .

1 . 1 −θ g 1 −1
r 1 −θ g
2
In this case the value of E depends on the green-time vector g which depends on T. The graph showing the variation of
E as T varies is illustrated in Fig. 4; three particular values of E are shown as dots in Fig. 4. The values taken when T = s1 ,
T = ½(s1 + s2 ) and T = s2 ; where of course s1 < ½(s1 + s2 ) < s2 ; are

E ( s1 ) = > E ( 1/2 ( s1 + s2 ) ) = > E ( s2 ) =
1 − r/a 1−r 1 − ra
Thus, as Fig. 4 shows, here E(T) is a decreasing function of T; see Appendix 5 for more detail.
2.7.1.4. Equilibrium performance of biased spatial P0 , Phk , with spatial queueing. Here ki = 1 − θ gi .
Again a similar argument shows that in this case
E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 – C1 = k2 b2 = [ - (h2 – h1 )/s1 ] / [(s2 /s1 ) – 1] + =
if we choose, for example, h = [s1 C1 , s1 C2 ].

2.7.2. A graphical performance comparison

Now we compare the performance measure E = E(T) in tests 2, 3, 4 graphically, as T increases from 0 to s2 . In this figure,
T < s2 and s1 < (1-θ )s2 . So here r = s1 /s2 < 1-θ where θ > 0.
The bold line in Fig. 4 shows that if the original responsive P0 policy is utilised when spatial queueing occurs then the
extra delay E will, at equilibrium when T is just > s1 , be /(1-r/a); which may be much greater than /(1-r) (s) (which is the
value of E with vertical queueing when T is just > s1 ; see Test 1). The bold line in Fig. 4 shows also that, when T > s1 , E(T)
decreases with T. This decreasing behaviour of E(T) may seem counterintuitive; it is the result of the three-way interaction
between (i) the responsive P0 policy, (ii) routing decisions by drivers, and (iii) the spatial queueing model adopted. The
horizontal dotted lines in Fig. 4 also show how performance is improved by switching to the spatial version of P0 , P0 k , and
how performance is further improved by switching to the biased spatial version of P0 , namely Phk , with suitable h, k.
In the analysis here it is assumed, as shown in Fig. 4, that
s1 < (1 − θ )s2 < s2.
The consideration of the case when s1 > (1−θ ) s2 is reserved for further study. The discussion in this section and espe-
cially our redesigned Fig. 4 is a response to a perceptive query by an anonymous referee concerning the case s1 = s2 .
3. A fixed demand capacitated traffic assignment model with blocking back
3.1. The link model with blocking back
In Section 2, we introduced a basic equilibrium with queues, taking some account of the spatial extent of queues within
the link. This condition makes sense so long as the equilibrium queue lengths are within the queueing capacities of the
links where the queue starts. We proved that a solution of the model certainly exists provided these queueing capacities
MaxQi are large enough.
Now we consider what happens if the bound MaxQi is exceeded. In this case we must allow for the possibility that a
link outflow is restricted by downstream queues filling a downstream link and overflowing onto the link.
Suppose that the flow along link i, vi , is constrained not by the exit saturation flow of link i but by a queue on a
downstream link exceeding the queue capacity of that downstream link and spilling back into link i. The maximum possible
number of queueing vehicles on link i is unchanged at MaxQi and so now the maximum bottleneck delay on link i, when
MaxQ
the flow is vi , is v i . This is the maximum possible bottleneck or queueing delay, so that if vi is very small the maximum
i
queueing delay can be very large with a given queue storage capacity.
We utilise Eq. (3) (as we did in Section 2) but now the flow is constrained to be less than si by a blocking back queue,
and we do not yet know how much it is; hence we assume the flow is vi (also yet unknown). In this case we obtain
vi ci ( vi )
ki = ki ( vi ) = 1 −
MaxQi
Thus, ki depends on vi and is constrained by the exit saturation flows of downstream links. Now, with blocking back, a
link model which may prove useful in estimating the sum of the travelling time and the queueing time on link i is:

vi ci ( vi )
ti ( vi , bi ) = ci ( vi ) + ki ( vi )bi = ci ( vi ) + 1− bi (21)
MaxQi
Link model (21) clearly includes expressions (5a) as a special case; of course, in Section 2, queues arise only when vi = si ,
whereas here queues may arise also when vi < si due to blocking back. Eq. (21) gives the set of all those pairs (vi , bi ) which
can occur at a solution of the model. (21) may perhaps be regarded as a rudimentary flow – density – cost relation.
Travel time function (21) here is again (like (5a)) a two-dimensional travel time function: the travel time for link i
depends on both the flow vi and the queueing delay bi . The total link travel time function here, ti (vi , bi ) in (21), is defined
only if
MaxQi
0 ≤ vi ≤ si and 0 ≤ bi ≤
vi
or
0 ≤ vi ≤ si and 0 ≤ vi bi ≤ MaxQi (22)
Constraint (22) is non-convex; this suggests that, even if equilibrium solutions exist, there may be many of them.
A further feasibility constraint apparently arises here from the form and sense of (21). This is:
vi ci (vi ) − MaxQi ≤ 0 for all vi ≤ si
Assuming that vi ci (vi ) is non-decreasing, this is always true if:
si ci (si ) − MaxQi ≤ 0
which is our condition (6); and so this, apparently new, condition is implied by (6), which, as remarked above in Section 2,
is entirely reasonable.
3.1.1. Supply-feasibility with blocking back

The link i flow vi may be constrained by
(i) the link i exit capacity or
(ii) the queue Qj on link j immediately upper-level downstream being MaxQj .
It is necessary now to take account of both possible constraining mechanisms within the definition of “link i supply
feasibility” and the consequent constraints imposed on v and the bottleneck delay vector b. For sake of clarity, let just the
two base links j and k be upper-network-downstream from base link i.
So we are here thinking, to be specific and to limit complexity, of a simple diverge network. Generalisation to multiple
outgoing flows is straightforward.
The set of feasible base link flows is now, with possible blocking back, to be defined as follows. First for each such base
(diverge) link i we put

BBSi = (v, Q ); vi − si ≤ 0 and Q j − MaxQ j ≤ 0 for all links j upper − level − downstream of link i
BBSi is the set of blocking-back supply-feasible (v, Q) pairs associated with the link i; of course, this supply-feasible
set now involves flows and bottleneck delays on downstream links j as well as on link i. Then the set of blocking-back
supply-feasible (v, Q) pairs is to be
BBS = {(v, Q ); (v, Q ) ∈ BBSi for all i} = ∩i BBSi = {(v, Q ); Q − MaxQ ≤ 0 and v − s ≤ 0}
This may be compared with (7b) which lacks the constraint Q – MaxQ ≤ 0.
3.2. Wardrop equilibrium (X, b) with bottleneck delays (and with blocking back)
Taking the two constraining effects in (i) and (ii) in 3.1.1 together (instead of just (i) in Section 2) we now insist that the
following queue-equilibrium condition holds for our link i:

bi ≥ 0 and vi = si or Q j = MaxQ j for some link j upper − level − downstream of link i , or
(23)
(vi < si ) and Q j < MaxQ j for all links upper − level − downstream of link i and (bi = 0 ) .
This is a more extensive condition than the simpler complementarity condition (11) above. Both (11) and (23) are saying
that the queueing delay may be positive if there is a cause; but that if there is no cause then the bottleneck delay is zero.
Non-negativity constraints apply to all variables including the bottleneck delays bi .
It is now natural to give a possible queue state for each possible cause as follows. Let
wii = vi − si ,
wi j = Q j − MaxQ j , and (24)
wik = Qk − MaxQk .
Then the three-level queueing delay equilibrium condition (24) may be written in this case with just links k and j down-
stream:
wii = 0, or [wii < 0 and bii = 0],

wi j = 0, or wi j < 0 and bi j = 0 , and (25)
wik = 0, or[wik < 0 and bik = 0]
Here the wih take on and extend the role of vi – si in (11) but wij depends on the downstream queue Qj .
The Wardrop / queueing equilibrium condition is now, with blocking back:
(v, b )belongs to [D × Rm
+ ] ∩ BBS,
−[c(v ) + k ◦ b] is normal at v to D, (26)

bi = h bih and wih ≤ 0, wih < 0 implies bih = 0, for all (i, h )with h = i or link h downstream of link i.
As is shown already by Daganzo (1998), we cannot hope to obtain the kind of general existence results previously stated.
So, in this paper we do not prove that an equilibrium exists in this case. On the other hand, finding controls (including
prices and signal green-times) and network conditions, which ensure existence of equilibrium in this framework is likely to
prove a profitable, yet arduous, research direction, which we aim start undertaking in the next section.
4. Existence and non-existence of equilibria in an illustrative example network
In this section, we illustrate how link models (5a) and (21), together with simple equilibrium analysis leads, in a simple
network model, to different quasi-dynamic equilibrium steady states. We also illustrate, using these two models, how long
queues back to an origin may happen, and then a quasi-dynamic equilibrium does not exist (even if demand is within the
capacity of the network and there is no merging).
Fig. 5. A network with three capacitated links (2, 3, 4) and two OD pairs [A, B] and [C, B].
We also consider different merging models in this section and show (using these two link models and a merge) that
many well-known merging models lead to the non-existence of quasi-dynamic equilibria (even if demand is within the
capacity of the network). Further, it is shown how the quasi-dynamic modelling framework here may be utilised to design
beneficial combinations of perimeter control, or gating, and merging.
4.1. The example network
Throughout Section 4 we consider the network in Fig. 5; two OD pairs [A, B] and [C, B] are joined by three routes as
follows.
1. Route 1 (the bypass) joins the OD pair [A, B], and has just link 1.
2. Route 2 joins OD pair [A, B], and has links 2, 3 and 0.
3. Route 3 joins OD pair [C, B], and has links 4, 3 and 0.
Link i has a saturation flow at the exit of si vehicles per hour for i = 2, 3 and 4. All other links will be uncapacitated so
that there is no limit to the flow exiting link 1 or link 0. We suppose throughout this section that s3 < s2 so that along route
2 the exit of link 3 must be the main bottleneck as indicated graphically in Fig. 5.
The (fixed or inelastic) AB flow rate will be TAB and the (fixed or inelastic) CB flow rate will be TCB . We will always
assume that
0 ≤ TCB < s3 < TAB .
If TCB > 0 then there is a flow TAB > 0 along link 2 and a flow TCB > 0 along link 4 and these must merge at node M.
If vi is the flow rate along link i (i = 0, 1, 2, 3, 4), we suppose that, with no queueing, the time taken to traverse link i
will be ci (vi ) where ci (vi ) ≥ 0. These cost functions are all to be defined for all flows and are all to be non-decreasing. ci (vi )
constant is allowed. We also suppose that c0 (v0 ) = 0 for all v0 .
4.2. Quasi-dynamic equilibrium with spatial queueing on link 3: the case with no blocking back and no merging traffic
In this first case we assume that there is no traffic flow coming from node C; so we assume that the steady flow from
C to B = TCB = 0. Hence there is no merging traffic. We now impose conditions which ensure that there is, at equilibrium,
no blocking back; so, in this section, there is spatial queueing on (only) link 3. We note that always TAB > s3 ; so not all AB
traffic can traverse link 3 and some AB traffic must, at a steady state (or quasi-dynamic) equilibrium, use route 1. To exclude
the possibility to have blocking back at equilibrium we suppose that (for all flows):
c2 (v2 ) + c3 (v3 ) < c1 (v1 ) < c2 (v2 ) + Maxb3 (27)
MaxQ3
where Maxb3 = s3 .
Since TAB > s3 , at a quasi-dynamic equilibrium route 1 must be used. Now, by (27), route 1 is more costly than route
2 (when there is no queueing) and so, at a quasi-dynamic equilibrium, there must be some flow and also some queueing
delay along route 2 to equilibrate the route choices on this network. Queueing delay can only occur on a saturated link, and
s3 < s2 , so link 3 must be the only saturated link on route 2 and v2 = v3 = s3 . To determine the queue and bottleneck delay
on link 3 at equilibrium here we initially utilise link performance model (5a), assuming that blocking back to M does not
occur. Then, at a quasi-dynamic equilibrium,
v2 = v3 = s3 , v1 + v2 = TAB , and c1 (v1 ) = c2 (v2 ) + c3 (v3 ) + k3 b3
Hence
c1 (TAB − s3 ) − (c2 (s3 ) + c3 (s3 ))
b3 =
k3
s3 c3 ( s3 )
where k3 = 1 − MaxQ3 by (5b) and so
c1 (TAB − s3 ) − (c2 (s3 ) + c3 (s3 ))

b3 = s3 c3 ( s3 )
1− MaxQ3
and
Q3 = b3 s3 .
It is easy to check, for any TAB > s3 , that Q3 < MaxQ3 (provided the condition (27) holds) and hence that a quasi-dynamic
[flow vector, queue vector, delay vector] = [v1 , v2 , v3 , v4 , Q1 , Q2 , Q3 , Q4 , b1 , b2 , b3 , b4 ]
= [TAB − s3 , s3 , s3 , 0, 0, 0, Q3 , 0, 0, 0, b3 , 0].
Here Q3 and b3 are given above. Thus, there is a quasi-dynamic equilibrium with queueing delay on just link 3 (and no
blocking back to M) if the specified conditions (27) hold. It is easy to see that this quasi-dynamic equilibrium triple is
unique for any given TAB > s3 .
4.3. Equilibrium with spatial queueing on links 2 and 3: the case with blocking back but still no merging traffic
In this second case we also assume that there is no merging traffic flow TCB . Suppose also now that
c2 (v2 ) + Maxb3 < c1 (v1 ) < Maxb2 + Maxb3 (28)
for all v1 , v2 . Thus link 1 is longer or more costly to traverse than before, in Section 4.1. Bottleneck delays on both links 2
and 3 will be enough to balance the network routeing since, by (28),
c1 (v1 ) < Maxb2 + Maxb3
However, in this case, the queue generated at the bottleneck at the exit of link 3 must be longer than before and must spill
back onto link 2 since, by (28),
c2 (v2 ) + Maxb3 < c1 (v1 )
The equilibrium queue on link 3 must now be a maximum and now also the queue on link 2 will be such that the total
delay incurred on the two queued links equals the uncongested travel time difference between the two alternative routes
joining A and B. So, now,
MaxQ3
c1 (v1 ) = c2 (s3 ) + k2 b2 + Maxb3 = c2 (s3 ) + k2 b2 +
s3
using (21) rather than (5a). In this case, by (21),
c1 (TAB − s3 ) − c2 (s3 ) − MaxQ3
s3
c1 (TAB − s3 ) − c2 (s3 ) − MaxQ3
s3
b2 = = s3 c2 ( s3 )
k2 1− MaxQ2
Also, then
Q2 = b2 s3
It is easy to check that now Q2 < MaxQ2 when the conditions imposed in this section hold. As in Section 4.2 above, it
follows that for any demand TAB > s3 , here is a quasi-dynamic equilibrium [flow, queue, delay] triple (now with blocking
back) given by:
[v1 = TAB − s3 , v2 = v3 = s3 , Q2 , MaxQ3 , b2 , Maxb3 ]
4.4. Non-existence of equilibrium due to blocking back (with still no merging traffic flow)
In this third case we still assume that there is still no merging traffic flow. But now suppose that link 1 is even longer
than before; here we suppose that, for all v1 ,
c1 (v1 ) > Maxb3 + Maxb2
It is clear that in this case no queueing within the specified queue storage limits can balance the network. This situation is
similar to that described by Daganzo (1998).
The sections here have shown that the model described in this paper is, at least in this simple network case, able to
determine equilibria with queueing both when there is no blocking back and also when there is blocking back. The model
clearly also reveals, with no merging traffic, when blocking back is so severe that queues reach origins in the network and
no quasi-dynamic equilibrium state exists.
The next section illustrates the behaviour and the utility of the model when there is merging traffic and we show that
the model may then be useful for comparing different merge models; and so may be utilised to help design or control
merges in real life.
4.5. Equilibrium with spatial queueing on links 2 and 3, blocking back and unrestricted merging flows
In this first merging case we do not impose a merge control policy; merges here will be unrestricted.
We now make the same assumptions as in 4.3, including the assumption that (28) holds for all flows except that now
we let 0 < TCB < s3 . Since TCB < s3 , this TCB flow does not completely flood the bottleneck on link 3, allowing some room at
the bottleneck for AB traffic, and so it allows us now to consider merging at M. In this case at a quasi-dynamic equilibrium
the flow v2 < s3 , since there must be, at a quasi-dynamic equilibrium, room for the flow v4 = TCB at the bottleneck on link
3. (This TCB flow has no route choice of course and so must traverse link 3.)
At any quasi-dynamic equilibrium both of the OD flows, TCB > 0 and TAB > 0, must be met, so that, by the previous
paragraph,
v1 + v2 = TAB , and v2 = s3 − v4 = s3 − TCB
In this case any quasi-dynamic equilibrium, again using (21), must satisfy

v2 c2 ( v2 ) MaxQ3
c1 (v1 ) = c2 (v2 ) + k2 b2 + Maxb3 = c2 (v2 ) + 1− b2 +
MaxQ2 s3
This equation must hold at a quasi-dynamic equilibrium, independently of any merge model, and implies that:
c1 (TAB − v2 ) − c2 (v2 ) − MaxQ3
s3
b2 =
2 c2 ( v2 )
1 − vMax Q2
and
Q2 = b2 v2
It is now easy to check, for any TAB > s3 , that Q2 < MaxQ2 (provided that the condition (28) holds) and hence that a
quasi-dynamic equilibrium exists and is the following vector:
[flow vector, queue vector, delay vector] = [v1 , v2 , v3 , v4 , Q1 , Q2 , Q3 , Q4 , b1 , b2 , b3 , b4 ]
= [TAB + TCB − s3 , s3 − TCB , s3 , TCB , 0, Q2 , MaxQ3 , Q4 , 0, b2 , Maxb3 , b4].
Here Q2 and b2 are given by the formulae above, Q4 is arbitrary and b4 = Q4 /s4 . Thus, there are (many) quasi-dynamic
equilibria with queueing delay on links 2, 3 and 4 if the specified conditions (28) hold.
4.6. Existence and non-existence of equilibrium with spatial queueing and a merge model
In Section 4.5 above the CB demand TCB = v4 takes a fixed total amount of the saturation flow s3 , thus reducing the
capacity available for v2 . This leads to the many quasi-dynamic equilibria given above in Section 4.5.
If now a merge model is given, then there is an extra condition (connecting link 3 and link 4 conditions) to be satisfied.
It may happen that none of the quasi-dynamic equilibria specified in Section 4.5 satisfies this extra condition.
In this light, we now consider the impact of two different merge models. These are:
merge model (a): the flow rates admitted from links 2 and 4 are adjusted so as to ensure that v2 = v4 , and
merge model (b): the flow rates admitted from links 2 and 4 are adjusted so as to ensure that b2 = b4 .
As shown in Section 4.5: for any merge at M, a quasi-equilibrium must have the form
[TAB + TCB − s3 , s3 − TCB , s3 , TCB , 0, Q2 , MaxQ3 , Q4 , 0, b2 , Maxb3 , b4 ]
So the flow on link 2 must satisfy:
v2 = s3 − TCB
Now merge model (a) implies that
v2 = v4 = TCB
Thus if merge model (a) holds at a quasi-dynamic equilibrium then both these equations above hold and so:
TCB = v2 = s3 − TCB
so that
TCB = s3 /2
Thus, if merge rule (a) is adopted at M, then a quasi-dynamic equilibrium with queueing in this network can only occur
when TCB = s3 / 2. If TCB
= s3 / 2 there is no quasi-dynamic equilibrium consistent with merge model (a).
With merge model (b) this does not happen. In this case the quasi-dynamic equilibria given in Section 4.5 satisfy (b) if
we insist that Q4 = s4 b3 so that b3 = Q4 /s4 = b4 .
Fig. 6. Graphical representation of two merging flows.
Note that the queueing delay on link 4 is not calculated in Section 4.5; because we do not specify a merge model in
Section 4.5 and there is then insufficient information to determine b4 .
So, in this proposed quasi-dynamic modelling framework, there exists a quasi-dynamic equilibrium consistent with merge
model (b). On the other hand, the merge model (a) does (almost always) rule out the existence of a quasi-dynamic equilib-
rium consistent with merge model (a).
4.6.1. General (fixed ratio) merge models

All merge policies for a simple merging node like in Fig. 5 can be represented graphically as in Fig. 6. Looking at the
figure, we can observe that, at equilibrium, just four possible states can occur when two merging links are involved:
A. No equilibrium queue exists at any of the entering links (point A)

B. Both links are queued. Point B is the only one representing this state, and the merge ratio in this case is the slope of the
line connecting B with the origin = v4 / v2 .
C. Only link 2 is queued. This is represented by one possible state C.
D. Only link 4 is queued. This is represented by one possible state D.
In fixed ratio merge models, merge ratios are assumed to be kept constant; the merge ratio p = v4 / v2 might be fixed
according to observed behaviour in real merging situations. Examples of these models in the literature include Daganzo’s
fixed merge model (Daganzo, 1995), where p can take any arbitrary value.
In our network in Fig. 5 the given merging fraction p at M may be represented by any point along the line for which
v2 + v4 = s3 , or inside the triangle delimited by this line and the two axes, as depicted in Fig. 6 (e.g., point A). One instance
of Daganzo’s fixed merge model follows the “zipper” rule introduced as merge model (a) above, which assumes that drivers
give way to the other approach in such a way that vehicles from the two approaches alternate with equal frequency. The
zipper model is obtained by letting the distribution fraction p = 1.
Another merge model that can be applied is the “fair shares” merge model, which assumes that merge behaviour is in
some way dependent on the traffic load of each merging link, or has some relationship with the importance of the roads,
normally represented by their capacity. One particular “fair shares” model is described by Jin (2010); in which p = v4 / v2 .
Another instance is the model of Ni and Leonard (2005), p = s4 / s2 ; so that more capacitated links get more priority.
Further research on the impact of various merge models within this overall framework, using the quasi-dynamic models
here would be natural. In this paper we have started to explore this research direction.
Various merge ratios are suitable for different layouts, different junction geometries and different control policies. In this
section, we show how, using the model developed in this paper, we can relate the demand entering the network with the
queueing delay caused by a specific merge model.
The definition of merging priority p in our case is as follows:
v4
p= (29)
v2
So, if in our model here (29) does not hold then there is no equilibrium consistent with merge model p.
4.6.2. Existence and non-existence of equilibrium with a merge model and no blocking back
Assume now that there is a quasi-dynamic equilibrium on the Fig. 5 network where link 1 is used and link 3 is fully
saturated, but the equilibrium queue is not causing blocking back. In this case no queues are observed at links 2 and 4, but
only at link 3. This corresponds to point A in Fig. 6. Then:

c1 ( v1 ) = c2 ( v2 ) + c3 ( s3 ) + k3 b3
s3
TAB = v1 + v2 = v1 +
p+1
p
TCB = v4 = s3
p+1
s3 c3 ( s3 )
Now k3 = 1 − MaxQ3 by (5b), and Q3 = b3 s3 , so we obtain:
c1 (TAB − s3
p+1
) − (c2 ( p+1
s3
) + c3 (s3 ))
b3 = s3 c3 ( s3 )
(30)
1− MaxQ3
and
Q3 = b3 s3
Our possible equilibrium must satisfy the above equations so plainly this equilibrium can only exist if
p
TCB = s3
p+1
with a fixed ratio merge model, typically, a quasi-dynamic equilibrium does not exist on this network.
4.6.3. Existence and non-existence of equilibrium with a merge model and blocking back
Suppose now that we have a quasi-dynamic equilibrium where Q3 = MaxQ3 and so there is blocking back at M. We must
now consider the 3 cases corresponding to B, C, D in Fig. 6.
Case C. Suppose that at our quasi-dynamic equilibrium link 2 is the only approach queued at M (thus we are at point C
in Fig. 6). Then, using Eq. (21), we obtain:
c1 (v1 ) = c2 (v2 ) + k2 b2 + Maxb3
s3
TAB = v1 + , and
p+1
p
TCB = v4 = s3
p+1
resulting in
c1 (TAB − s3
p+1
) − (c2 ( p+1
s3
)+ MaxQ3
s3
)
b2 = s3 s3 (31)
p+1 c2 ( )
1− MaxQ2
p+1
and
s3
Q2 = b2 v2 = b2
p+1
p
and these equations specify an equilibrium, but only if TCB = p+1 s3 .
Thus with a fixed ratio merge model in case C, typically, a quasi-dynamic equilibrium does not exist.
Case D. Suppose that at our quasi-dynamic equilibrium link 4 is the only approach queued at M (thus we are at point D
in Fig. 6). Then, using Eq. (21), we obtain:
c1 (v1 ) = c2 (v2 ) + Maxb3
s3
TAB = v1 + v2 = v1 +
p+1
p
TCB = v4 = s3
p+1
Thus, just like in case C above, with a fixed ratio merge model in this case D, typically, a quasi-dynamic equilibrium does
not exist.
Case B. Finally if both merging links have queues at equilibrium (and we are at point B in Fig. 6) then:
c1 (v1 ) = c2 (v2 ) + k2 b2 + Maxb3
s3
TAB = v1 + v2 = v1 +
p+1
p
TCB = v4 = s3
p+1
Thus with a fixed ratio merge model in case B, typically, a quasi-dynamic equilibrium does not exist.
4.7. Combining demand (or perimeter) and merging control
In the previous sub-sections we have shown that, given any specific priority p in a fixed ratio merge model, in all four
cases equilibrium does not exist unless the demand TCB satisfies:
p
TCB = v4 = s3 (32)
p+1
Apart from this specific case a quasi-dynamic equilibrium does not exist with any fixed priority merge model, unless we
control also the demand flowing into the controlled area. This can be done by combining the merge model with perimeter
control or gating, which has been proven to be an effective strategy for reducing congestion and preventing blocking back
inside the controlled area (e.g., Keyvan-Ekbatani et al., 2012; Abdoulas and Geroliminis, 2013; Haddad and Zheng, 2018;
Zhong et al., 2018). Eq. (32) can in this case indicate a desirable demand condition for managing a (sub-)network. A pos-
sibility would then be to regulate both the demand entering into the controlled area (e.g. via coordinated ramp metering)
and the merging priorities in such a way that equilibrium is likely to be reached inside the controlled area.
This combined strategy could therefore prevent congestion within the controlled area propagating back to the entrances
of the controlled area, so that the “damage” of blocking back could be contained to a maximum acceptable extent within
the controlled network. Even though in principle these are only conditions for our model to be controllable, we may assume
that they are equally relevant for reality. It seems indeed undesirable if the real traffic system gets uncontrollable under our
two basic modelling assumptions: if drivers tend to user-optimize (i.e. equilibrate) and current demand conditions would
pertain (i.e. stationary demand). Relaxing the latter assumption, it could be even more desirable if an automatic and local
control policy would be able to adapt the priority p parameter such that a wide range of feasible demand conditions could
be handled and equilibrated. In this paper we only briefly present the idea arising from the model developed in this paper,
leaving the elaboration of a general combined demand and merge control policy to future papers; however, the merge rule
(b) suggested above in Section 4.6 does seem like a good starting point.
For example, we can derive analytically the relationship between inflow and priority p if certain targets are to be met in
the example in Fig. 5. Considering only the case of a queue blocking node M and back-propagating onto links 2 and 4, if we
assume desired queue states Q2 ≤ MaxQ2 to avoid spillback onto node A, and no restriction is imposed on Q4 , then using
the value of the equilibrium Q2 given by Eq. (31):
s3 MaxQ3
s
c1 (TAB − )− 3
< MaxQ2
p+1 s3 p+1
This inequality determines a range of possible feasible priority fractions p, given TAB . Each p in this range will prevent
blocking back to the origin A.
5. Conclusions
The central model introduced in this paper is a steady-state, link-based, fixed (or inelastic) demand equilibrium model
with explicit link-exit capacities, explicit bottleneck or queueing delays and explicit bounds on queue storage capacities. The
model is a quasi-dynamic model. The link model at the heart of this equilibrium model takes account of the space taken
up by queues both when there is no blocking back and also when there is blocking back. We have called this link model
a “spatial queueing” model. It has been proved that if the quasi-dynamic model is utilised then for any feasible demand
there is an equilibrium solution, provided prices are used to impose capacity restrictions. The paper has utilised this result
to show that there is also an equilibrium solution with spatial queueing if the queue storage capacities are large; and also
that there is an equilibrium without this condition, provided prices are used to reduce spatial queues so that blocking back
does not occur.
Other similar results, involving the three responsive control policies: P0 , P0k (spatial P0 ) and Phk (biased spatial P0 ) have
been proved. These results include: (1) an existence of equilibrium result if signal green-time proportions respond to prices
using P0 , (2) an existence of equilibrium result if signal green-time proportions respond to spatial queues using the spatial
queueing Phk and (3) an existence of equilibrium result if also prices are used to reduce queue lengths in (2). We have
also tested the P0 control policy, the spatial P0 control policy, and the biased spatial P0 control policy, on a simple network.
These results show that with modest spatial queueing all three P0 – like control policies maximise the capacity of this
simple network; and also that the biased spatial version of P0 is much the best in reducing delays at equilibrium on this
network.
The paper has shown how the spatial queueing model works on a simple network both without and with blocking back
and with a simple merge. Different merges models and their impacts on the equilibrium results have also been studied;
it has been demonstrated that equilibrium may be prevented by certain (fixed ratio) merge models. It is also shown how
in this case gating strategies may be employed to keep the non-equilibrium, growing, queues outside the controlled area;
equilibrium may in this case be imposed on just the controlled area itself by a variety of (merge model, gating strategy)
combinations. Opportunities for developing such combined gating and merging control strategies have been discussed.
Fig. 7. This diagram illustrates a point v in the convex demand set D (the triangle) and a vector u normal, at v, to the set D. This means that, for all w in
D, u•(w – v) ≤ 0 as illustrated for a particular w in D. (Thus standing at v and looking in direction u all of D, including the w illustrated, is behind you.).
Directions for further research include (a) the development of efficient solution algorithms in the steady state and the
dynamic state, (b) the description of flow-dependent queueing delays and queue evolutions over time, (c) merge constraints
(we have just scratched the surface of this here), (d) the ramifications of extensive blocking back, (e) the application of
combined demand and signal control policies on real sized networks, and further, there are large opportunities for (f) con-
sidering how dynamically responsive prices (as well as green times) may be included in a dynamical version of the equi-
librium theory developed here. This paper has outlined several methods of seeking to allow correctly for spatial queueing
without going via dynamic assignment and so the approach is complementary to the approach adopted by Bliemer et al.
(2012, 2014).
Note
This paper is a substantially edited and substantially expanded version of the paper “Equilibrium in Capacitated Network
Models with Queueing Delays, Queue-storage, Blocking Back and Control” presented at the 20th International Symposium of
Traffic and Transport Theory, held in Noordwijk in July 2013 (Smith et al., 2013). While the base link model has not been
modified from the original version, all results in terms of three types of control polies, merging policies and in combina-
tion with gating control are new, and the variational inequality formulations and existence analyses have been revised and
extended.
Acknowledgement
The authors wish to express their thanks to the anonymous referees for their helpful comments on an earlier version
of this paper. The study is partially supported by General Research Fund # 16204518 of the Research Grants Council of the
HKSAR Government.
Appendix 1
This appendix illustrates the definition of normal.

Definition. The vector u is normal at v to the set D if and only if u • (w – v) ≤ 0 for all w ∈ D.
This condition is illustrated in Fig. 7. Standing at v and looking in direction u every point in D is behind you.
Appendix 2
This appendix gives a proof of Theorem 1.
Theorem 1. Suppose given the (closed, bounded and convex) set D ∩ Sa defined above in (9b) and (7b). Suppose:
(i) that D ∩ Sa is non-empty and

(ii) that each ci (.) is continuous, positive, non-decreasing and defined throughout [0, si ].
Then there exists a pair (v, p) satisfying (13b).

Consider the standard Beckmann et al. (1956) objective function Z where, for all v belonging to D ∩ Sa,

vi
Z (v ) = ci (u )du
0
i
The cost functions ci (.) are all continuous, non-decreasing and defined throughout [0, si ]. So Z(v) is defined for all v in
D ∩ Sa. D ∩ Sais closed and bounded, hence it is compact. Also Z is continuous and so Z must have a global minimum over
D ∩ Sa. Let v∗ be any such minimum.
Given v∗ , which minimises Z in the convex set D ∩ Sa, consider any w in D ∩ Sa. Since D ∩ Sa is convex the line [v∗ , w],
which joins v∗ and w, lies in D ∩ Sa. Therefore, since v∗ minimises Z in D ∩ Sa, Z cannot decrease as v moves away from v∗
heading for w in D ∩ Sa along the line [v∗ , w], Hence
−c(v∗ )[w − v∗ ] = −[gradZ (v∗ )][w − v∗] ≤ 0
(– c(v∗ ) is the steepest descent direction for Z at v∗ ) Thus, by the definition of normality above,
−c(v∗ ) is normal at v∗ to D ∩ Sa
Consider any one such Z-minimising v∗ ∈ D ∩ Sa. Sa, defined in (7b), is the intersection of m closed half spaces. (There is
one closed half space for each link in the base network and there are m of these; see (7b)). D, defined in (9b), is also the
intersection of a finite number of closed half spaces; suppose that there are N(D) of these. Thus D ∩ Sa is the intersection of
m + N(D) closed half spaces; suppose that:
D ∩ Sa = [H1 ∩ H2 ∩ H3 ∩ ... ∩ Hm ] ∩ [Hm+1 ∩ Hm+2 ∩ Hm+3 ∩ ... ∩ Hm+N (D ) ]
= H1 ∩ H2 ∩ H3 ∩ ... ∩ Hm ∩ Hm+1 ∩ Hm+2 ∩ Hm+3 ∩ ... ∩ Hm+N (D )
Consider any one of these closed half spaces, Hk say. Then Hk contains D ∩ Sa and so it contains v∗ as v∗ belongs to D ∩ Sa.
Given v∗ , if v∗ is on the boundary of Hk there is a unique unit-length normal to the closed half space Hk at v∗ ; for such an
Hk , let nk be this unique unit normal. On the other hand if Hk is such that v∗ is interior to Hk then the only normal at v∗
to Hk is the zero vector; for such Hk let nk be this zero vector.
Each Hk contains D ∩ Sa (which is the intersection of all the Hj ), so each nk is not only normal at v∗ to Hk but is also
normal at v∗ to D ∩ Sa. Therefore

n1 , n2 , n3 , . . ., nm +N (D ) is a set of m + N(D )vectors; and each nk is normal at v∗ to Hk and D ∩ Sa
Now D ∩ Sa equals the intersection of the m + N(D) closed half spaces Hk ; so any normal at v∗ to D ∩ Sa must be a non-
negative combination of these m + N(D) specially selected normals {n1 , n2 , n3 , …, nm+N(D) }. Therefore – c(v∗ ), being normal
at v∗ to D ∩ Sa, must be a non-negative combination of these m + N(D) normals at v∗ to D ∩ Sa.
Hence – c(v∗ ), being a non-negative combination of n1 , n2 , n3 , . . ., nm +N(D) , is the sum of
[a non-negative combination (= nSa say) of n1 , n2 , n3 , . . ., nm ] and

[a non-negative combination (= nD say) of nm +1 , nm +2 , nm +3 , . . ., nm +N(D) ]
By this construction, the vector nSa , being a non-negative combination of m unit normals at v∗ to Sa, must itself be
normal at v∗ to Sa and the vector nD , being a non-negative combination of N(D) normals at v∗ to D, must itself be normal
at v∗ to D. So
−c(v∗ ) = nS + nD
where nD is normal at v∗ to D and nSa is normal at v∗ to Sa. Subtracting the vector nS from both sides:
− ( c ( v∗ ) + nS ) = nD
and so is normal at v∗ to D. Now choose p∗ = nSa (which is, as shown above, normal at v∗ to Sa). Then:
v∗ belongs to D ∩ Sa
−[c(v∗ ) + p∗ ] = nD is normal at v∗ to D and also
p∗ = nSa is normal at v∗ to Sa
And so (v∗ , p∗ ) satisfies (13b). Note that, by the specification of Sa in (7b), Sa has no non-negative constraints and so,
since p∗ is normal at v∗ to Sa, p∗ ≥ 0.
Appendix 3
This appendix give the proof of Theorem 2.
Theorem 2. Suppose that (D × F) ∩ S∧ is non-empty, that

and
Suppose that each ci (vi ) is continuous and non-decreasing and that each hi (gi ) is continuous and non-increasing, that each
ci (•) is defined for all vi satisfying 0 ≤ vi ≤ si and that each hi (•) is defined for all gi satisfying 0 ≤ gi ≤ 1 Then there exists a
triple (v, g, p) with (v, g) in S^ and p ≥ 0 satisfying (14). (The conditions here may be weakened significantly without losing
the existence result.)
Proof. This follows from Theorem 1 above by considering the new augmented Beckmann objective function

vi
gi
Za(v, g ) = ci (u )du − hi (u )du
0 0
i i
defined for each (v, g) ∈ (D × F) ∩ S∧ . Using exactly the same arguments as those utilised in the proof of Theorem 1 leads to
a proof; but certain variables and sets must be replaced as follows:
S^ replaces Sa (so the augmented supply feasible set involves both link flows and link green-time proportions);
D × F replaces D (so the new “demand” set D × F includes the constraints on both link flows and link green time propor-
tions);
(c(v), -h(g)) replaces c(v) (so the new “extended link cost vector” (c(v), -h(g)) has a second component, -h(g), and the
extended link cost vector is felt by both the link flow vector v in D (-c(v) pushes v) and the link green-time proportion
vector g in F (h(g) pushes g); and
(p, −s◦p) replaces p (so as to ensure that we have a normal to the new augmented constraint set S^).
In applying the method of proof utilised in Theorem 1, it is critical here to observe that any normal nS ^ to the
new augmented supply-feasible S^ must be of the form (p, −s◦p) where p ≥ 0. This is central to the proof outlined in
Theorem 1 when applied in this new setting with the new augmented supply-feasible set S^ and the augmented cost func-
tion (c(v), -h(g)).
Appendix 4
This appendix gives the proof of Corollary 2.1.
and
is defined for all vi satisfying 0 ≤ vi ≤ si and that each hi (•) is defined for all gi satisfying 0 ≤ gi ≤ 1; as previously. Suppose also
that each ki > 0.
Then, provided MaxQ is sufficiently large, there exists a triple (v, g, b) satisfying (16).
Proof. Under the conditions stated in Corollary 2, Theorem 2 shows that there is a triple (v∗ , g∗ , p∗ ) satisfying (14). Regard
v∗ and g∗ as now fixed. Then define b∗ by putting
bi ∗ = pi ∗ /ki for i = 1, 2, 3, . . ., m
and b∗ = (b1 ∗ , b2 ∗ , b3 ∗ , . . ., bi ∗ , . . ., bm ∗ ). Thus we have a triple (v∗ , g∗ , b∗ ) satisfying (16).

Note: In the derivation of (5b) for the special ki there each link i has just a saturation flow si ; but here we also have
a green-time proportion gi . We would like to apply the result above to the consequent queue-dependent k with the link
outflow changed from si to si gi , as in formula (5c),
si gi ci ( si gi )
ki = ki ( si gi ) = 1 −
MaxQi
Plainly, the special ki given by (5c) is no longer fixed but now depends on gi . However, this does not matter in the proof
of the previous corollary as gi ∗ has already been fixed ahead of bi ∗ .
Thus there is a triple (v∗ , g∗ , b∗ ) satisfying (16) where each ki satisfies (5c).
Appendix 5
This appendix gives the detailed proofs in Tests 2, 3 and 4 in Section 2.7.1.2, 2.7.1.3, and 2.7.1.4. In the results here we
assume that ki = 1 − θ gi for i = 1, 2. We also utilise the variables r and a, where
r = s1 /s2 and a = 1 − θ .
Note: we insist that s1 < s2 and so r < 1 throughout this appendix.

2. Equilibrium performance of spatial P0 , P0k , with spatial queueing.
With the aboveki = 1 − θ gi , at equilibrium with spatial queueing: C1 + k1 b1 = C2 + k2 b2 .
Thus at equilibrium with spatial queueing: k1 b1 = k2 b2 + .
The spatial P0 policy is P0k (with the above ki ) so s1 k1 b1 = s2 k2 b2 .
Therefore: s1 k2 b2 + s1 = s1 k1 b1 = s2 k2 b2 ,
and so: s1 = (s2 - s1 )k2 b2
or: = [(s2 /s1 ) – 1]k2 b2 .
Thus, since r = s1 /s2 : k2 b2 = /[(s2 /s1 ) – 1] = r/[1 – r]
and so
E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 - C1 = k2 b2 + = r/[1 – r] + = /(1-r).
3. Equilibrium performance of P0 with spatial queueing. Here ki = 1 − θ gi for i = 1, 2.

At equilibrium with spatial queueing: C1 + k1 b1 = C2 + k2 b2 .
The P0 policy ensures that: s1 b1 = s2 b2 .
Therefore: s1 k2 b2 + s1 = s1 k1 b1 = k1 s1 b1 = k1 s2 b2 ,
and so: s1 = (k1 s2 - s1 k2 )b2
or: = [(k1 /k2 )(s2 /s1 ) – 1]k2 b2 .
Thus: k2 b2 = /[(k1 /k2 )(s2 /s1 ) – 1].
Now ki = 1 − θ gi for i = 1, 2. (k satisfies (5c), but here with ci = Ci (constant).) Also suppose in this example that MaxQ1
and MaxQ2 satisfy:
s1C1 s2C2
= = θ.
MaxQ1 MaxQ2
Then
E = [the travel time at equilibrium minus C1 ]
= C2 + k2 b2 − C1 = k2 b2 + = /[(k1 /k2 )(s2 /s1 ) − 1] +

= 1 1−s1 g1 C1 /MaxQ1
+
r . 1−s2 g2 C2 /MaxQ2 −1

= 1 1−θ g1
+
r . 1−θ g2 −1
In this case it is natural to consider three values of E by putting
(g1 = 0, g2 = 1), (g1 = 1, g2 = 0), (g1 = 1/2, g2 = 1/2).

These green-time proportions correspond to values of T slightly less than s2 , slightly greater than s1 and = ½(s1 +s2 ) (v/s).
Here we denote these T values as follows: T = s2 -, T = s1 + and T = ½(s2 +s2 ). If we let θ = ¼ corresponding values for E are as
follows:
r
T = s2 −, g1 = 0 and g2 = 1 : E= += += +=
1 1 −θ g 1 1
. 1 −1 1/a − r 1 − ra
.
r 1 −θ g 2
−1 r 1 −θ
r (r/a )
T = s1 +, g1 = 1 and g2 = 0 : E= += += += +=
1 1 −θ g 1 1
(1 − θ ) − 1 a−r 1 − r/a 1 − r/a
.
r 1 −θ g 2
−1 r
r r
T = 1/2(s1 + s2 ), g1 = 1/2 and g2 = 1/2 : E= += += +=
1 1 −θ g 1 1 −θ / 2 1−r 1−r
.
r 1 −θ g 2
−1 1 −θ / 2
−r
4. Equilibrium performance of biased spatial P0 , Phk , with spatial queueing. Here ki = 1 − θ gi .

With above ki , at equilibrium with spatial queueing: C1 + k1 b1 = C2 + k2 b2 .
The spatial biased P0 policy is Phk so: s1 k1 b1 + h1 = s2 k2 b2 + h2
Therefore: s1 k2 b2 + s1 = s1 k1 b1 = s2 k2 b2 + h2 – h1
and so: s1 = (s2 - s1 )k2 b2 + h2 – h1
or: = [(s2 /s1 ) – 1]k2 b2 + (h2 – h1 )/s1
Thus: k2 b2 = [ - (h2 – h1 )/s1 ]/[(s2 /s1 ) – 1].
So
E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 – C1 = k2 b2 + = [ - (h2 – h1 )/s1 ]/[(s2 /s1 ) – 1] + =
if we choose, for example, h = [s1 C1 , s1 C2 ].

References
Aashtiani, H., Magnanti, T., 1983. Equilibrium on a congested transport network. SIAM J. Algebraic Discrete Methods 2, 213–216.
Abdoulas, K., Geroliminis, N., 2013. Perimeter and boundary flow control in multi-reservoir heterogeneous networks. Transp. Res. Part B 55, 265–281.
Bar-Gera, H., 2002. Origin-based algorithm for the traffic assignment problem. Transp. Sci. 36 (4), 398–417.
Bar-Gera, H., 2010. Traffic assignment by paired alternative segments. Transp. Res. Part B 44 (8–9), 1022–1046.
Beckmann, M., McGuire, C.B., Winsten, C.B., 1956. Studies in the Economics of Transportation. Yale University Press, New Haven, CT.
Bliemer, M.C.J., Brederode, L., Wismans, L., Smits, E., 2012. Quasi-dynamic network loading: adding queuing and spillback to static traffic assignment.
Transportation Research Board 91st Annual Meeting (TRB) 26th January.
Bliemer, M.C.J., Raadsen, M.P.H., Smits, E.S., Zhou, B., Bell, M.G.H., 2014. Quasi-dynamic traffic assignment with residual point queues incorporating a first
order node model. Transp. Res. Part B 68, 363–384.
Bliemer, M.C.J., Raadsen, M.P.H., 2017. Static traffic assignment with residual queues and spillback. In: Proceedings of 17th Swiss Transportation Research
Conference, Monte Verita, Switzerland. Transport and Mobility Laboratory, EPF Lausanne, pp. 1–32.
Bundschuh, M., Vortisch, P., Van Vuren, T., 2006. Modelling queues in static traffic assignment. In: Proceedings of the European Transport Conference 2006.
Charnes, A., Cooper, W.W., 1961. Multicopy traffic network models. Theory of Traffic Flow. In: Herman, R. (Ed.), Proceedings of the Symposium on the Theory
of Traffic Flow, General Motors Research Laboratories. Elsevier, Warren, MI, pp. 85–96.
Dafermos, S., 1980. Traffic equilibrium and variational inequalities. Transp. Sci. 14 (1), 42–54.
Daganzo, C.F., 1995. The cell transmission model, part II: network traffic. Transp. Res. Part B 29 (2), 79–93.
Daganzo, C.F., 1998. Queue spillovers in transportation networks with route choice. Transp. Sci. 32 (1), 3–11.
Gentile, G., Meschini, L., Papola, N., 2005. Macroscopic arc performance models with capacity constraints for within-day dynamic traffic assignment. Transp.
Res. Part B 39 (4), 319–338.
Gentile, G., Meschini, L., Papola, N., 2007. Spillback congestion in dynamic traffic assignment: a macroscopic flow model with time-varying bottlenecks.
Transp. Res. Part B 41 (10), 1121–1138.
Haddad, J., Zheng, Z., 2018. Adaptive perimeter control for multi-region accumulation-based models with state delays. Transp. Res. Part B. https://doi.org/10.
1016/j.trb.2018.05.019.
Hajiahmadi, M., van de Weg, G., Tampère, C.M.J., Corthout, R., Hegyi, A., De Schutter, B., Hellendoorn, H., 2016. Integrated predictive control of freeway
networks using the extended link transmission model. IEEE Trans. Intell. Transp. Syst. 17 (1), 65–78.
He, X., Guo, X., Liu, H., 2010. A link-based day to day traffic assignment model. Transp. Res. Part B 44 (4), 597–608.
Hegyi, A., De Schutter, B., Hellendoorn, J., 2005. Model predictive control for optimal coordination of ramp metering and variable speed limits. Transp. Res.
Part C 13 (3), 185–209.
Huang, W., Viti, F., Tampére, C.M.J., 2016. Repeated anticipatory network traffic control using iterative optimization accounting for model bias correction.
Transp. Res. Part C 67, 243–265.
Huang, W., Viti, F., Tampére, C.M.J., 2017. An iterative learning approach for anticipatory traffic signal control on urban networks. Transp. B 5 (4), 402–425.
Jin, W.-L., 2010. Continuous kinematic wave models of merging traffic flow. Transp. Res. Part B 44, 8–9.
Keyvan-Ekbatani, M., Kouvelas, A., Papamichail, I., Papageorgiou, M., 2012. Exploiting the fundamental diagram of urban networks for feedback-based gating.
Transp. Res. Part B 46, 1393–1403.
Lam, W.H.K., Zhang, Y., 20 0 0. Capacity-constrained traffic assignment in networks with residual queues. J. Transp. Eng. 126 (2), 121–128.
Larsson, T., Patriksson, M., 1995. An augmented Lagrangean dual algorithm for link capacity constrained traffic assignment problems. Transp. Res. Part B 29,
433–455.
Lin, S., 2011. Efficient Model Predictive Control for Large-Scale Urban Traffic Networks Ph.D. thesis. Delft University of Technology, Delft, The Netherlands.
Marcotte, P., Nguyen, S., Schoeb, A., 2003. A strategic flow model of traffic assignment in static capacitated networks. Oper. Res. 52, 191–212.
Nesterov, Y., de Palma, A., 2003. Stationary dynamic solutions in congested transportation networks. Netw. Spat. Econ. 3, 371–395.
Ni, D., Leonard, J.D., 2005. A simplified kinematic wave model at a merge bottleneck. Appl. Math. Modell. 29, 1054–1072.
Nie, Y., Zhang, H., Lee, D., 2004. Models and algorithms for the traffic assignment problem with link capacity constraints. Transp. Res. Part B 38, 285–312.
Rinaldi, M., Tampère, C.M.J., Himpe, W., Holvoet, T., 2013. Centralized and decomposed anticipatory model predictive control for network-wide ramp me-
tering. In: Proceedings of the 16th International IEEE annual conference on Intelligent Transportation Systems. the Hague, the Netherlands.
Smith, M.J., 1979a. A local traffic control policy which automatically maximises the overall travel capacity of an urban road network. In: Proceedings of the
International Symposium on Traffic Control Systems, 21. University of California, Berkeley, pp. 298–302. Volume 2A, 11–32; and Traffic Engineering and
Control.
Smith, M.J., 1979b. The existence, uniqueness and stability of traffic equilibria. Transp. Res. Part B 13, 295–304.
Smith, M.J., 1984a. A descent algorithm for solving a variety of monotone equilibrium problems. In: Volmuller, J., Hamerslag, R. (Eds.), Proceedings of the
Ninth International Symposium on Transportation and Traffic Theory, pp. 273–297.
Smith, M.J., 1984b. A descent method for solving monotone variational inequalities and monotone complementarity problems. J. Optim. Theory Appl. 44,
485–496.
Smith, M.J., 1987. Traffic control and traffic assignment in a signal-controlled network with queueing. In: Gartner, N.H., Wilson, N.H.M. (Eds.), Proceedings
of the Tenth International Symposium on Transportation and Traffic Theory. Elsevier, pp. 61–68.
Smith, M.J., 2013. A link-based elastic demand equilibrium model with capacity constraints and queueing delays. Transp. Res. Part C 29, 131–147.
Smith, M.J., Huang, W., Viti, F., Tampere, C.M.J., 2013. Equilibrium in capacitated network models with queueing delays, queue-storage, blocking back and
control. In: 20th International Symposium of Traffic and Transport Theory, 80, Noordwijk, pp. 860–879. July 2013, and Transport Research Procedia.
Smith, M.J., Liu, R., Mounce, R., 2015. Traffic control and route choice: capacity maximization and stability. Transportation Research Part B, 81, 863-885. Also
in Proceedings of the 21st International Symposium on Transportation and Traffic Theory (Kobe, 2015). Transp. Res. Procedia 7, 556–577.
Taale, H., 2008. Integrated Anticipatory of Road Networks: A Theoretical Approach. Delft University of Technology Ph.D. Dissertation.
Tampère, C.M.J., Corthout, R., Cattryse, D., Immers, L.H., 2011. A generic class of first order node models for dynamic macroscopic simulation of traffic flows.
Transp. Res. Part B 45, 289–309.
Thompson, W.A., Payne, H.J., 1975. Traffic assignment on a transportation network with capacity constraints and queueing. Paper Presented at the 47th
National ORSA/TIMS North American Meeting 1975.
Viti, F., Van Zuylen, H.J., 2010. Probabilistic models of queues at fixed control signals. Transp. Res. Part B 44 (1), 120–135.
Wardrop, J.G., 1952. Some theoretical aspects of road traffic research. In: Proceedings of the Institute of Civil Engineers Part II, 1, pp. 325–378.
Zhong, R., Chen, C., Huang, Y., Sumalee, A., Lam, W.H.K., Xu, D., 2018. Robust perimeter control for two urban regions with macroscopic fundamental
diagrams: a control-Lyapunov function approach. Transp. Res. Part B 117, 687–707.

Transportation Research B: Mike Smith, Wei Huang, Francesco Viti, Chris M.J. Tampère, Hong K. Lo

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Transportation Research B: Mike Smith, Wei Huang, Francesco Viti, Chris M.J. Tampère, Hong K. Lo

Uploaded by

Copyright:

Available Formats

Transportation Research Part B 122 (2019) 140–166

Contents lists available at ScienceDirect

Transportation Research Part B

Quasi-dynamic traﬃc assignment with spatial queueing,

1.1. A motivational example

1.2. Context and contribution

Fig. 2. A single link model representing a single traﬃc lane.

2. A ﬁxed demand capacitated traﬃc assignment model with spatial queues

As is usual in traﬃc modelling, each real-life traﬃc lane is here represented by

and going with (5c) we have

2.2.1. Supply side conditions and supply feasibility

where ki is given by (5b).

2.2.2. Demand feasibility

(ii) X is also supply feasible or belongs to SaU and

(This condition is illustrated in Fig. 7 in Appendix 1.)

v ∈ D and − c(v ) · (w − v ) ≤ 0 for all w ∈ D (10b)

v belongs to D ∩ Sa, b ≥ 0 and

[the ﬂow vi = r Lir Xr = si and the bottleneck delay bi ≥ 0] or

This condition may be written as a complementarity condition:

s ◦ b < MaxQ where MaxQ = [MaxQ1 , MaxQ2 , ..., MaxQm ]

(i) that D ∩ Sa is non-empty and

Then there exists a pair (v, p) satisfying (13b).

Proof. The interested reader can ﬁnd the proof in Appendix 2.

Theorem 2. Suppose that (D × F) ∩ S∧ is non-empty, that

c(v ) = [c1 (v1 ), c2 (v2 ), c3 (v3 ), . . ., ci (vi ), . . ., cm (vm )]

h(g ) = [h1 (g1 ), h2 (g2 ), h3 (g3 ), . . ., hi (gi ), . . ., hm (gm )].

−grad Z(v∗ , g∗ ) = (−c(v∗ ), h(g∗ ) )is normal at (v∗ , g∗ ) to (D × F ) ∩ S∧ .

Proj(D×F )∩S∧ (−c(v∗ ) − p, h(g∗ ) + s ◦ p )

Corollary 2.1. Suppose that (D × F) ∩ S∧ is non-empty, that

Proof. This is given in Appendix 4.

Corollary 2.2. Suppose that (D × F) ∩ S∧ is non-empty, that

2.7. A simple example

Here we also suppose that

E = [the travel time at equilibrium minus C1 ] (s).

2.7.1.1. Equilibrium performance of the P0 control policy, with vertical queues.

So, in this case, if we put r = s1 /s2 :

E = [the travel time at equilibrium minus C1 ] = C2 + b2 – C1 = b2 +  = r/[1 – r] +  = /(1 − r).

E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 – C1 = k2 b2 +  = /(1-r).

Again, this formula only applies when r < 1.

2.7.1.3. Equilibrium performance of P0 with spatial queueing. Here ki = 1 − θ gi for i = 1, 2.

E = [the travel time at equilibrium minus C1 ] =  + .

E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 – C1 = k2 b2 = [ - (h2 – h1 )/s1 ] / [(s2 /s1 ) – 1] +  = 

if we choose, for example, h = [s1 C1 , s1 C2 ].

2.7.2. A graphical performance comparison

3. A ﬁxed demand capacitated traﬃc assignment model with blocking back

3.1. The link model with blocking back

3.1.1. Supply-feasibility with blocking back

−[c(v ) + k ◦ b] is normal at v to D, (26)

4. Existence and non-existence of equilibria in an illustrative example network

4.1. The example network

c1 (TAB − s3 ) − (c2 (s3 ) + c3 (s3 ))

Fig. 6. Graphical representation of two merging ﬂows.

4.6.1. General (ﬁxed ratio) merge models

A. No equilibrium queue exists at any of the entering links (point A)

only at link 3. This corresponds to point A in Fig. 6. Then:

4.7. Combining demand (or perimeter) and merging control

This appendix illustrates the deﬁnition of normal.

This appendix gives a proof of Theorem 1.

(i) that D ∩ Sa is non-empty and

Then there exists a pair (v, p) satisfying (13b).

[a non-negative combination (= nSa say) of n1 , n2 , n3 , . . ., nm ] and

This appendix give the proof of Theorem 2.

Theorem 2. Suppose that (D × F) ∩ S∧ is non-empty, that

E = [the travel time at equilibrium minus C1 ] = C2 + b2 – C1 = b2 + = r/[1 – r] + = /(1 − r).

E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 – C1 = k2 b2 + = /(1-r).

E = [the travel time at equilibrium minus C1 ] = + .

E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 – C1 = k2 b2 = [ - (h2 – h1 )/s1 ] / [(s2 /s1 ) – 1] + =

E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 - C1 = k2 b2 + = r/[1 – r] + = /(1-r).

E = [the travel time at equilibrium minus C1 ] = C2 + k2 b2 – C1 = k2 b2 + = [ - (h2 – h1 )/s1 ]/[(s2 /s1 ) – 1] + =