Position weighted backpressure intersection control for urban
networks
Li Lia , Saif Eddin Jabaria,b,∗
arXiv:1810.11406v2 [cs.SY] 3 Feb 2019
a New
b New
York University Tandon School of Engineering, Brooklyn NY
York University Abu Dhabi, Saadiyat Island, P.O. Box 129188, Abu Dhabi, U.A.E.
Abstract
This paper proposes a position weighted backpressure (PWBP) control
policy for network traffic. This is a decentralized policy that considers the
spatial distribution of vehicles along the length of network roads to account
for spill-back dynamics. The approach can be implemented in real time,
without a priori knowledge of network traffic demands, and can scale to
large urban networks. The capacity region of the network is formally defined in the context of urban network traffic and it is proved, when exogenous arrival rates are within the capacity region, that PWBP control
is network stabilizing. We demonstrate that PWBP can outperform adaptive signal control taken from a real-world setting. We perform comparisons against fixed signal timing, standard backpressure, and the capacityaware backpressure control policies using a calibrated microscopic simulation model of Abu Dhabi Island in the United Arab Emirates. We demonstrate the superiority of PWBP over the three other policies in terms of capacity region, network-wide delay, congestion propagation speed, recoverability from heavy congestion (outside of the capacity region), and response
to incidents.
Keywords: Decentralized control, backpressure, stochastic traffic flow,
urban networks, intersection control
1. Introduction
Various approaches have been proposed to optimize signal timing for
isolated intersections, including mixed-integer linear models, rolling hori∗ Corresponding
author, e-mail: sej7@nyu.edu
Preprint submitted to Elsevier
February 7, 2022
zon approaches, and store-and-forward models based on model predictive
control; see (Yann et al., 2011; Gartner, 1983; Tettamanti and Varga, 2010;
Mirchandani and Head, 2001; You et al., 2013; Ma et al., 2013) for examples.
On the one hand, isolated intersection approaches fail to account for spillback from adjacent road segments, which can eventually lead to gridlock
throughout a road network (Cervero, 1986). On the other hand, centralized techniques that include coordination between intersection controllers
(Heung et al., 2005; Gettman et al., 2007) are not scalable and difficult to
implement in real-world/real-time settings (Papageorgiou et al., 2003). For
example, ACS-Lite (Gettman et al., 2007) can handle no more than 12 intersections in real-time.
Recent articles in traffic control focus on connected-automated vehicles
(CAVs). This is well-justified considering the vast opportunities and challenges that CAVs have to offer. The ability to control both trajectories and
signals is one such opportunity that CAVs have to offer; we refer to (Li and
Zhou, 2017; Yu et al., 2018; Feng et al., 2018) and references therein for recent examples. For more information, we refer to a recent review article
that covers many aspects of intersection control: (Guo et al., 2019).
Decentralized control techniques have been proposed to overcome the
scalability issues associated with network control optimization. These techniques expect intersection controllers to be able to measure/estimate local
traffic information in real-time. This information includes expected traffic
demand at the intersection in the next cycle for heuristic approaches, e.g.,
(Smith, 1980; Lämmer and Helbing, 2008, 2010; Smith, 2011) or the queue
sizes along the intersection arcs in max pressure based approaches (Wongpiromsarn et al., 2012; Varaiya, 2013; Xiao et al., 2014; Le et al., 2015). According to (De Gier et al., 2011), control strategies that use traffic conditions
along both upstream and downstream arcs are more efficient and reliable
than those that utilize upstream traffic conditions only. Among the decentralized techniques, backpressure (BP) based approaches serve as examples
of techniques that utilize both upstream and downstream information.
They were first independently proposed in (Wongpiromsarn et al., 2012)
and (Varaiya, 2013) based on seminal work in communications networks
(Tassiulas and Ephremides, 1992) (see (Neely et al., 2005; Georgiadis et al.,
2006; Neely, 2010) for more details).
In general, BP based approaches are scalable and come with theoretical
guarantees of network stability. However, as it was originally developed
for packet queueing in communications networks, the assumptions are not
tailored to traffic problems and in some cases the assumptions are not suitable for traffic networks. Specifically, these assumptions include: (i) point
2
queues and (more critically) queue with infinite buffer capacities, (ii) separate queues for different commodities (corresponding to vehicles with different turning desires in traffic) and no interference between commodities.
A consequence of the first assumption is that the models do not account for
the spatial distribution of the queues, which has great impact on traffic control. For instance, Fig. 1 illustrates three different spatial distributions of
vehicles with the same queue size. Clearly, signal control decisions at the
downstream end should be very different for these three cases. A key point
here is how vehicle flux out of road segments are affected by the vehicle
distribution along the length of the road. While communications networks
assume that such “transmission rates” will not be influenced by the distribution of packets along the channels, in vehicular traffic the situation is
quite different. Point queueing techniques suffer this same drawback. For
example, the flow rate over the course of a short time interval (e.g. 10 sec)
at the downstream end of the road segment depicted in Fig. 1 should be
very different in the three cases. A serious consequence of assumption (i)
Queue
Build-‐up
Dynamics
Dense
Traffic
Free-‐flowing
Traffic
Queue
Dissipation
Dynamics
Free-‐flowing
Traffic
Dense
Traffic
Queue
Build-‐up
and
Dissipation
Dynamics
Dense
Traffic Free-‐flowing
Traffic
Dense
Traffic
Fig. 1: Three different spatial distributions of queues with same queue size.
is loss of work conservation, in which flow is prohibited across the intersection despite the availability of (spatial) capacity in the outbound roads.
Fig. 2 shows three cases in which BP control favors the eastbound approach
(Q a to Qb ), despite the fact that flow rates will be zero along this approach
if given priority. Recognizing the finite (spatial) capacity issue, (Gregoire
et al., 2015) proposed an improvement, referred to as capacity aware back
pressure (CABP) control. However, due to failure to account for the queue’s
3
spatial distribution, their approach can only avoid the case illustrated in
Fig. 2a, but not the two depicted in Fig. 2b and Fig. 2c (in the former, the
downstream queue is concentrated at the ingress of the road segment). As-
(a)
(b)
(c)
Fig. 2: Three non-work conserving cases (adopted from (Gregoire et al., 2015) and reproduced)
sumption (ii) could be easily violated in traffic networks, e.g., shared lanes.
Even when there are no shared lanes, road widening near the exits of inter4
section inbound roads, a very common geometrical features in urban networks, can create bottlenecks at the lane-branching point. Different turning
movements (commodities) interact at the bottleneck, and one queue may
block another if it gets too long as illustrated in Fig. 3. Work conservation
may also be violated here, as traditional BP (and CABP) control would favor the through movement, despite the fact that no through vehicles can
actually cross the intersection. This, in fact, serves as one physical mechanism that can lead to the scenario depicted in Fig. 2c. Loss of work conservation is a result of zero outflow if the through movement is given priority
and the prime culprit is the fact that the spatial distribution of vehicles is
not taken into consideration.
Fig. 3: Bottleneck at the lane-branching point.
This paper proposes decentralized intersection control techniques that
apply macroscopic traffic theory to overcome the issues described above.
We refer to this approach as position-weighted backpressure (PWBP). PWBP
considers the spatial distribution of vehicles along the road, applying higher
weights to queues that extend to the ingress of the road, thereby accounting
for the possibility of spillback. Rates of flow across the intersection depend
on both the control (signal status) and vehicle densities profiles (spatial distribution) along the inbound and outbound roads, capturing diminished
flow rates at signal phase startups (startup lost times). We perform comparisons in isolated intersection settings against a real-world implementation
of SCOOT as well as in a network setting against fixed intersection control,
5
standard BP, and CABP. We demonstrate superiority of PWBP in terms of
capacity region, delay, congestion propagation speed, recoverability from
heavy congestion and response to an incident.
The type of control proposed is applied to intersection signal control
(modern day traffic lights), but it can also be thought of as a prioritization
scheme for connected vehicles at network intersections that can guarantee
network stability. In both cases, when accurate measurement of the distribution of vehicles along the roads is not possible, one may employ a
light-weight traffic state estimation technique. We refer to (Jabari and Liu,
2013; Seo et al., 2017; Zheng et al., 2018) for recent examples.
The remainder of this paper is organized as follows: Sec. 2 describes the
traffic dynamics model, macroscopic intersection control, and the proposed
PWBP control policy. Sec. 3 rigorously demonstrates the network-wide stability properties of the PWBP approach using Lyapunov drift techniques.
A comparison with adaptive control at the isolated intersection level and
simulation experiments at the network level are provided in Sec. 4. Sec. 5
concludes the paper.
2. Problem formulation
2.1. Notation
Consider an urban traffic network represented by the directed graph
G = (N , A), where N is a set of network nodes, representing intersections
and A ⊂ N × N is a set of network arcs, representing road segments.
Each element of A is in one-to-one correspondence with an ordered pair
of elements in N . For each node, n ∈ N , Πn and Σn denote, respectively,
the set of (predecessor) arcs terminating in n and the set of (successor) arcs
emanating from n. We also use Π( a) ⊆ A to denote the set of predecessor
arcs to arc a ∈ A. That is, if n is the ingress node of arc a, then Π( a) = Πn .
Similarly, Σ( a) is the set of successor arc to arc a.
Fictitious source arcs are appended to the network to represent exogenous network arrivals. A new junction with indegree zero and outdegree
one is created for each exogenous inflow and the new source arc connects
this new node to the network boundary node; see Fig. 4. When exogenous
inflows occur at the interior of the network (i.e., at a junction with non-zero
in-degree) representing, for example, a parking ramp/lot, the associated
arc can be broken into two arcs with a new node placed at the position
of the merge; see Fig. 5. Source arcs will be assumed to have infinite jam
densities (i.e., they serve as fictitious reservoirs), but the flow rates in and
6
Boundary
junction
Fictitious
(source)
arc
Fig. 4: Fictitious boundary source arcs
Arc
with
exogenous
arrivals
Fictitious
(source)
arc
Fig. 5: Fictitious interior source arcs
out of these arcs are assumed to be finite (i.e., finite capacities). They shall
also be assumed to have zero physical length. Therefore, the traffic states
associated with fictitious source arcs are point queues concentrated at the
source node. We shall denote the set of (fictitious) source arcs by Asrc ⊂ A.
Arcs in Asrc serve two purposes: the first (mentioned above) is to model exogenous network inflows. The second is to capture instabilities in the network: Roads have finite spatial capacities and traffic densities are always
finite. Source arcs with infinite storage capacities are capable of capturing
network instabilities. For example, a signal control policy that results in instabilities is one where congestion propagates to the source arcs and builds
up there and exogenous arrivals can no longer be accommodated.
2.2. Stochastic arc dynamics
We denote the length of each arc a ∈ A by la . With slight notation
abuse, the upstream-most position (the entrance position) for each arc a in
7
the network is x = 0, while the downstream-most position (the arc exit
position) is x = la (that these coordinates pertain to arc a only should be
understood implicitly). We consider a multi-commodity framework, where
ρba ( x, t) denotes the traffic density at position x along arc a that is destined
to outbound arc b ∈ Σ( a) at time instant t. Similarly, qba ( x, t) denotes the
flow rate at x along a that is destined to b at time t. We define the state
of the system at time t as the vector of commodity-specific network traffic
densities. This is denoted as1 ρ(t) ≡ {ρba (·, t)}(a,b)∈M .
On the interiors of network arcs, we have the following conservation
equation: for each a ∈ A and b ∈ Σ( a)
∂ρba ( x, t)
∂qb ( x, t)
=− a
x ∈ (0, la ), t ≥ 0.
∂t
∂x
(1)
In a first order context, one sets qba ( x, t) ≡ φab ( x, t)Q a ρ a ( x, t) , where
φab ( x, t) is the fraction of vehicles at position x along arc a that is destined
to arc b at time t and Q a is a (stochastic) stationary flow-density relation
pertaining to arc a. In a higher order context, qba ( x, t) = ρba ( x, t)v a ( x, t),
where v is the macroscopic speed and
∂
dv a ( x, t)
∂
=
+ v a ( x, t)
v a ( x, t)
dt
∂t
∂x
∂ρ a ( x, t) ∂V a ρ a ( x, t)
loc
,
,
= A a ρ a ( x, t), V a ρ a ( x, t) ,
∂x
∂x
(2)
where Aloc
a are ‘local’ macroscopic acceleration models (Treiber and Kesting, 2013, Chapter 9) and V a is a stationary stochastic speed-density relation. The stochasticity in Q a and V a is parametric, that is, they can be described as generalizations of equilibrium fundamental relations that capture heterogeneity in the driving population as described in (Jabari et al.,
2014). For example, a generalization of Newell’s simplified relation (Newell,
2002):
Q a (ρ) = min v a ρ, wa (ρ − ρ a ) ,
(3)
is one where the parameters v a , wa , and ρ a , denoting free-flow speed, backward wave speed, and jammed traffic density, respectively, are random
1 We use the ‘·’ notation as a function argument to represent the entire curve in the dimension in which it is used. In other words, ρba (·, t) denotes the traffic density curve along
arc a destined to adjacent arc b at time instant t.
8
variables. We refer to (Jabari et al., 2018; Zheng et al., 2018) for the properties of the stochastic dynamics that arise as a result of a parametric treatment.
Remark 1. We make no assumptions about the relationship between flux and density. The proposed approach is equally valid in first and second order contexts. The
only assumptions we make are (i) flow conservation, (ii) probabilistic upper bounds
on flux and density, and (iii) that arc parameters do not change along the length of
the arc. The last assumption is easy to honor in a general network by splitting arcs
with varying parameters into more than one arc.
2.3. Boundary dynamics and junction control
At the arc boundaries, i.e., for x ∈ {0, la }, we employ a node model.
Node models represent the coupling between adjacent arcs and are responsible for capturing queue spillback dynamics.
Notation. For each node n ∈ N , let Mn denote the set of allowed movements between inbound and outbound road segments. The set Mn consists
of ordered pairs ( a, b) such that a ∈ Πn and b ∈ Σn , i.e., Mn ⊆ Πn × Σn .
The set of all network movements is denoted by M ≡ M1 ⊔ · · · ⊔ M|N | .
A signal phase consists of junction movements that do not conflict with
one another. We denote by Pn ⊆ 2Mn the set of allowable phases and
by P ⊆ ⊗n∈N Pn the set of allowable network phasing schemes. Essentially, an allowable phase is one that does not allow crossing conflicts and
only allows merging conflicts between a protected movement and a permitted movement. Example allowable phases are depicted in Fig. 6.
Fig. 6: Example phases for a four-leg isolated intersection.
Exogenous arrivals. For (fictitious) source arcs, we assume random arrivals; for commodity a ∈ Asrc and b ∈ Σ( a), let Aba (t) be a random (cumulative) arrival process with (instantaneous) rate λba (t) = E
9
dAba (t)
dt .
We, thus
have that2
dAba (t)
for a ∈ Asrc , b ∈ Σ( a).
(4)
dt
Junction control. Let p a,in (t) and p a,out (t) denote the upstream and
downstream control state at boundaries of arc a. The control state at time t
is defined as the set of movements that are active at time t as implied by the
active phases at the network
junctions. The boundary flows
are given by
b
b
b
b
q a (0−, t) = q a,in p a,in (t) and q a (la +, t) = q a,out p a,out (t) , where qba,in and
qba,out are boundary flux functions, which depend on the (boundary) control
variables and, implicitly, on the node dynamics (for instance, qba,in and qba,out
cannot exceed local supplies and demands at the arc boundaries).
We denote by q a,b pb,in (t) or equivalently q a,b p a,out (t) the rate of flow
that departs arc a ∈ Π(b) into arc b at time t. These are related to the
commodity flows at the arc boundaries as follows:
qc,a p a,in (t)
(5)
qba,in p a,in (t) = π a,b (t)
∑
qba (0−, t) =
c∈Π( a):(c,a)∈M
and
qba,out p a,out (t) = q a,b p a,out (t) ,
(6)
where π a,b (t) is the percentage of flow into a at time t that is destined
to adjacent arc b ∈ Σ( a). In the context of signalized urban networks, it
was demonstrated in (Jabari, 2016) that the node coupling, represented by
movement flows, is given uniquely by
(7)
q a,b p a,out (t) = 1 {(a,b)∈ pa,out (t)} min δa,b ρba (la +, t) , σb ρb (0, t) ,
where 1 {(a,b)∈ pa,out (t)} is an indicator function that returns 1 if the movement ( a, b) belongs to the phase p a,out (t) and returns 0 otherwise, δa,b is
a commodity-specific (local) demand function that depends on the traffic
density at the egress of arc a, σb is a (local) supply function that depends on
the total traffic density at the ingress of arc b:
ρb (0, t) =
∑
ρcb (0, t).
(8)
c:(b,c)∈M
processes Aba may have jumps. To be more accurate in such situations, one defines
Rt
Rt
Λba (t1 , t2 ) ≡ E t12 Aba (t)dt = t12 λba (t)dt. The boundary flux is then given by qba (0−, t−) =
R t+∆t b ′ ′
Rt
A a (t )dt .
lim t−∆t Aba (t′ )dt′ and qba (0−, t+) = lim t
2 The
∆t↓0
∆t↓0
10
Note that we adopt modified demand functions in order to account for startup
lost times; see (Jabari, 2016) and references therein for more details. The
local demand and supply functions are derived from the stationary flowdensity relations Q a . Thus, the source of randomness in qba,in p a,in (t) and
qba,out p a,out (t) is also parametric (i.e., the stochasticity is driven by the random parameters).
Finally, at the arc boundaries the conservation law (1) is given, for a ∈
A/Asrc , by
b
∂qba ( x, t)
∂ρba ( x, t)
q a,in p a,in (t) − qba (0, t) x = 0
(9)
=−
=
qba (la , t) − qba,out p a,out (t) x = la
∂t
∂x
and for a ∈ Asrc by
dρba (t)
dqb (t)
dAba (t)
=− a
=
− qba,out p a,out (t) .
dt
dx
dt
(10)
2.4. Network capacity region
Under any network-wide phasing scheme, p ∈ P , the traffic network
can “support” arrival processes with certain rates. Beyond these arrival
rates, queues along the source arcs will grow indefinitely. For each p ∈
P , we say that the network can support an arrival rate vector λ( p) =
[λ a ( p) · · · λ|A| ( p)]⊤ if
lim
T →∞
∑
a∈A
1
T
Z T
0
λ a ( p) + q a,in ( p) − q a,out ( p) dt = 0,
(11)
where with slight abuse of notation, q a,in ( p) and q a,out ( p) are the inflow and
outflow rates obtained when the network phasing scheme p is active. This
is interpreted as follows: the phasing scheme p is such that the total arc
outflow exceeds the total arc inflow in the long run.
In accord with (11), each p ∈ P defines a set of admissible arrival rates;
denote these (convex) polytopes by Ω( p). Taking the union of these sets,
∪ p∈P Ω( p), we get the vectors of all possible arrival rates that the network
can support under all p ∈ P . This is formally defined next.
Definition 1 (Maximal throughput region). The maximal throughput region
(a.k.a., capacity region) of the network, denoted by Λ, is the convex hull of all sets
of admissible flows. That is,
Λ ≡ Conv ∪ Ω( p) .
(12)
p∈P
11
Arrival rates that lie in Λ but not in ∪ p∈P Ω( p) are interpreted as arrival
rates that can be supported by switching between phasing schemes that lie
in the latter (i.e., time-sharing). A control policy that can support all possible
arrival rates in Λ is referred to as a throughput-maximal control policy. We
denote a control policy by a vector of network control states: at time t the
network control state is denoted by p(t) ≡ [· · · p a,in (t) p a,out (t) · · · ]⊤ , a
policy is an entire curve p(·).
We give two examples to illustrate the notion of capacity region. The
first is the simple isolated intersection of two one-way streets depicted in
Fig. 7a. If Phase p is active, the maximum arrival rate that can be accommodated is Ω( p). In the example below, each phase consists of only one
movement as shown in Fig. 7b. Consequently, the maximal arrival rate that
can be accommodated when Phase 1 (Phase 2) is active is a singleton denoted Ω( p1 ) (Ω( p2 )). Ω( p1 ) would consist of the saturation flow rate of
the Arc1-Arc3 (eastbound) movement and Ω( p2 ) would consist of the saturation flow rate of the Arc2-Arc4 (northbound) movement. The capacity
region, depicted in Fig. 7c, is the set of maximal arrival rates (λ1 , λ2 ) that
can be accommodated by switching between phases p1 and p2 .
λ2
<latexit sha1_base64="(null)">(null)</latexit>
Arc 4
Ω(p2 )
<latexit sha1_base64="(null)">(null)</latexit>
!
"
Λ = Conv Ω(p1 ) ∪ Ω(p2 )
<latexit sha1_base64="(null)">(null)</latexit>
Arc 3
Arc 1
Arc 2
Phase 1
(p1 )
Phase 2
(p2 )
<latexit sha1_base64="(null)">(null)</latexit>
<latexit sha1_base64="(null)">(null)</latexit>
λ1
<latexit sha1_base64="(null)">(null)</latexit>
Ω(p1 )
<latexit sha1_base64="(null)">(null)</latexit>
(a)
(b)
(c)
Fig. 7: Example isolated intersection and the associated capacity region, (a) intersection
layout (arcs 1 and 2 are source arcs), (b) the two possible phases, (c) the capacity region.
The second example is borrowed from (Varaiya, 2013, Example 3) and
illustrates how an instability forms in the proposed model, namely given
that we consider finite spatial arc capacities. In this example, depicted in
Fig. 8, there is one source arc. Hence, the capacity region is one dimensional. The initial conditions depicted in the figure are such that the network is in a state of gridlock at time t = 0. Moreover, the turning desires
shown in the figure prevent all vehicles from moving into their desired destination arcs. In this case, the capacity region consists of the singleton set
12
λ = 0. That is, the maximal arrival rate the network can accommodate is
zero. Any other arrival rate is outside of the capacity region and cannot be accommodated by any control policy, not BP, not CABP, not PWBP, nor any signal
timing optimization technique. In such cases, the only way to relieve gridlock
is to re-route vehicles; the subject of control+routing is beyond the scope of
the present paper, we leave it to future research.
Arc 4
ρ24 (x, 0) = ρ4
<latexit sha1_base64="(null)">(null)</latexit>
ρ42 (x, 0) = ρ2
<latexit sha1_base64="(null)">(null)</latexit>
Arc 1
Arc 2
ρ32 (x, 0) = 0
<latexit sha1_base64="(null)">(null)</latexit>
Arc 3
π2,3 (0) = π2,4 (0) = 0.5
<latexit sha1_base64="(null)">(null)</latexit>
π4,2 (0) = π1,2 (0) = 1.0
<latexit sha1_base64="(null)">(null)</latexit>
Fig. 8: A gridlock scenario (adopted from (Varaiya, 2013) and reproduced).
2.5. Position-weighted back-pressure (PWBP)
For any intersection n ∈ N , we assume that controllers are capable
of assessing the (average) movement fluxes associated with
any possible
phase p ∈ Pn . That is, for any ( a, b) ∈ Mn , E q a,b ( p) ρ(t) ≡ Eρ(t) q a,b ( p)
is known or can be estimated by the controller (locally). Omitting depenρ( t )
dence of δa,b and σb in (7) to simplify notation, define Pa,b ≡ P δa,b − σb ≤
0|ρ(t) . Then
Eρ(t) q a,b ( p) = 1 {(a,b)∈ p} Eρ(t) min δa,b , σb
ρ( t )
ρ( t )
= 1 {(a,b)∈ p} Pa,b Eρ(t) δa,b + 1 − Pa,b Eρ(t) σb .
ρ( t )
(13)
Note that Pa,b , Eρ(t) δa,b , and Eρ(t) σb are deterministic functions of ρ(t) that
depend on the distributions of the parameters of δa,b and σb . These distributions can be established empirically using historical data (Jabari et al., 2014).
The splits π a,b (t) are also treated as random quantities that are to be estimated or measured. In a fully automated system, these random variables
may degenerate, that is, it is easy to imagine that they can be measured
with high accuracy and become deterministic quantities. In present day
settings they need to be estimated. The setting envisaged in this paper is
one with mixed automated/connected and traditional vehicles. Connected
13
vehicles announce their turning desires upon entering arc a and may serve
as probes to allow the controller to estimate traffic conditions along the
arc and the split variables. Empirical techniques may also be employed for
this purpose; we refer to (Zheng and Liu, 2017) for a recent approach and to
(Rodriguez-Vega et al., 2019) for a recent article on reconstructing turning
movements.
The traffic state at time t, ρ(t), requires a traffic state estimation procedure that is capable of producing real-time estimates under present day
instrumentation in the real world. We refer to (Seo et al., 2017; Zheng et al.,
2018; van Erp et al., 2018) and references therein for recent research on traffic state estimation tools.
For each n ∈ N and each ( a, b) ∈ Mn , we define the weight variable
wa,b (t) = c a,b
Z la
x
0
la
ρba ( x, t)dx
Z lb
lb − x
−
lb
0
∑
cb,c πb,c (t)ρcb ( x, t)dx ,
c∈Σ(b):
(b,c)∈M
(14)
which depends on the (commodity) density curves along arcs a and b. To
interpret this, first note that
Z la
0
ρba ( x, t)dx
(15)
is just the total traffic volume (queue size) along arc a that is destined to
arc b. Then the first integral inside the square brackets in (14) can be interpreted as a weighted queue size, where traffic densities at the downstream
end of arc a (at x = la ) have the (maximal) weight of one, while traffic densities at the upstream end of a (at x = 0) have a weight of zero. In between,
the weights increase linearly with x. Similarly, the second integral inside
the square brackets in (14) can also be interpreted as a weighted queue size,
but with the weights decreasing linearly with x. Hence, the weight associated with movement ( a, b) decreases as the traffic densities at upstream
end (ingress) of arc b increase and vice versa, it increases when the traffic
densities are concentrated at the downstream end of arc a and vice versa.
The movement constants c a,b in (14) allow for assigning higher weights to
certain movements. The phase that is active at node n at time t under PWBP
control, denoted pn,PWBP (t), is given by
pn,PWBP (t) ∈ arg max
p∈Pn
∑
( a,b)∈Mn
14
wa,b (t)Eρ(t) q a,b ( p).
(16)
Since the number of possible phases at any intersection tend to be small
(typically four-eight), (16) can be easily solved by direct enumeration. This
allows for real-time distributed implementation of the proposed approach.
When there exists more than one solution to (16), a randomization procedure that applies equal weight to all the maximizers is employed. This
helps ensure work conservation as discussed below. Implementation of
PWBP control for node n at time instant t is summarized in Algorithm 1.
Algorithm 1: Position weighted backpressure phasing for node n at
time t: PWBP(n, t)
Input: Road geometry: {la } a∈Πn , {lb }b∈Σn ; movement constants: {c a,b }( a,b)∈Mn ;
(estimated) traffic state at time t: ρ(t); (estimated) splits at time t:
{π a,b (t)}(a,b)∈M ; the distributions of the parameters of {δa,b }(a,b)∈Mn and
{σb }b∈Σn
Iterate:
1: for ( a, b) ∈ Mn do
Rl
2:
wa,b (t) ←[ c a,b 0 a
x
la
ρba ( x, t)dx −
R lb
0
lb − x
lb
∑ c∈Σ(b): cb,c πb,c (t)ρcb ( x, t)dx
3:
for p ∈ Pn do
4:
Eρ(t) q a,b ( p) ←[ 1 {( a,b)∈ p} Eρ(t) min δa,b , σb
5:
end for
6: end for
Output:
7: if
arg max p∈Pn ∑( a,b)∈Mn wa,b (t)Eρ(t) q a,b ( p)
8:
(b,c)∈M
= 1 then
pn,PWBP (t) ←[ arg max ∑( a,b)∈Mn wa,b (t)Eρ(t) q a,b ( p)
p∈Pn
9: else
10:
Select pn,PWBP (t) at random from the set
n
arg max
p∈Pn
∑
wa,b (t)Eρ(t) q a,b ( p)
( a,b)∈Mn
o
assigning equal probabilities to each element in the set.
11: end if
One of the advantages of a continuous time formulation is that Algorithm 1 can be implemented at pre-specified cadence. Moreover, the cadence can vary from one intersection to another in order to accommodate
such constraints as minimum greens (to avoid aggressive oscillations in the
control dynamics), pedestrian movements, and so on. To elaborate, let τn
denote the minimum phase length for node n. The signal phasing sequence
is given by pn,PWBP (kτn ) = PWBP(n, kτn ) where k is a positive integer. An15
other advantage of the proposed approach is that it is decentralized; that
is, the calculations can be parallelized over the network nodes.
PWBP and work conservation. Work conservation of PWBP control
follows from two features of the proposed approach. The first feature is
the node model used: It was demonstrated in (Jabari, 2016) that the node
model produces holding-free solutions. Hence, for any chosen phase p ∈ P
(which dictates the node supplies), as long as there exist supply along the
outbound arcs, demands at the inbound are guaranteed to be served. The
second feature is that the phase chosen by PWBP depends on both the
movement weights wa,b and the expected movement fluxes, Eρ(t) q a,b ( p).
Since the weights are non-negative, the phase chosen is guaranteed to result in (holding-free) flow across the node as long as at least one of the
movements has a non-negative expected flux, Eρ(t) q a,b ( p). Holding only
occurs when Eρ(t) q a,b ( p) = 0 for all movements ( a, b) ∈ Mn . However,
this is gridlock scenario and no work is lost. In the case where the expected
fluxes are positive but the weights are zero, if all other expected fluxes are
zero, work may be lost. This corresponds to an alternative scenario where
max
p∈Pn
∑
wa,b (t)Eρ(t) q a,b ( p) = 0.
(17)
( a,b)∈Mn
In this case, the randomization procedure ensures with probability 1 that
loss of work does not persist.
3. Network stability
3.1. Lyapunov functional and stability
The traffic network is said to be strongly stable if (Neely, 2010, Definition
2.7):
Z
Z la
1 T
|ρba ( x, t)|dx dt < ∞.
(18)
E
lim sup
∑
0
T →∞ T 0 ( a,b)∈M
Since the network traffic densities depend (implicitly) on the control decisions at the network nodes, strong stability implies that the control in place
ensure that the network densities do not grow without bound in the long
run. This section demonstrates that as long as such a control policy exists3 ,
the PWBP algorithm ensures strong stability.
3 Otherwise,
there does not exist a control policy capable of stabilizing the network.
Hence, this is a feasibility assumption.
16
Since the spatial capacities of non-source arcs in the network are naturally bounded, the stability condition in (18) can be restated in terms of
source movements only, that is, the traffic network is said to be strongly
stable if
Z
1 T
lim sup
(19)
∑ src E|ρba (t)|dt < ∞.
T
0
T →∞
( a,b)∈M:a∈A
Consider the network-wide energy functional V : D → R with domain
D being an appropriately defined |A|-dimensional set of curves. V is defined as
Z la Z la
la − x − x′ b ′
1
V ρ( t ) ≡
c a,b
ρ a ( x , t)ρba ( x, t)dx ′ dx, (20)
∑
2 (a,b)∈M
l
0
0
a
where {c a,b }(a,b)∈M are non-negative finite constants. It can be easily shown
that V is a Lyapunov functional: (i) V ρ(t) ≥ 0 almost surely since traffic densities are non-negative (with probability 1) and (ii) V (ρ) = 0 if and
only if ρba ( x, t) = 0 almost surely for all ( a, b) ∈ M and all x ∈ [0, la ]4 .
Lemma 1 below provides a sufficient condition for strong stability using
the definition of Lyapunov functionals. A notational convention used below is V̇ ≡ dV/dt. For source nodes, la = 0 (no physical length); consequently, source arc commodities (i.e., when a ∈ Asrc ) contribute ρba (t)2 only
to the Lyapunov function, where dependence on position is dropped as an
argument since these are point queues.
Lemma 1. For the Lyapunov functional (20), suppose EV ρ(0) < ∞. If there
exist constants 0 < K < ∞ and 0 < ǫ < ∞ such that
Z la
|ρba ( x, t)|dx
(21)
Eρ(t) V̇ ρ(t) ≤ K − ǫ ∑ E
( a,b)∈M
0
holds for all t ≥ 0 and all possible ρ(t), then the traffic network is strongly stable.
P ROOF. We first integrate both sides of (21) over the interval [0, T ] and take
expectation of both sides of the inequality to obtain
E
Z T
0
E
ρ( t )
V̇ ρ(t) dt ≤ KT − ǫ
Z T
0
∑
( a,b)∈M
E
Z la
0
|ρba ( x, t)|dxdt.
(22)
4 One can construct pathological density curves with non-zero density spikes, where V =
0. However, such densities occur with probability zero. Technically, these are overcome by
using equivalence classes of density curves, but we shall avoid this level of technicality to
promote readability.
17
Reversing the order of the (outer) expectation and the integral, the lefthand side becomes
Z T
0
Z
EEρ(t) V̇ ρ(t) dt =
T
0
Z
EV̇ ρ(t) dt = E
V̇ ρ(t) dt
0
= EV ρ( T ) − EV ρ(0) . (23)
T
Combining this with (22), we get the inequality
EV ρ( T ) − EV ρ(0) ≤ KT − ǫ
Z T
0
∑
( a,b)∈M
E
Z
la
0
|ρba ( x, t)|dx dt. (24)
Dividing both sides by Tǫ and noting that EV ρ( T ) ≥ 0 (by definition of
V), we get the inequality
1
T
Z T
0
∑
( a,b)∈M
E
Z
la
0
K
1
|ρba ( x, t)|dx dt ≤ +
EV ρ(0) .
ǫ
ǫT
(25)
Noting that EV ρ(0) < ∞, the result follows immediately upon taking
the limit on both sides as T → ∞.
3.2. Stability of PWBP
According to Lemma 1, finding (finite) constants K and ǫ that satisfy the
condition (21) will ensure strong stability of the dynamics at the network
level. The constant K is established using the boundedness of the fluxes
q a (·, ·), which is a property of traffic flow (i.e., a physical property that must
be ensured by any model). On the other hand, ǫ depends on the intersection
control polices. Lemma 2 provides a necessary ingredient that will be used
later to establish the constant K.
Lemma 2. Let a ∈ A/Asrc and suppose there exist constants 0 ≤ q a < ∞ and
0 ≤ ρ a < ∞ such that P (qba ( x, t) ≤ q a ) = 1 and P (ρba ( x, t) ≤ ρ a ) = 1 for
any ( a, b) ∈ M, any x ∈ [0, la ], and any t ≥ 0. Then, there exist constants
(1)
(2)
0 ≤ Ka < ∞ and 0 ≤ Ka < ∞ such that, with probability 1,
(i) for any ( x1 , x2 ) ⊆ [0, la ]
Z
Eρ(t) qba (la , t)
x2
x1
and
18
(1)
ρba ( x, t)dx ≤ Ka
(26)
(ii) for any ( x1 , x2 ) ⊆ (0, la ) and any ( x3 , x4 ) ⊆ [0, la ]
− E ρ( t )
Z x4 Z x2
x3
x1
ρba ( x, t)
∂qba ( x ′ , t) ′
(2)
dx dx ≤ Ka .
∂x
(27)
P ROOF. First, note that
− E ρ( t )
Z x2
∂qba ( x, t)
dx = Eρ(t) qba ( x1 , t) − Eρ(t) qba ( x2 , t) ≤ q a
∂x
x1
(28)
with probability 1. By the boundedness properties of ρba (·, ·) for a ∈ A/Asrc ,
it holds that
Z x
4
b
P
(29)
ρ a ( x, t)dx ≤ la ρ a = 1.
x3
Hence, with probability 1, we have that
−E
ρ( t )
Z x4 Z x2
x3
x1
∂q a ( x ′ , t) ′
dx dx
∂x
Z
Z x4
b
ρ( t )
ρ a ( x, t)dx E
=−
ρba ( x, t)
x3
x2
x1
∂qba ( x, t)
dx ≤ la ρ a q a . (30)
∂x
The bound in (27) follows immediately and (26) follows from (29) along
with the boundedness of qba ( x, t) and then applying the Cauchy-Schwartz
inequality.
Corollary 1. Let a ∈ A/Asrc and assume the probabilistic bounds of Lemma 2.
Then, there exists a constant 0 ≤ K < ∞ such that
Z la
x b
ρ( t )
b
c
E
q
(
l
,
t
)
ρ
(
x,
t
)
dx
a,b
∑
a a
la a
0
( a,b)∈M:a6∈Asrc
Z la Z la − l − x − x ′
∂qba ( x ′ , t) ′
a
ρ( t )
b
ρ a ( x, t)
−E
dx dx ≤ K. (31)
la
∂x
0+
0
P ROOF. Since |(la − x − x ′ )/la | ≤ 1 for all ( x, x ′ ) ∈ [0, la ] × (0, la ), it follows
(1)
(2)
from Lemma 2 that there exists constants 0 ≤ Ka < ∞ and 0 ≤ Ka < ∞
for each a ∈ A/Asrc that bind each of the terms in the sums in (31) from
above. Defining
(1)
(2)
(32)
c a,b Ka + c a,b Ka ,
K≡
∑
( a,b)∈M:a6∈Asrc
the result follows immediately.
19
Theorem 3.1 (Stability of PWBP). Assume that the boundedness conditions
of Lemma 2 hold for all a ∈ A/Asrc , assume that arrival rates lie in Λ, that is,
there exists a control policy p∗ (·) that can stabilize the network in the sense defined
in Lemma 1, and let wa,b (·) for each movement ( a, b) ∈ M be as defined in (14).
Then the policy
wa,b (t)Eρ(t) q a,b ( p) ∀n ∈ N ,
∑
pn,PWBP (t) ≡ arg max
p∈Pn
(33)
( a,b)∈Mn
ensures strong stability of the traffic network.
P ROOF.
Eρ(t) V̇ ρ(t)
d
1
= Eρ(t) ∑ c a,b
2
dt
( a,b)∈M
=
1
c a,b Eρ(t)
2 (a,b∑
)∈M
∑
=−
Z la Z la
la − x − x′
ρba ( x ′ , t)ρba ( x, t)dx ′ dx
la
0
0
Z la Z la
la − x − x′ ∂
c a,b Eρ(t)
( a,b)∈M
0
la
0
∂t
Z la Z la
la − x − x′
0
0
ρba ( x ′ , t)ρba ( x, t) dx ′ dx
ρba ( x, t)
la
∂qba ( x ′ , t) ′
dx dx,
∂x
(34)
where the last equality follows from (1). For each a, the integrals inside the
expectation can be decomposed as
∂qba (0, t)
∂x
Z la
la − x
0
la
ρba ( x, t)dx +
+
Z la Z la −
0
0+
Z
∂qba (la , t) la x b
ρ ( x, t)dx
∂x
la a
0
∂qb ( x ′ , t) ′
la − x − x′ b
dx dx. (35)
ρ a ( x, t) a
la
∂x
Then from (9), we have that
Eρ(t) V̇ ρ(t) =
∑
( a,b)∈M:a∈Asrc
c a,b Eρ(t)
dAb (t)
a
dt
Z
c a,b Eρ(t) qba,in p a,in (t)
ρba (t) − qba,out p a,out (t) ρba (t)
la − x b
ρ
(
x,
t
)
dx
∑
a
la
0
( a,b)∈M:a6∈Asrc
Z la
la − x b
ρ( t )
b
−
c
E
q
(
0,
t
)
ρ
(
x,
t
)
dx
a,b
∑
a
a
la
0
( a,b)∈M:a6∈Asrc
Z la x b
ρ( t )
b
−
c
E
q
p
(
t
)
ρ
(
x,
t
)
dx
a,out
a,b
∑
a,out
la a
0
( a,b)∈M:a6∈Asrc
+
20
la
∑
+
( a,b)∈M:a6∈Asrc
−E
ρ( t )
Z
c a,b Eρ(t) qba (la , t)
Z
la
0
Z la −
la − x − x′
la
0+
0
la
x b
ρ a ( x, t)dx
la
∂qb ( x ′ , t) ′
dx dx
ρba ( x, t) a
∂x
. (36)
Appeal to Corollary 1 and noting that the third sum on the right-hand side
is non-negative (and does not involve control variables), we have that there
e < ∞ such that
exists a constant 0 < K
Z l
a
x b
ρ( t )
ρ( t )
e
ρ a ( x, t)qba,out p a,out (t) dx
E V̇ ρ(t) ≤ K − ∑ c a,b E
la
0
( a,b)∈M
Z la
dA a (t)
la − x b
b
−
dx , (37)
ρ a ( x, t) q a,in p a,in (t) +
la
dt
0
where qba,in ( p a,in (t)) ≡ 0 for a ∈ Asrc and dAba (t)/dt ≡ 0 for a ∈ A/Asrc .
Rl
Also, for a ∈ Asrc the traffic density is concentrated at x = 0, i.e., 0 a |(la −
x )/la |ρba ( x, t)dx = ρba (t). Upon re-arranging terms on the right-hand side
of (37) and utilizing the properties of conditional expectation, we have that
Z la
x b
ρ( t )
ρ( t )
e
ρ a ( x, t)E
q a,out p a,out (t)
E V̇ ρ(t) ≤ K − ∑
c a,b
la
( a,b)∈M 0
la − x b
la − x dAba (t)
dx. (38)
q a,in p a,in (t) − c a,b
− c a,b
la
la
dt
By assumption, we have that there exist constants 0 < K ∗ < ∞ and ǫ∗ > 0
associated with the policy p∗ (·) such that
E
ρ( t )
∗
V̇ ρ(t) ≤ K − ǫ
∗
∑
( a,b)∈M
E
Z la
0
|ρba ( x, t)|dx.
(39)
By definition, we have for each t ≥ 0 that
ǫ
∗
∑
Z la
( a,b)∈M 0
ρba ( x, t)dx
≤ max
− c a,b
p∈P
∑
Z la
( a,b)∈M 0
la − x b
q a,in ( p) − c a,b
la
x b
q
( p)
la a,out
la − x dAba (t)
dx. (40)
la
dt
ρba ( x, t)Eρ(t)
c a,b
e), we have by appeal to Lemma 1 that the
Hence, setting K ≡ max(K ∗ , K
control policy, p(·) ∈ P , which maximizes the right-hand side of (40) for
each t ≥ 0 is also network stabilizing.
21
It remains to show that is equivalent to (33). We have from (40) that
Z la
la − x b
x b
b
ρ( t )
q
( p) −
q a,in ( p) dx
ρ a ( x, t)E
arg max ∑ c a,b
la a,out
la
0
p∈P
( a,b)∈M
(41)
for each t ≥ 0 is network stabilizing. (The last term on the right-hand side
of (40) is dropped from the optimization problem since it constitutes an
additive constant to the problem.) From (5) and (6), the objective function
can be written as
Z la
la − x
x
b
ρ( t )
q ( p) −
∑ πa,b (t)qc,a ( p) dx,
∑ ca,b 0 ρa (x, t)E
la a,b
la
( a,b)∈M
c∈Π( a):
(c,a)∈M
(42)
which upon re-arranging terms and the orders of summation and integration becomes
Z la
x
∑ ca,b 0 la ρba (x, t)dx
( a,b)∈M
Z lb
lb − x
c
−
cb,c πb,c (t)ρb ( x, t)dx Eρ(t) q a,b ( p). (43)
∑
l
0
b
c∈Σ(b):
(b,c)∈M
The latter is bounded from above by
∑
( a,b)∈M
c a,b
Z la
x
la
0
−
ρba ( x, t)dx
Z lb
lb − x
0
lb
∑
cb,c πb,c (t)ρcb ( x, t)dx Eρ(t) q a,b ( p).
(44)
c∈Σ(b):
(b,c)∈M
Since intersection movements do not interact across nodes instantaneously,
the optimization problem naturally decomposes by intersection. That is,
maximizing (44) is equivalent to solving the |N | problems
arg max
p∈Pn
∑
wa,b (t)Eρ(t) q a,b ( p) ∀n ∈ N ,
( a,b)∈Mn
22
(45)
where
wa,b (t) ≡ c a,b
Z la
x
0
la
ρba ( x, t)dx
−
Z lb
lb − x
0
lb
∑
cb,c πb,c (t)ρcb ( x, t)dx .
c∈Σ(b):
(b,c)∈M
(46)
This completes the proof.
4. Experiments
4.1. Real-world isolated intersection experiment
In this section we will compare the SCOOT and PWBP using a calibrated intersection. The reason we only use a single intersection for the
comparison between SCOOT and PWBP is that we want to make our simulation precisely reproduce the demand and corresponding SCOOT signal
timing in real world, which is difficult to achieve in a network. Nevertheless, the comparison of network will be conducted in the following section
among PWBP, BP, CABP and fixed timing using different virtual demands.
Fig. 9 shows the layout of the calibrated single intersection: the Hamdan
Bin Mohammed Street - Fatima Bint Mubarak Street intersection in Abu
Dhabi in the United Arab Emirates (UAE). This intersection uses SCOOT
to optimize the signal timing in real world. We chose December 6, 2017 as a
typical working day and calibrated the traffic demand every 1 min for the
whole day, using high-resolution detector data. Then we use the calibrated demand as the demand input in our simulation. The real time SCOOT signal
timing was also extracted from historical data and imported to our simulator. In this way, our simulation reproduced what was happening in real
world. In the second step, by keeping the demand unchanged, we used the
PWBP to determine the signal timing in place of SCOOT.
Fig. 10a shows the number of vehicles passing the intersection under
both SCOOT and PWBP control over a 24 hour period, and Fig. 10b shows
their average delay. The figures demonstrate that PWBP outperforms SCOOT
substantially in terms of delay under both low and high demands. The average delay over the entire day for SCOOT is 95 seconds, while the average delay of PWBP is only 35 seconds. In addition, when demand is high,
PWBP control has higher throughput.
4.2. Network experiments
Network description. We utilize a microscopic traffic simulation network of a part of the city of Abu Dhabi consisting of eleven signalized intersections but also containing unsignalized intermediate junctions. The
23
Fig. 9: Layout of the calibrated single intersection in Abu Dhabi
network layout is shown in Fig. 11. We compare PWBP control with three
other control policies: fixed time, standard BP control, and CABP control.
The fixed timing plans are optimized and include optimal offsets (i.e., signal coordination). BP, CABP, and PWBP are all implemented using a software interface. To simplify the experiments, we utilize a uniform demand
at the boundaries, which we vary to gauge the capacity region of the network. Using a uniform (average) demand level allows us to use a single
number (namely the demand) as a way to gauge the capacity region.
Average network delay and network capacity region. Fig. 12 shows
the total network delay under different demand scenarios (ranging from
500 to 1800 veh/h on average) for BP, CABP, and PWBP using two types of
phasing schemes: one with four phases (“4-phase” scheme) and a scheme
with eight phases (“8-phase scheme”). The 4-phase scheme includes phases
1-4 in Fig. 6, while the 8-phase scheme is all eight phases in Fig. 6. We observe that 40 s/veh is a threshold delay, beyond which the delay increases
dramatically. We can hence treat 40 s/veh as indicative of reaching the
boundary of the capacity region. From Fig. 12, with the 8-phase scheme,
24
Number of vehicles passing the intersection
1000
SCOOT
PWBP
900
800
700
600
500
400
300
200
100
0
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time (h)
(a)
200
SCOOT
PWBP
180
Average delay (s/veh)
160
140
120
100
80
60
40
20
0
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time (h)
(b)
Fig. 10: Comparison between SCOOT and PWBP in (a)flow, (b)delay.
we see that delays begin to increase rapidly at a higher average demand
levels for the PWBP: 1620 veh/h for the 8-phase scheme vs. 1580 veh/h for
the 4-phase scheme. However, this is not the case for BP and CABP control, since they do not distinguish left-turning and through queues, which
results in blocking at the points where roads widen (left-turn lane addition). This indicates that BP and CABP have a wider capacity region using
a 4-phase scheme compared to the 8-phase scheme. All subsequent experiments use an 8-phase scheme with PWBP and 4-phase schemes with
25
150
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
PWBP_8 Phase
PWBP_4 Phase
Delay (s/veh)
Delay (s/veh)
Fig. 11: Simulation network in Abu Dhabi.
1580
1620
150
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
BP_8 Phase
BP_4 Phase
1370
(b)
CABP_8 Phase
CABP_4 Phase
Delay (s/veh)
Delay (s/veh)
(a)
150
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
1555
500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800
Demand (veh/h)
500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800
Demand (veh/h)
1400
1570
150
140
130
120
110
100
90
80
70
60
50
40
30
20
10
0
Fixed timing
1225
500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800
Demand (veh/h)
500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800
Demand (veh/h)
(c)
(d)
Fig. 12: Delay patterns at varying demand levels for different control policies.
BP and CABP. The demands at which delays begin to increase quickly for
fixed signal timing, BP, CABP, and PWBP are 1225, 1555, 1570, and 1620
26
veh/h, respectively. Fig. 13 shows a comparison of network delays for the
four control policies under varying demands.
160
Fixed timing
BP
CABP
PWBP
140
Delay (s/veh)
120
100
80
60
40
20
0
500
600
700
800
900
1000
1100
1200
1300
1400
1500
1600
1700
Demand (veh/h)
Fig. 13: Network delays associated with different control policies.
Congestion propagation. In the following experiments, we set the demand levels to the deterioration bounds of the control policies and compare
how congestion levels propagate over time. Since the deterioration bounds
for BP and CABP are close, we just use CABP’s bound (1570 veh/h); we,
hence, compare three demand scenarios. Fig. 14 – 15 show how the speeds
of all vehicles within the network are distributed under demand levels
1225, 1570 and 1620 veh/h. The horizontal axes in these figures are time
and the vertical axes are percentage of vehicles traveling at or below the
color-coded speeds. Under the different demand levels, the network eventually becomes grid-locked (at different levels for the different control policies). Specifically, it takes about four hours until total network gridlock
under a fixing timing plan when the demand reaches 1225 veh/h, under
BP and CABP it takes approximately six hours (at 1570 veh/h) until gridlock, and for PWBP, it takes approximately seven hours. This indicates that
PWBP is more resilient than the other policies. Fig. 16 shows how the total
number of vehicles (stuck) in the network evolves with time.
Recoverability from congestion. Fig. 17 shows how different control
policies recover from congestion. The total simulation time is eight hours,
the time interval from t = 120 min to t = 240 min is set as a congested period, during which demand levels are set to the deterioration bounds. We
set a demand of 1000 veh/h for the remainder of the eight-hour simulation
27
100%
20
40 km/h
10
30 km/h
0.1
40%
60
60
50 km/h
40 km/h
60%
50
50
40
30
20
10
0.1
40
30
20
10
0.1
20 km/h
20%
10 km/h
0
50
100
150
200
Time (min)
250
0 km/h
0
50
100
100%
50 km/h
200
250
40 km/h
60%
50
30 km/h
60
60
0 km/h
60 km/h
50 km/h
80%
Speed Percentile
Speed Percentile
100%
60 km/h
80%
50
10 km/h
(b) BP@1225vph
60
60
150
20 km/h
Time (min)
(a) FT@1225vph
40%
40 km/h
60%
50
50
40
40
30
20
10
30 km/h
40%
50
100
150
200
250
30
20
10
0.1
20%
10 km/h
0.1
0.1
0
20 km/h
40
30
20
10
40
30
20
10
20%
0 km/h
0
50
100
150
200
250
0 km/h
(d) PWBP@1225vph
100%
60 km/h
50 km/h
40 km/h
0.1
30 km/h
40%
60 km/h
50
40
30
80%
Speed Percentile
80%
60
60
50 km/h
20
50
10
40 km/h
60%
4
300
20
10
50040 20
3
0.1
2
10 0
10 km/h
Time (min)
(c) CABP@1225vph
100%
20 km/h
0.1
Time (min)
Speed Percentile
30 km/h
40%
20%
60%
60 km/h
80%
Speed Percentile
60%
100%
50 km/h
10
1
40
30
40
30
20
0.
80%
Speed Percentile
60 km/h
50
50
10
0.1
0.1
30 km/h
40%
0.1
20 km/h
20%
10 km/h
0
50
100
150
200
250
300
350
20 km/h
20%
0 km/h
10 km/h
0
Time (min)
50
100
150
200
250
300
350
0 km/h
Time (min)
(e) FT@1570vph
(f) BP@1570vph
Fig. 14: Network speed evolution, (a) fixed timing under a demand level of 1225 veh/h, (b)
BP under a demand level of 1225 veh/h, (c) CABP under a demand level of 1225 veh/h,
and (d) PWBP under a demand level of 1225 veh/h, (e) fixed timing under a demand level
of 1570 veh/h, (f) BP under a demand level of 1570 veh/h.
28
60
60
50 km/h
80%
0.1
40 km/h
40
30 km/h
0.1
0.1
100
150
200
250
300
350
0.1
0 km/h
0
50
100
0.1
150
250
300
20%
60
60 km/h
150
200
250
300
350
400
20
10
30 km/h
10
0.1
20 km/h
0 km/h
10 km/h
0
50
100
150
Time (min)
200
250
300
350
100%
0.1
60 km/h
60
60
50
0.
50
60%
40 0
3
20
40 km/h
10
30 km/h
0.1
20 km/h
20 km/h
20%
20%
10 km/h
0
40 0
3 20
40%
0.1
50 km/h
50
10
Speed Percentile
20
30 km/h
40%
80%
0.1
Speed Percentile
40 km/h
0.1
10
50
100
150
200
250
60 km/h
1
3
12000
50
40
50 km/h
40
30
0 km/h
(d) BP@1620vph
50
60%
400
Time (min)
(c) FT@1620vph
60
40 km/h
0.1
60%
20%
10 km/h
100
0 km/h
50 km/h
40 0
3
20 km/h
80%
350
50
40%
100%
20 km/h
2400350
0
0.1
30 km/h
40%
50
200
60
80%
Speed Percentile
40 km/h
10
0.1
Speed Percentile
100%
60 km/h
50 km/h
0
0.1
(b) PWBP@1570vph
80%
60%
30 km/h
Time (min)
(a) CABP@1570vph
4300500
2
40
30
20
10
10 km/h
Time (min)
100%
40 km/h
40
30
20
10
20%
10 km/h
50
50
50
60%
20 km/h
20%
0
60 km/h
50 km/h
40%
10
20
60
40 30 20 10 0.1
40%
Speed Percentile
3040
20 0
1
60%
60
80%
50
30
Speed Percentile
100%
60 km/h
0.1
50
4
103020
100%
300
350
400
0 km/h
10 km/h
0
Time (min)
50
100
150
200
250
300
350
400
0 km/h
Time (min)
(e) CABP@1620vph
(f) PWBP@1620vph
Fig. 15: Network speed evolution, (a) CABP under a demand level of 1570 veh/h, (b) PWBP
under a demand level of 1570 veh/h, (c) fixed timing under a demand level of 1620 veh/h,
and (d) BP under a demand level of 1620 veh/h, (e) CABP under a demand level of 1620
veh/h, (f) PWBP under a demand level of 1620 veh/h.
29
Number of vehicles in the network
7000
Fixed timing
BP
CABP
PWBP
6000
5000
4000
3000
2000
1000
0
0
50
100
150
200
250
300
350
400
450
300
350
400
450
300
350
400
450
Time (min)
(a)
Number of vehicles in the network
7000
Fixed timing
BP
CABP
PWBP
6000
5000
4000
3000
2000
1000
0
0
50
100
150
200
250
Time (min)
(b)
Number of vehicles in the network
7000
Fixed timing
BP
CABP
PWBP
6000
5000
4000
3000
2000
1000
0
0
50
100
150
200
250
Time (min)
(c)
Fig. 16: Evolution of total numbers vehicles in the network under different control policies
and demand levels of (a) 1225 veh/h, (b) 1570 veh/h, and (c) 1620 veh/h.
30
60
Fixed timing
BP
CABP
PWBP
Delay (s/veh)
50
40
30
20
10
0
0
30
60
90
120 150 180 210 240 270 300 330 360 390 420 450 480
Time (min)
(a) demand@1225vph
200
Fixed timing
BP
CABP
PWBP
180
160
Delay (s/veh)
140
120
100
80
60
40
20
0
0
30
60
90
120 150 180 210 240 270 300 330 360 390 420 450 480
Time (min)
(b) demand@1570vph
200
Fixed timing
BP
CABP
PWBP
180
160
Delay (s/veh)
140
120
100
80
60
40
20
0
0
30
60
90
120 150 180 210 240 270 300 330 360 390 420 450 480
Time (min)
(c) demand@1620vph
Fig. 17: Average network delay under varying peak period demands.
time. Fig. 17a, b and c only differ in the demand levels during the congested
period. The congested period demand levels are 1225, 1570 and 1620 veh/h
31
in Fig.17a, b and c, respectively. According to Fig. 17, for all tested scenarios, PWBP outperforms the other three control policies in terms of both delay and recovery time. Even when the peak demand reaches 1620 veh/h,
PWBP only needs 30 min to recover from the congestion, while fixed timing
needs about 90 min to recover with a peak demand of 1225 veh/h. Note
that when the peak demand reaches 1570 and 1620 veh/h, the delay levels under fixed timing becomes too high and hence cannot be shown in
Fig. 17b and c. We also see that using fixed timing, the network does not
eventually recover from congestion.
Response to an incident. We investigate the performance of PWBP in
the presence of an incident located at the yellow spot in Fig. 11. The incident is located half-way between intersections A and B, along a 3-lanes arc.
We test scenarios where one lane and two lanes are blocked for a duration
of one and two hours, and under different demand levels. Fig. 18 shows
the results for one-lane blocked cases when demand is 1500 veh/h. Fixed
140
120
BP_NI
CABP_NI
BP_1h
CABP_1h
BP_2h
CABP_2h
PWBP_NI
PWBP_1h
PWBP_2h
Delay (s/veh)
100
80
60
40
20
0
0
30
60
90
120
150
180
210
240
270
300
330
360
Time (min)
Fig. 18: Delays associated with different policies with one lane blocked by the incident
under a demand level of 1500 veh/h.
timing is not included here since 1500 veh/h is beyond its capacity region
and the delays will only increase without bound. Dotted lines represent
the non-incident cases, while dashed and solid lines represent the incident
cases with one and two hour durations, respectively. The incident starts at
the 60th min in both cases. When the incident duration is one hour, we see
that the network recovers within 30 minutes after the incident is cleared
32
under BP, CABP and PWBP. However, when the incident duration is two
hours, PWBP only needs one hour to completely recover, while congestion
in the network persists for significantly longer under BP and CABP: the effects of the incident are still felt in the network three hours after the incident
is cleared (compared to the no-incident scenarios).
Fig. 19 shows the two-lanes-blocked cases when demand is 1200 veh/h.
The network fails to recover under fixed timing, BP and CABP control
90
Fixed timing_NI
Fixed timing_1h
80
BP_NI
CABP_NI
PWBP_NI
BP_1h
CABP_1h
PWBP_1h
70
Delay (s/veh)
60
50
40
30
20
10
0
0
10
20
30
40
50
60
70
80
90
100
110
120
130
140
150
Time (min)
Fig. 19: Delays associated with different policies with two lanes blocked by the incident
under a demand level of 1200 veh/h.
when the incident blocks two of the three lanes. The delays increase sharply
and the whole network becomes gridlocked. In contrast, using PWBP control the incident hardly has any impact at all on network delay.
The reason of the performance difference between BP, CABP and PWBP
originates from how the model deals with scenarios in Fig. 2b and Fig. 2c.
With an incident located half-way between intersections A to B, the incident results in congested conditions (queueing) between the incident location and intersection A and low volume traffic between incident location
and intersection B. When the queue spills back to intersection A (similar
to Fig. 2b), PWBP will forbid the movements from A to B, while BP and
CABP fail to capture the spillback dynamics. In addition, PWBP does not
allocate green time at intersection B to the movement from A when there
are actually no vehicle near the stop line (similar to Fig. 2c), while BP and
CABP may still allocate green time to this movement.
33
5. Conclusion and outlook
Backpressure (BP) based intersection control is a control policy that
was originally developed for communications networks. Many of the assumptions made in the original theory were adopted in the BP applications
to traffic networks despite them not being applicable to vehicular traffic.
Specifically, infinite arc capacities, point queues, independence of commodities (turning movements), and there being no analogue for start-up
lost times in communications networks. These are critical features in intersection control. To accommodate these features, we develop a backpressure
control technique that is based on macroscopic traffic flow, which we refer
to as position-weighted backpressure (PWBP). PWBP considers the spatial
distribution of vehicles when calculating the backpressure weights.
The proposed PWBP control policy is tested using a microscopic traffic
simulation model of an eleven-intersection network in Abu Dhabi. Comparisons against coordinated and optimized fixed signal timing, standard
BP, and a capacity-aware variant of BP (CABP) were carried out. The results
indicate that PWBP can accommodate higher demand levels than the other
three control policies and outperforms them in terms of total network delay, congestion propagation speed, recoverability from heavy congestion,
and response to an incident.
This paper has focused on prioritization of movements at network intersections. As a possible future research direction, this can be extended
to include real-time route guidance. Another possible avenue for future
research is a combined perimeter/interior control policy. Perimeter control (Yang et al., 2017; Chiabaut et al., 2018; Ortigosa et al., 2014; Ambühl
et al., 2018; H. et al., 2018; Yang et al., 2018) is emerging as a useful tool for
network control at a macroscopic level. A study of the trade-offs between
the capacity region of an intersection control policy and perimeter control
could serve as a powerful network-wide control tool.
Acknowledgments
This work was funded in part by the C2 SMART Center, a Tier 1 USDOT University Transportation Center, and in part by the New York University Abu Dhabi Research Enhancement Fund. The authors also wish to
acknowledge the Abu Dhabi Department of Transportation for their support.
34
References
References
Ambühl, L., Loder, A., Menendez, M., Axhausen, K., 2018. A case study of
Zurichs two-layered perimeter control, in: 7th Transport Research Arena
(TRA 2018).
Cervero, R., 1986. Unlocking suburban gridlock. Journal of the American
Planning Association 52, 389–406.
Chiabaut, N., Küng, M., Menendez, M., Leclercq, L., 2018. Perimeter control as an alternative to dedicated bus lanes: A case study. Transportation
Research Record: Journal of the Transportation Research Board (in press,
DOI: 10.1177/0361198118786607), 0361198118786607.
De Gier, J., Garoni, T., Rojas, O., 2011. Traffic flow on realistic road networks
with adaptive traffic lights. Journal of Statistical Mechanics: Theory and
Experiment 2011, P04008.
van Erp, P., Knoop, V., Hoogendoorn, S., 2018. Macroscopic traffic state
estimation using relative flows from stationary and moving observers.
Transportation Research Part B 114, 281–299.
Feng, Y., Yu, C., Liu, H., 2018. Spatiotemporal intersection control in a
connected and automated vehicle environment. Transportation Research
Part C 89, 364–383.
Gartner, N., 1983. OPAC: A demand-responsive strategy for traffic signal
control. Transportation Research Record 906, 75–81.
Georgiadis, L., Neely, M., Tassiulas, L., 2006. Resource Allocation and
Cross-Layer Control in Wireless Networks. Now Publishers, Hanover,
MA.
Gettman, D., Shelby, S., Head, L., Bullock, D., Soyke, N., 2007. Data-driven
algorithms for real-time adaptive tuning of offsets in coordinated traffic
signal systems. Transportation Research Record 2035, 1–9.
Gregoire, J., Qian, X., Frazzoli, E., Fortelle, A.D.L., Wongpiromsarn, T.,
2015. Capacity-aware backpressure traffic signal control. IEEE Transactions on Control of Network Systems 2, 164–173.
35
Guo, G., Li, L., Ban, X., 2019. Urban traffic signal control with connected
and automated vehicles: A survey. Transportation Research Part C (To
appear).
H., H., Yang, K., Liang, H., Menendez, M., Guler, S., 2018. Providing public transport priority at urban network perimeters: A bi-modal perimeter
control approach, in: Transportation Research Board 97th Annual Meeting.
Heung, T., Ho, T., Fung, Y., 2005. Coordinated road-junction traffic control
by dynamic programming. IEEE Transactions on Intelligent Transportation Systems 6, 341–350.
Jabari, S., 2016. Node modeling for congested urban road networks. Transportation Research Part B 91, 229–249.
Jabari, S., Liu, H., 2013. A stochastic model of traffic flow: Gaussian approximation and estimation. Transportation Research Part B 47, 15–41.
Jabari, S., Zheng, F., Liu, H., Filipovska, M., 2018. Stochastic Lagrangian
modeling of traffic dynamics, in: 97th Annual Meeting of the Transportation Research Board.
Jabari, S., Zheng, J., Liu, H., 2014. A probabilistic stationary speed-density
relation based on Newells simplified car-following model. Transportation Research Part B 68, 205–223.
Lämmer, S., Helbing, D., 2010. Self-Stabilizing Decentralized Signal Control
of Realistic, Saturated Network Traffic (Technical Report No. 10-09-019).
Santa Fe Institute.
Lämmer, S., Helbing, S., 2008. Self-control of traffic lights and vehicle flows
in urban road networks. Journal of Statistical Mechanics: Theory and
Experiment 2008, P04019.
Le, T., Kovács, P., Walton, N., Vu, H., Andrew, L., Hoogendoorn, S., 2015.
Decentralized signal control for urban rroad networks. Transportation
Research Part C 58, 431–450.
Li, P., Zhou, X., 2017. Recasting and optimizing intersection automation
as a connected-and-automated-vehicle (CAV) scheduling problem: A sequential branch-and-bound search approach in phase-time-traffic hypernetwork. Transportation Research Part B 105, 479–506.
36
Ma, W., Xie, H., Liu, Y., Head, L., Luo, Z., 2013. Coordinated optimization
of signal timings for intersection approach with presignals. Transportation Research Record: Journal of the Transportation Research Board 2355,
93–104.
Mirchandani, P., Head, L., 2001. A real-time traffic signal control system:
Architecture, algorithms, and analysis. Transportation Research Part C
9, 415–432.
Neely, M., 2010. Stochastic network optimization with application to communication and queueing systems. Synthesis Lectures on Communication Networks 3, 1–211.
Neely, M., Modiano, E., Rohrs, C., 2005. Dynamic power allocation and
routing for time-varying wireless networks. IEEE Journal on Selected
Areas in Communications 23, 89–103.
Newell, G., 2002. A simplified car-following theory: A lower order model.
Transportation Research Part B 36, 195–205.
Ortigosa, J., Menendez, M., Tapia, H., 2014. Study on the number and
location of measurement points for an MFD perimeter control scheme: A
case study of Zurich. EURO Journal on Transportation and Logistics 3,
245–266.
Papageorgiou, M., Diakaki, C., Dinopoulou, V., Kotsialos, A., Wang, Y.,
2003. Review of road traffic control strategies. Proceedings of the IEEE
91, 2043–2067.
Rodriguez-Vega, M., Canudas-De-Wit, C., Fourati, H., 2019. Location of
turning ratio and flow sensors for flow reconstruction in large traffic networks. Transportation Research Part B 121, 21–40.
Seo, T., Bayen, A., Kusakabe, T., Asakura, Y., 2017. Traffic state estimation
on highway: A comprehensive survey. Annual Reviews in Control 43,
128–151.
Smith, M., 1980. A local traffic control policy which automatically maximises the overall travel capacity of an urban road network. Traffic Engineering & Control 21.
Smith, M., 2011. Dynamics of route choice and signal control in capacitated
networks. Journal of Choice Modelling 4, 30–51.
37
Tassiulas, L., Ephremides, A., 1992. Stability properties of constrained
queueing systems and scheduling policies for maximum throughput in
multihop radio networks. IEEE transactions on automatic control 37,
1936–1948.
Tettamanti, T., Varga, I., 2010. Distributed traffic control system based on
model predictive control. Periodica Polytechnica Civil Engineering 54,
3–9.
Treiber, M., Kesting, A., 2013. Traffic Flow Dynamics: Data, Models, and
Simulation. Springer–Verlag, Berlin.
Varaiya, P., 2013. Max pressure control of a network of signalized intersections. Transportation Research Part C 36, 177–195.
Wongpiromsarn, T., Uthaicharoenpong, T., Wang, Y., Frazzoli, E., Wang, D.,
2012. Distributed traffic signal control for maximum network throughput, in: Intelligent Transportation Systems (ITSC), 2012 15th International IEEE Conference on, pp. 588–595.
Xiao, N., Frazzoli, E., Li, Y., Wang, Y., Wang, D., 2014. Pressure releasing
policy in traffic signal control with finite queue capacities, in: Decision
and Control (CDC), 2014 IEEE 53rd Annual Conference on, pp. 6492–
6497.
Yang, K., Zheng, N., Menendez, M., 2017. Multi-scale perimeter control
approach in a connected-vehicle environment. Transportation Research
Part C: Emerging Technologies (in press, DOI: 10.1016/j.trc.2017.08.014),
101–120.
Yang, K., Zheng, N., Menendez, M., 2018. A perimeter control approach integrating dedicated express toll lanes, in: Transportation Research Board
97th Annual Meeting.
Yann, D., Boillot, F., Vanderpooten, D., Vinant, P., 2011. Multiobjective and
multimodal adaptive traffic light control on single junctions, in: Intelligent Transportation Systems (ITSC), 2011 14th International IEEE Conference on, pp. 1361–1368.
You, X., Li, L., Ma, W., 2013. Coordinated optimization model for signal
timings of full continuous flow intersections. Transportation Research
Record: Journal of the Transportation Research Board 2356, 23–33.
38
Yu, C., Feng, Y., Liu, H., Ma, W., Yang, X., 2018. Integrated optimization
of traffic signals and vehicle trajectories at isolated urban intersections.
Transportation Research Part B 112, 89–112.
Zheng, F., Jabari, S., Liu, H., Lin, D., 2018. Traffic state estimation using
stochastic Lagrangian dynamics. Transportation Research Part B 115,
143–165.
Zheng, J., Liu, H., 2017. Estimating traffic volumes for signalized intersections using connected vehicle data. Transportation Research Part C 79,
347–362.
39