This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 1

Sliding Mode Network Perimeter Control

Youssef Bichiou, Maha Elouni, Hossam M. Abdelghaffar , and Hesham A. Rakha , Fellow, IEEE

Abstract— Urban traffic congestion is a chronic problem

faced by many cities in the US and worldwide. It results in
inefficient infrastructure use as well as increased vehicle fuel
consumption and emission levels. Congestion is intertwined with
delay, as road users waste precious hours on the road, which
in turn reduces productivity. Researchers have developed, and
continue to develop, tools and systems to alleviate this problem.
Network perimeter control is one such tool that has been
studied extensively. It attempts to control the flow of vehicles
entering a protected area to ensure that the congested regime
predetermined by the Network Fundamental Diagram (NFD) is
not reached. In this paper, an approach derived from sliding
mode control theory is presented. Its main advantages over
proportional-integral controllers include (1) minimal tuning, (2)
no linearization of the governing equations, (3) no assumptions
with regard to the shape of the NFD, and (4) ability to handle
various demand profiles without the need to retune the controller.
A sliding mode controller was implemented and tested on a
congested grid network. The results show that the proposed
controller produces network-wide delay savings and disperses
congestion effectively. Fig. 1. The different segments of a typical NFD.

Index Terms— Network fundamental diagram, network

perimeter control, sliding mode control, traffic signal control. as the fundamental diagram [1], [2]. In addition, a number
of states can be identified on the fundamental diagram. These
I. I NTRODUCTION include the free-flow speed (u f ), which is the traffic stream
space-mean speed when the roadway is empty, and the speed-
R OADWAY traffic is a complex phenomenon that requires
serious modeling scalability. Nevertheless, various prop-
erties can be directly observed. These properties include
at-capacity (u c ), which is the traffic stream space-mean when
the flow is at maximum (prior to the onset of congestion).
(1) traffic stream density (k): the number of vehicles per unit These are the key properties used to model the behavior of
length of the road or lane; (2) traffic stream flow (q): the vehicles in any given network. The flow continuity equation
number of vehicles passing a fixed point per unit of time; and (i.e., conservation of mass) demonstrates that the flow through
(3) space-mean speed (u): the traffic stream density weighted a network is directly related to the density on the differ-
average speed. These three parameters are related using the ent roads. This relationship is characterized by the Network
hydrodynamic equation q = ku. Furthermore, it has been Fundamental Diagram (NFD), as presented in Fig. 1. If the
hypothesized that these variables along a roadway segment distance-weighted arithmetic means of the link densities and
are related to one another, forming what is commonly known flows are computed, a scaled NFD can be obtained. In general,
obtaining a clear and well-behaved NFD can be difficult. The
Manuscript received July 22, 2018; revised April 17, 2019 and scatter of points (as presented in Fig. 1) might not generate
December 24, 2019; accepted February 24, 2020. This work was supported
by the Department of Energy through the Office of Energy Efficiency and a visible curve. However, if the distribution of congestion is
Renewable Energy (EERE), Vehicle Technologies Office, Energy Efficient uniform across the network, the existence of a well-defined
Mobility Systems Program, under Award DE-EE0008209 [Product # DOE- NFD is guaranteed [3]–[5]. Godfrey [6] were the first to
VT-0008209-J05]. The Associate Editor for this article was N. Geroliminis.
(Corresponding author: Hesham A. Rakha.) introduce the concept of the NFD for the center of London.
Youssef Bichiou and Maha Elouni are with the Center for Sustainable The authors demonstrated that the relationship between the
Mobility, Virginia Tech Transportation Institute, Virginia Tech, Blacksburg, average velocity and the vehicle-travelled distance is parabolic
VA 24061 USA (e-mail: [email protected]; [email protected]).
Hossam M. Abdelghaffar is with the Department of Computer Engineering and that the speed is inversely proportional to the density.
and Systems, Faculty of Engineering, Mansoura University, Mansoura 35516, Geroliminis and Daganzo [7] observed the NFD in the con-
Egypt, and also with the Center for Sustainable Mobility, Virginia Tech gested urban network of Yokohama, Japan, demonstrating that
Transportation Institute, Virginia Tech, Blacksburg, VA 24061 USA (e-mail:
[email protected]). under homogeneous conditions, even with large discrepancies
Hesham A. Rakha is with the Department of Civil and Environmental between the different link fundamental diagrams, the NFD of
Engineering, Virginia Tech, Blacksburg, VA 24061 USA, and also with the network can be reproduced. This means that the NFD is
the Center for Sustainable Mobility, Virginia Tech Transportation Institute,
Virginia Tech, Blacksburg, VA 24061 USA (e-mail: [email protected]). a direct result of the infrastructure and thus is independent
Digital Object Identifier 10.1109/TITS.2020.2978166 of the number of vehicles in circulation, the demand, and the
1524-9050 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

selected routes (i.e., independent of the road taken by each techniques that alleviated the tuning problem [48]. In addition,
individual vehicle and the origin-destination table). the PIC developed in [38] requires linearization of the control
Leclercq and Geroliminis [8] acknowledges that the shape function, which further complicates and limits its use.
and scatter of the NFD is subject to local traffic hetero- Mirkin et al. [49] developed an adaptive sliding mode
geneities. This is achieved through a non-homogeneous dis- controller (SMC) that takes into account uncertainties and
tribution of congestion due to heterogeneous local capacities unknowns in the NFD, delayed input, and adjustable gains.
and route choice. He also concludes that it is difficult to link However, it requires numerous design parameters and needs
and understand the connection between local phenomena and model linearization.
the NFD. In this work, we developed an SMC [50] that has com-
Many clustering algorithms have been developed in order parable performance to the PIC, avoids the need for tuning,
to determine small regions where there is small variance in and makes no assumptions about the governing model (i.e., no
density and an NFD can be generated [9]–[15]. With the linearization is needed for the NFD). The present effort shows
determination of multi-region networks, researchers developed that only a set point (i.e., a target network vehicle density) is
multi-region controllers [16]–[18]. needed. This value is obtained only once from the NFD. The
Other studies focused on controlling only one region net- other parameters for this new controller can be evaluated in a
work. In order to avoid congestion in a specific region of a systematic manner.
given network (i.e., a protected network [PN]), flow optimiza- This paper is composed of seven sections. Section II
tion [19] can be useful. However, we will focus on gating, also presents the NFD equations. Section III is a brief description
referred to as perimeter control. There are gating solutions to of the PIC used in the literature. The introduced SMC is
reduce congestion, travel time, and delay, some of which are presented in Section IV. Section V presents a case study
based on the NFD and others that are not [20]–[28]. This to evaluate both controllers. The results are discussed in
paper focuses on methods that rely, in one form or another, Section VI, followed by concluding remarks.
on the NFD. The idea behind these methods is to control the
access points to a “protected” area (i.e., protected subnetwork) II. D ERIVATION OF THE NFD E QUATIONS
by regulating the entering traffic to ensure that the network NFD equations are used to plot the NFD curve in order to
will not operate beyond the capacity regime (i.e., maximum determine the set point for the SMC and the PIC used for
throughput). This can be achieved given information provided comparison purposes. It has to be noted that NFD equations
by the NFD. The aim is to avoid oversaturation (i.e., the are used in the PIC modelling but not in the SMC.
congested regime, as shown in Fig. 1). The NFD is computed based on the average link density (k)
Various researchers have tried to achieve this objec- in vehicles per unit length and the average vehicle flow (q)
tive [29]–[36]. Li et al. [37] investigated a perimeter control inside the network in vehicles per unit time. These quantities
strategy for an oversaturated network. They optimized the can be computed from loop detectors placed throughout the
green duration allocation in order to maximize the throughput network, where k is computed using (1).
using a genetic algorithm to minimize queues and delays 1
by optimizing phase sequences and offsets. However, their k[n] = k z [n].l z (1)
L
z∈Z
method used fixed signal timings, which does not reflect
typical real-time traffic conditions [29]–[36]. where z is the index for the link; Z is the set of all links
Many studies have overcome this issue and achieved real- belonging to the protected area where measurements are taking
time perimeter control using techniques such as the stan- place; n is the time index and reflects the cycle number; L is
dard proportional-integral controller (PIC) [38]–[41], a robust the total length
 of the PN (i.e., the sum of the length of all
PIC [10], a linear quadratic controller [14], and a model links L = z∈Z l z ; these links also feature loop detectors.); l z
predictive controller. Haddad and Mirkin [30], Haddad and is the length of link z; and k z [n] is the traffic stream density
Zheng [31], and Haddad and Mirkin [32] introduced various on link z during cycle n and is calculated using (2).
adaptive perimeter controller schemes that take into account oz [n − 1]
k z [n] = nl z k j z (2)
model uncertainty and NFD scatter. However, due to the 100
nonlinear nature of the NFD, model linearization is essential where nl z is the number of lanes of link z; k j z is the jam
for controller design. density of link z; oz is the measured time-occupancy (in
The PIC requires a number of input parameters, namely: a percent) on link z.
set point (the desired network density) and two gains that have The flow inside the network (q) is calculated using (3):
to be tuned a priori. The optimum gain values are typically 1
difficult to compute using the various methods to estimate the q[n] = qz [n].l z (3)
L
best gains [42]–[44]. Anandanatarajan et al. [45] and Sung and z∈Z
Lee [46] showed further limitations of PI and proportional- where qz [n] is the measured flow on link z.
integral-derivative (PID) controllers. Elouni and Rakha [47] The NFD is the plot relating the flow q[n] to the network
demonstrated that using a weather-tuned perimeter control, density k[n].
(i.e., retuning the PIC gains with respect to weather), improved
q[n] = ϕ (k[n]) (4)
the network’s performance metrics in terms of average speed
and delay. Kouvelas et al. developed robust adaptive tuning where ϕ is an unknown NFD function.

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BICHIOU et al.: SLIDING MODE NETWORK PERIMETER CONTROL 3

where n is the time index and K p = μ/ζ and

K I = (1 − μ)/ζ [38].
Based on measured data and the least squares method, μ
and ζ are estimated using (8) [40].
 
k [n + 1] − k̄ = μ. k [n] − k̄ + ζ.(q in [n] − q̄in ) + ε[n]
(8)
where q̄in is the average measured inflow of vehicles at the
capacity regime and ε[n] is an error term.
The values of μ and ζ are determined using data from
the NFD (i.e., scatter points) in the capacity regime [40].
Fig. 2. An abstract schematic of an urban network that features a congestion Tuning, in practice, could deliver negative and/or zero values.
protected subnetwork. Particular care must be taken with respect to the calibration
data so that appropriate μ and ζ values are obtained. There
also exist some robust tuning techniques that alleviate this
It is important to mention here that knowledge of the exact problem [48].
ϕ function is not necessary. The plot of the NFD scatter During simulation, the values of density k at the previous
(measurements of q and k during the simulation) is used to time step, as calculated using loop detectors, and the current
determine the set point that corresponds to the density having time step are needed. These are used to calculate the flow that
the highest flow. should enter the PN while avoiding congestion.
In general, the tuning process and the estimation of the
III. P ROPORTIONAL I NTEGRAL C ONTROLLER (PIC) density (k) add a layer of complexity to the controller.
This section introduces the PIC. Any given network can be Consequently, developing an approach that does not require
represented by an abstract schematic (Fig. 2). This network tuning, does not require linearization of the control equation,
experiences moderate to heavy congestion during the course and provides robustness is of interest.
of the day, leading to delays, excessive fuel consumption,
and pollution. To alleviate these effects, the identified area IV. P ROPOSED S LIDING M ODE C ONTROLLER (SMC)
of congestion (PN) is protected. That is, all the inflows (i.e., This section introduces the developed SMC. For this pur-
1 , q 2 , etc.) are monitored and in some cases restricted to
qin pose, we simplify the notation. Equation (5) can be re-written
in
the extent that the protected area operates at capacity. as
The time rate of change of the number of vehicles inside dk(t) qin (t) − qout (t) + qd (t)
the PN is computed using (5). =
dt L
d N V (t) = u(t) − Vout (t) + Vd (t) (9)
= qin (t) − qout (t) + qd (t) (5) 
dt where we assume that u = qin L is the system’s input. This
 input governs the maximum number of vehicles allowed to
where qin (t) = mi i=1
i
qin (t) is the sum of all inflows at the
instant t; qout (t) = mo i enter the protected area at instant t (i.e., qin ). For discrete
i=1 qout (t) is the sum of all vehicle
outflows from the subnetwork; qd (t) is the disturbance flow time steps, u[n] (i.e., n is the corresponding time step) would
that might occur inside the protected area; and N V = k × L determine the maximum number of vehicles allowed over a
is the number of vehicles inside the PN. N V is the product of given time horizon t. Equation (9) governs the time rate
the vehicle density k and the total length of the links L inside of change of the density (k) in the protected area. Here k is
the protected region. the state variable used in this work and represents the vehicle
In this section, we describe a few details of the PIC used density inside the protected area. Specifically, it is the length
in the literature [38, 40]. Since the objective is to avoid weighted average network density computed as the sum of the
congestion, a desired density k̄ is sought within a PN. This product of the link densities and lengths divided by the total
density is usually chosen to be the density at capacity (Fig. 1). length of the links in the protected area.
Therefore, the error defined in (6) is driven to zero when the It is important to note here that u(t) is proportional to the
control is activated. sum of all inflows entering the protected area through the mi
access points (i.e. access links).
e(t) = k(t) − k̄ (6)
1 i
mi
qin (t)
where k is the current density of vehicles in the protected area, u(t) = = qin (t) (10)
L L
and k̄ is the desired density (density at capacity). i=1

The PIC equation for discrete time steps is given by [40] To derive an expression for this input, we use sliding mode
as control theory [51] and introduce the error e as e = k − k̄,
where k̄ is a set point around which the PN is desired to
qin [n] = qin [n − 1] − K p (k[n] − k[n − 1]) operate. In this case, this point corresponds to the density at
+ K I (k̄ − k[n]) (7) capacity.

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

We introduce the variable x defined as For a discrete time step n,

 t  t
x(t) = e(τ )dτ = (k(τ ) − k̄)dτ (11) u s [n] = V̂out [n] − V̂d [n] − λ(k[n] − k̄) − γ sign k[n] − k̄
0 0

and S as 
n
d x(t) −λ (k[i ] − k̄)t (20)
S= + λx(t) = ẋ(t) + λx(t) (12) i=0
dt
where λ is a strictly positive real number. To ensure the surface defined by (13) is a stable surface
The sliding surface S is defined in Equation (13). This leads for the chosen controller (19), we introduce the Lyapunov
to the relationship ẋ(t) + λx(t) = 0. In other words, x decays function, defined by L F(S).
exponentially to zero given that λ is strictly positive. 1 1
S(x)T S(x) = S(X)2
L F(S(x)) = (21)
S(x(t)) = 0 (13) 2 2
The equilibrium (13) is stable if
Using (9), (11), and (12), we obtain d
(L F(S)) ≤ 0 (22)
d S(t)   dt
= u (t) − Vout (t) + Vd (t) + λ k (t) − k̄ (14)
dt At this point it is assumed that the actual control u(t) input
Since the trajectories are expected to remain on the surface to the system is bounded:
⎧
(i.e., (13)) for all time, d S(t)/dt has to remain at zero (i.e., ⎪
⎨ Umax i f u s ∈ [Umax , ∞[
d S(t)/dt = 0). This in turn leads to (15). u (t) = u s i f u s ∈ [Umin , Umax ] (23)
⎪
⎩
u ∗ (t) = Vout (t) − Vd (t) − λ(k(t) − k̄) (15) Umin i f u s ∈] − ∞, Umin ]
The values of Vout (t) and Vd (t) are not known since they Therefore, using Equation (14) we derive
represent the current outflow and disturbance flow in the d
network. Consequently, we use bounded estimates V̂out and (L F (S)) = S (u(t) − u s (t)) + Su s (t) − SVout (t)
dt  
V̂d with + SVd (t) + λS k (t) − k̄ (24)
 
 
V̂out (t) − Vout (t) ≤ α If u(t) does not hit the bounds, then u (t) = u s (t). Using
 
  Equations (17) and (18) we compute
V̂d (t) − Vd (t) ≤ β (16)  
d
(L F (S)) = S V̂out (t) − Vout (t)
where α and β are positive real numbers. It is important to note dt  
here that the bounds might not be always available or known. + S Vd (t) − V̂d (t) − γ |S| (25)
In this case, adaptive SMC can be used. In this approach,
the control gains are adapted dynamically to counteract the Using Equation (16) we obtain
uncertainties. For further details, the reader is referred to the d
work of Utkin and Poznyak [52]. In this effort, we assume (L F (S)) ≤ α |S| + β |S| − γ |S| (26)
dt
that α and β can be determined.
Choosing γ = α + β + η (with η strictly positive) leads to
Using the previous estimates, the new estimated controller
û becomes d L F(S)
≤ −η |S| . (27)
  dt
û(t) = V̂out (t) − V̂d (t) − λ k(t) − k̄ (17) If u(t) does hit the bounds, then u (t) = Ub where Ub refers
It is important to note here that we chose our estimates to any of the bounds. In this case using Equations (17) and
to be in the following form: V̂out (t) = Vout (t − t) and (18)
 
V̂d (t) = Vd (t − t). That is, our estimates are the observed d
(L F (S)) = S (Ub − u s ) + S V̂out (t) − Vout (t)
measurements of the mentioned quantities at the previous time dt  
step. This assumption requires loop detectors placed at the + S Vd (t) − V̂d (t) − γ |S| (28)
exits and on links in the PN.
We add to the estimated controller the term γ sign(S) as Using Equation (16) we obtain
u s = û − γ sign(S), (18) d
(L F (S)) ≤ |S| |Ub − u s |+α |S|+β |S|−γ |S| (29)
dt
where γ is a positive real number, to force the controller to There exists a positive real number θ such that θ >
always move toward the sliding surface. |Ub − u s |.
That leads to
 t Choosing γ = α + β + θ + η (with η strictly positive) leads
to
u s = V̂out − V̂d −λ(k − k̄)−γ sign k − k̄ −λ k − k̄dt
0 d L F(S)
≤ −η |S| . (30)
(19) dt

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BICHIOU et al.: SLIDING MODE NETWORK PERIMETER CONTROL 5

In other words, the distance to the sliding surface decreases

with time. This sliding surface S = ẋ +λx =  0 will be reached
in a finite time that is bounded by  S(t = 0) η. Consequently,
the choice of η will impact the time it takes the system to
reach the sliding surface. Once on the surface, the system will
remain there and will slide to the desired state, which in this
case is x = 0 exponentially with a time constant λ−1 .
It is important to note here that the value of k[n] is unknown
at the current time step n. Therefore, quantities measured from
the previous time steps were used as shown in equation (31).

V̂out [n] = qout [n − 1] L

V̂d [n] = qd [n − 1] L (31)
Also, the value of k[n] was replaced with k[n − 1]. Hence,
the controller becomes
qout [n − 1] qd [n − 1] Fig. 3. Grid network modeled in INTEGRATION.
u[n] = − − λ(k[n − 1] − k̄)
L L of its one-way streets and block lengths. This network includes
a protected region, also referred to as the PN. The PN
− (α + β + η) sign k[n − 1] − k̄
corresponds to the region surrounded by the green rectangle
(Fig. 3). All the access points to this subnetwork are identified

n−1
with the yellow arrows. In total there are eight links that feed
−λ (k[i ] − k̄)t (32)
directly into the protected area. For the purpose of this paper,
i=0
u min = 480 veh/h and u max = 12, 960 veh/h. These bounds
It is important to note here that the SMC suffers from a represent the minimum and maximum allowable vehicle flow
known problem—control/input chattering. This can be solved into the protected area via those eight links (Fig. 3).
by changing the “sign()” function in (18) into a saturation The PN contains 48 links, all of which are one-way road-
“sat()” function or a “tanh()” function. Chattering usually ways and each of which has only one lane of the same
results in the failure of components in mechanical systems. length of 150 m. The full network contains 36 signalized
However, in this research, the values of the inflow allowed in intersections. The network was modeled using INTEGRA-
the protected area are transformed into green times (explained TION software. The computation of various measures of
later in the paper; (34)). For each new cycle t (assumed to be effectiveness within INTEGRATION is beyond the scope of
60 s in this paper) a new value of the green time is obtained. this paper. The delays were validated in [53], the computation
Chattering in this case will result in the change in the green of vehicle stops was validated in [54], and the estimation of
time by a large or small value from one cycle to the next. This vehicle fuel consumption and emission levels are computed
change is expected not to cause a failure of any component using the VT-Micro model, which has also been extensively
and thus is less of an issue in this application. In contrast validated [55]–[57]. Origin and destination zones for trips are
to mechanical systems, chattering (i.e., sudden change of the represented by blue circles. Loop detectors are placed on each
control input) might cause the failure of components, for link of the network and collect measurements every cycle.
instance a sudden variation of the velocity input signal to a The cycle length is taken to be 60 s (i.e., t = 60 s).
motor might damage it. The demand for this network (D1) is generated during the
For the controller to function in an ideal manner, detectors first 75 minutes of the simulation, increasing during the first
are needed on the majority of the links of the protected area 37.5 minutes and decreasing after that to model the buildup
as well as at the entrance points. This is also true for the PIC and decay of traffic congestion (Fig. 4). The simulation time is
to monitor the state of the area in terms of associated vehicle taken to be 176 minutes in order to provide sufficient time for
density, inflow, and outflow. This information can then be used all vehicles to clear the network. Dynamic traffic assignment
to determine the estimated values needed by the controller (i.e., is also activated during the simulation to reflect realistic
V̂out (t) and V̂d (t) (31)). The developed controller features four driver behavior during congested conditions (i.e., rerouting of
hyper-parameters, namely α, β, λ, and η. α and β quantify vehicles is activated).
how close the estimated outflow and disturbance flows are to The NFD associated with the network presented in
the actual values at a particular instant in time. In the next Fig. 3 is shown in Fig. 5. During all “No Perimeter
section, this control law will be implemented and applied to Control” case simulations, the green time for the differ-
a grid network. ent traffic signals was optimized using phase split opti-
mization (Webster method [58]). The offset is also opti-
V. C ASE S TUDY mized using the procedures described in [59]. The point
In this section, the introduced grid network is used to cloud, shown in Fig. 5, presents snapshots of the simu-
conduct the controllers testing (Fig. 3). The network was lated sub-network state. Note that congestion is observed
created to replicate downtown Washington, DC, both in terms beyond an average network density of k̄ = 48 veh/km.

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

Fig. 6. Time series of the density k from the protected area when the PIC
Fig. 4. Base demand profile used in the simulation where each demand is activated and using the tuned parameters.
period spans 300 s. This demand profile is referred to as demand D1 in the
text.

Fig. 7. Time series of the density k from the protected area when the SMC
Fig. 5. NFD of the grid network presented in Fig. 3 for the demand profile control is activated. The five best cases are shown.
presented in Fig. 4.

Beyond this point, vehicles experience significant delays and Once on the surface (i.e., ẋ +λx = 0), the system converges
consume additional fuel. To avoid reaching this state, the exponentially to zero with a time constant of λ−1 with the
proposed SMC was implemented as described earlier in requirement that λ must be less than the minimum frequency
Section IV. of the un-modeled dynamics of the system [51]. This means
that we can choose λ as small as adequately possible. From
VI. R ESULTS AND S ENSITIVITY A NALYSIS
the previous results, we can conclude that almost no tuning
This section describes the various numerical simulations is needed for the parameters of the SMC. The model presents
that were performed. In the first subsection, the response only ranges of validity.
of the SMC is shown and compared to the PIC for var- Table I presents various responses of the system using
ious controller parameters. The inflow and outflow from different control parameters for the PIC and SMC for the
the protected area for the case of No Perimeter Control traffic demand presented in Fig. 4. The table also presents
(NPC) and SMC are presented in the second subsection. the base case (i.e., NPC with optimum signal timing for each
In the third subsection, the behavior of the SMC for a intersection) relative to which all comparisons are performed.
range of uncertainty is presented. Finally, the performance of A negative change means a decrease in value, while a positive
the SMC and PIC with respect to various demand profiles change indicates an increase in value. For the SMC, we tested
is shown. various combinations, in all cases we get a reduction with
respect to the base case. Fig. 6 and Fig. 7 present a comparison
A. Performance of the SMC and PIC for Different between the evolution of the density k inside the protected
Parameters area in the presence and absence of control. In Fig. 6, the
Tuning is required for the PIC. This requirement is not five tuned cases of the PIC were picked, and in Fig. 7 the
needed for the SMC. Only two parameters, λ and η, are best performing SMC cases in terms of travel time reduction
considered. η governs
 how fast
  the system converges to the
were picked. It is important to mention here that for the SMC
sliding surface.  S(x (t = 0)) η is the maximum time it takes the density in the protected area is generally below the target
for convergence to occur. density k̄ with the exception of SMC 6, which is slightly above
 Since we choose the control to be
activated when(k − k̄) k  ≤ 0.15 (i.e., control is activated k̄ for the first few time steps when the control is activated.
when k reaches 0.85k̄), the value of  S(x (t = 0)) η is small PIC 1 and 2 are clearly above the target value by a significant
for a wide range of η. margin and over the entire duration of the control interval
  (i.e., the time steps where the density is above 0.85 k̄). This
 S (x(t = 0)) 
  ≤ 0.15 k̄ (33) is clear for PIC 1, which exceeds the NPC case in some time
 η  η steps (i.e., congestion forms in the protected area). This in

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BICHIOU et al.: SLIDING MODE NETWORK PERIMETER CONTROL 7

TABLE I
S ENSITIVITY OF SMC AND PIC W ITH R ESPECT TO NPC (BASE C ASE ) FOR
VARIOUS VALUES OF C ONTROLLER PARAMETERS (λ, η, μ, AND ζ )

Fig. 9. Time series of the total vehicle inflow to the protected area from
all the entry points with and without control as well as the computed control
flow by the introduced logic.

Fig. 10. Time series of the outflow from the protected area when the control
is not activated and activated.

to mention here that inflow computed by the controller (qin )

is transformed into green times (G) that are allocated to the
different approaches feeding into the protected area. This is
performed using Equation (34). This transformation creates
discrepancies between the computed control inflow and the
Fig. 8. NFD of the grid subnetwork without and with control.
actual inflow when the control is activated (Fig. 9), given that
turn underscores the criticality of the tuning process and its the effective green time is different from the displayed green
importance in the PIC performance. time because of start loss and end gain (vehicles discharging
In the rest of the paper, and since the performance of most during the yellow indication).
of the various controllers with different parameters is almost
the same with respect to the reduction in travel time (Table I), qin
G=C (34)
the parameters of SMC 6 and PIC 5, which deliver the highest qs
reduction in delay, will be used.
Fig. 8 shows a comparison between two controllers: the
where qs is the saturation flow rate and C is the traffic signal
proposed SMC and the PIC for the demand profile D1. The
cycle length.
plot shows similar performance between the two controllers.
Fig. 9 shows that the flow of vehicles when control is
The tuned parameters for the PIC are taken to be μ = 0.847
activated is, on average, lower than when it is not activated.
and ζ = 0.002 (i.e., K I = 73.7, K P = 408), and k̄ = 48.76
Fig. 9 also shows that the controller attempts to maintain a
veh/km). For the SMC, the parameters are λ = 15 and
constant inflow that targets the optimum operating point of
η = 200. In the absence of control, we notice that the
(k̄, qmax ).
average network density of the protected area exceeds the
Fig. 10 shows the outflow of vehicles from the protected
optimum value k̄. Once the control is activated, congestion
area when the control is deactivated (NPC) and activated
is consistently eliminated from the protected area.
(SMC). It is important to note here that when the control
is activated from the time step of 33 to 62, the outflow of
B. Inflow and Outflow Profiles With and Without SMC
vehicles from the protected area (in the control case) is higher
Fig. 9 presents the time series of the inflow of vehicles than the NPC. This demonstrates that congested occurs in
without perimeter control (NPC) and with perimeter control the protected area for NPC, which results in lower system
for the demand profile D1 presented in Fig. 4. It is important throughput.

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

TABLE II
SMC P ERFORMANCE W ITH R ESPECT TO THE BASE C ASE
FOR VARIOUS VALUES OF α AND β

Fig. 13. Evolution of the density inside the protected area for the demand
profile D2 presented in Fig. 11 (a) without and with control.

Fig. 14. Evolution of the density inside the protected area for the demand
profile D3 presented in Fig. 11 (b) without and with control.

TABLE III
Fig. 11. Additional network demand profiles; each demand period spans AVG . S IMULATION R ESULTS FOR THE D EMAND P ROFILES OF F IG . 11
300 s. These demand profiles are referred to in the text as demand D2 and
D3, respectively.

Fig. 12. Evolution of the density inside the protected area for the demand
profile D1 presented in Fig. 4 without and with control.
shown in Fig. 11. Fig. 11(a) shows a dome-shaped demand
(i.e., demand D2), and Fig. 11(b) shows a sinusoidal demand
C. Performance of the SMC for Various Uncertainty Ranges
profile (i.e., demand D3). The evolution of the vehicle density
on qout and qd
for the different demands D1, D2, and D3 in the protected area
Table II shows the effects of the uncertainty on qout , α and is presented in Fig. 12, Fig. 13, and Fig. 14. It is important to
on qd , β. It is important to mention here that after careful note here that both controllers start from the 0.85k̄ threshold
examination of Equation (32), α and β play exactly the same to regulate the density inside the protected area to values
role in the controller. around k̄. Table III shows the results for the PIC and SMC.
As expected, as the uncertainty bound increases, the per- Table III presents the change in travel time, delay, fuel
formance of the controller degrades (i.e., lower reduction in consumption, and speed with respect to a base case for the
travel time, delay, fuel consumption, and lower increase in the PIC and SMC for the demand profiles D1, D2, and D3.
average vehicular speed relative to the base case; Table II). For the demand profile D1, we notice that both controllers have
quite similar performance, with one or the other exceeding
D. Response to Different Demand Profiles slightly in one or two measures (i.e., for instance, the PIC
To further demonstrate the effectiveness of the SMC, is slightly better in travel time with respect to the SMC).
we considered other demand profiles for this study. These are For demand profile D2, we notice a slight advantage of the

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

BICHIOU et al.: SLIDING MODE NETWORK PERIMETER CONTROL 9

SMC with respect to the PIC. This advantage is clear in [11] Y. Ji and N. Geroliminis, “On the spatial partitioning of urban
demand D3. The SMC outperforms the PIC by a relatively transportation networks,” Transp. Res. B, Methodol., vol. 46, no. 10,
pp. 1639–1656, Dec. 2012.
significant margin in the case of D3. This shows that the SMC [12] M. Saeedmanesh and N. Geroliminis, “Clustering of heterogeneous
adapts to a changing demand pattern whereas the PIC is less networks with directional flows based on ‘Snake’ similarities,” Transp.
adaptable. It should be noted here that dynamic re-routing is Res. B, Methodol., vol. 91, pp. 250–269, Sep. 2016.
[13] C. Lopez, P. Krishnakumari, L. Leclercq, N. Chiabaut, and
activated. Therefore, the demands at the various traffic signals H. van Lint, “Spatiotemporal partitioning of transportation network using
might vary. travel time data,” Transp. Res. Rec., J. Transp. Res. Board, vol. 2623,
no. 1, pp. 98–107, Jan. 2017.
VII. C ONCLUSION [14] K. Aboudolas and N. Geroliminis, “Perimeter and boundary flow control
in multi-reservoir heterogeneous networks,” Transp. Res. B, Methodol.,
In this paper, an SMC that attempts to alleviate traffic con- vol. 55, pp. 265–281, Sep. 2013.
gestion in regions within large urban networks was developed. [15] J. Haddad and N. Geroliminis, “On the stability of traffic perimeter
This controller was applied to the traffic stream flow continuity control in two-region urban cities,” Transp. Res. B, Methodol., vol. 46,
no. 9, pp. 1159–1176, Nov. 2012.
equation that governs the density of vehicles on any given [16] A. Kouvelas, M. Saeedmanesh, and N. Geroliminis, “A convex formula-
link and thus a congested region. The controller was tested tion for model predictive perimeter flow control in multi-region cities,”
on a grid network and the results suggest that it has similar presented at the 16th Swiss Transp. Res. Conf. (STRC), 2016.
[17] C. Menelaou, S. Timotheou, P. Kolios, and C. G. Panayiotou, “Joint route
performance to the PIC and in some cases outperforms it. The guidance and demand management for multi-region traffic networks,” in
SMC, however, does offer additional benefits, namely it makes Proc. 18th Eur. Control Conf. (ECC), Jun. 2019, pp. 2183–2188.
no assumptions with regards to 1) the form or simplifications [18] S. F. A. Batista and L. Leclercq, “Regional dynamic traffic assignment
framework for macroscopic fundamental diagram multi-regions models,”
of the governing equations, 2) the existence or nonexistence Transp. Sci., vol. 53, no. 6, pp. 1563–1590, Nov. 2019.
of the NFD, or 3) the shape of the NFD (i.e., linearization of [19] R. Mohebifard and A. Hajbabaie, “Distributed optimization and coor-
the NFD around the set point, then calibrating the PIC gains dination algorithms for dynamic traffic metering in urban street net-
using the linearized data). This controller also has consistent works,” IEEE Trans. Intell. Transp. Syst., vol. 20, no. 5, pp. 1930–1941,
May 2019.
performance with varying demand patterns. Unlike the PIC, [20] K. Aboudolas, M. Papageorgiou, and E. Kosmatopoulos, “Store-and-
it requires no tuning, but its parameters need be within a forward based methods for the signal control problem in large-scale
specified range. The user decides how fast the system should congested urban road networks,” Transp. Res. C, Emerg. Technol.,
vol. 17, no. 2, pp. 163–174, Apr. 2009.
converge to the sliding surface, then once on the sliding surface [21] E. Christofa, K. Ampountolas, and A. Skabardonis, “Arterial traffic
how fast it converges to the desired control (error of zero). The signal optimization: A person-based approach,” Transp. Res. C, Emerg.
controller consists of a single equation with various inputs that Technol., vol. 66, pp. 27–47, May 2016.
[22] D. Manolis, T. Pappa, C. Diakaki, I. Papamichail, and M. Papageorgiou,
are collected from the detectors assumed to be available on “Centralised versus decentralised signal control of large-scale urban road
all links within and feeding to the congested region. Future networks in real time: A simulation study,” IET Intell. Transp. Syst.,
work will look into ways to reduce the number of detector vol. 12, no. 8, pp. 891–900, Oct. 2018.
[23] L. Bu, F. Wang, X. Zhou, and C. Yin, “Managed gating control strategy
measurements needed to run the controller. for emergency evacuation,” Transportmetrica A, Transp. Sci., vol. 15,
no. 2, pp. 963–992, Dec. 2018.
R EFERENCES [24] H. M. Abdelghaffar, H. Yang, and H. A. Rakha, “Isolated traffic signal
[1] H. Rakha and M. Arafeh, “Calibrating steady-state traffic stream and control using a game theoretic framework,” in Proc. IEEE 19th Int. Conf.
car-following models using loop detector data,” Transp. Sci., vol. 44, Intell. Transp. Syst. (ITSC), Nov. 2016, pp. 1496–1501.
no. 2, pp. 151–168, May 2010. [25] H. M. Abdelghaffar, H. Yang, and H. A. Rakha, “Developing a de-
[2] H. Rakha, “Validation of van Aerde’s simplified steadystate car- centralized cycle-free Nash bargaining arterial traffic signal controller,”
following and traffic stream model,” Transp. Lett., vol. 1, no. 3, in Proc. 5th IEEE Int. Conf. Models Technol. for Intell. Transp. Syst.
pp. 227–244, 2009. (MT-ITS), Jun. 2017.
[3] C. Buisson and C. Ladier, “Exploring the impact of homogeneity [26] R. Mohebifard, S. M. A. B. A. Islam, and A. Hajbabaie, “Coopera-
of traffic measurements on the existence of macroscopic fundamental tive traffic signal and perimeter control in semi-connected urban-street
diagrams,” Transp. Res. Rec., J. Transp. Res. Board, vol. 2124, no. 1, networks,” Transp. Res. C, Emerg. Technol., vol. 104, pp. 408–427,
pp. 127–136, Jan. 2009. Jul. 2019.
[4] Y. Ji, W. Daamen, S. Hoogendoorn, S. Hoogendoorn-Lanser, and [27] R. Mohebifard and A. Hajbabaie, “Dynamic traffic metering in urban
X. Qian, “Investigating the shape of the macroscopic fundamental street networks: Formulation and solution algorithm,” Transp. Res.
diagram using simulation data,” Transp. Res. Rec., J. Transp. Res. Board, C, Emerg. Technol., vol. 93, pp. 161–178, Aug. 2018.
vol. 2161, no. 1, pp. 40–48, Jan. 2010. [28] R. Mohajerpoor, M. Saberi, H. L. Vu, T. M. Garoni, and M. Ramezani,
[5] N. Geroliminis and J. Sun, “Properties of a well-defined macroscopic “H∞ robust perimeter flow control in urban networks with partial
fundamental diagram for urban traffic,” Transp. Res. B, Methodol., information feedback,” Transp. Res. B, Methodol., to be published.
vol. 45, no. 3, pp. 605–617, Mar. 2011. [29] J. Haddad, M. Ramezani, and N. Geroliminis, “Cooperative traffic
[6] J. W. Godfrey, “The mechanism of a road network,” Traffic Eng. control of a mixed network with two urban regions and a freeway,”
Control, vol. 8, no. 8, pp. 323–327, 1970. [Online]. Available: Transp. Res. B, Methodol., vol. 54, pp. 17–36, Aug. 2013.
https://trid.trb.org/view/117139 [30] J. Haddad and B. Mirkin, “Coordinated distributed adaptive perimeter
[7] N. Geroliminis and C. F. Daganzo, “Existence of urban-scale macro- control for large-scale urban road networks,” Transp. Res. C, Emerg.
scopic fundamental diagrams: Some experimental findings,” Transp. Res. Technol., vol. 77, pp. 495–515, Apr. 2017.
B, Methodol., vol. 42, no. 9, pp. 759–770, Nov. 2008. [31] J. Haddad and Z. Zheng, “Adaptive perimeter control for multi-
[8] L. Leclercq and N. Geroliminis, “Estimating MFDs in simple networks region accumulation-based models with state delays,” Transp. Res. B,
with route choice,” Procedia-Social Behav. Sci., vol. 80, pp. 99–118, Methodol., to be published.
Jun. 2013. [32] J. Haddad and B. Mirkin, “Adaptive perimeter traffic control of urban
[9] C. F. Daganzo, “Urban gridlock: Macroscopic modeling and mitigation road networks based on MFD model with time delays,” Int. J. Robust
approaches,” Transp. Res. B, Methodol., vol. 41, no. 1, pp. 49–62, Nonlinear Control, vol. 26, no. 6, pp. 1267–1285, Jan. 2016.
Jan. 2007. [33] I. I. Sirmatel and N. Geroliminis, “Economic model predictive control
[10] J. Haddad and A. Shraiber, “Robust perimeter control design for of large-scale urban road networks via perimeter control and regional
an urban region,” Transp. Res. B, Methodol., vol. 68, pp. 315–332, route guidance,” IEEE Trans. Intell. Transp. Syst., vol. 19, no. 4,
Oct. 2014. pp. 1112–1121, Apr. 2018.

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

[34] A. Kouvelas and N. Geroliminis, “Hierarchical control for large-scale [57] K. Ahn, H. Rakha, A. Trani, and M. Van Aerde, “Estimating vehicle
urban road traffic networks,” in Proc. 7th Int. Conf. Transp. Netw. fuel consumption and emissions based on instantaneous speed and
Rel., 2018, Art. no. 18-04080. [Online]. Available: https://www.research- acceleration levels,” J. Transp. Eng., vol. 128, no. 2, pp. 182–190,
collection.ethz.ch/handle/20.500.11850/275976 Mar. 2002.
[35] J. Haddad, M. Ramezani, and N. Geroliminis, “Model predictive perime- [58] R. P. Roess, E. S. Prassas, and W. R. McShane, Traffic Engineering.
ter control for urban areas with macroscopic fundamental diagrams,” in London, U.K.: Pearson, 2010.
Proc. Amer. Control Conf. (ACC), Jun. 2012, pp. 5757–5762. [59] H. Rakha and M. V. Aerde, “An off-line emulator for estimating the
[36] G. Mariotte and L. Leclercq, “Heterogeneous perimeter flow distribu- impacts of SCOOT,” presented at the 74th Transp. Res. Board Annu.
tions and MFD-based traffic simulation,” Transportmetrica B, Transp. Meeting, Washington, DC, USA, 1995.
Dyn., vol. 7, no. 1, pp. 1378–1401, Jun. 2019.
[37] Y. Li, J. Xu, and L. Shen, “A perimeter control strategy for oversatu-
Youssef Bichiou received the Ph.D. degree in
rated network preventing queue spillback,” Procedia-Social Behav. Sci.,
engineering mechanics from the Virginia Polytech-
vol. 43, pp. 418–427, 2012.
nic Institute and State University (Virginia Tech),
[38] M. Keyvan-Ekbatani, A. Kouvelas, I. Papamichail, and M. Papageorgiou, Blacksburg, VA, USA. He finished his graduate
“Exploiting the fundamental diagram of urban networks for feedback- studies in the field of fluid structure interaction in
based gating,” Transp. Res. B, Methodol., vol. 46, no. 10, pp. 1393–1403, the Department of Mechanical Engineering, Tunisia
Dec. 2012. Polytechnic School, La Marsa, Tunisia. He is cur-
[39] M. Elouni, H. A. Rakha, and Y. Bichiou, “Implementation and investiga- rently a Research Associate with the Center for Sus-
tion of a weather-and jam density-tuned network perimeter controller,” tainable Mobility, Virginia Tech Transportation Insti-
in Smart Cities, Green Technologies and Intelligent Transport Systems. tute. His research interests include traffic flow the-
Cham, Switzerland: Springer, 2018, pp. 266–278. ory, autonomous vehicle control, traffic congestion
[40] M. Keyvan-Ekbatani, M. Papageorgiou, and V. L. Knoop, “Controller alleviation, and intersection control.
design for gating traffic control in presence of time-delay in urban road
networks,” Transp. Res. Procedia, vol. 7, pp. 651–668, 2015.
[41] M. Keyvan-Ekbatani, X. Gao, V. V. Gayah, and V. L. Knoop, Maha Elouni received the B.Sc. and M.Sc. degrees
“Traffic-responsive signals combined with perimeter control: Investi- in computer science from the National School of
gating the benefits,” Transportmetrica B, Transp. Dyn., vol. 7, no. 1, Computer Science, Tunisia, in 2012, and the M.Sc.
pp. 1402–1425, Jun. 2019. degree in applied mathematics from Virginia Tech
in 2015, where she is currently pursuing the Ph.D.
[42] T. A. M. Euzébio and P. R. Barros, “Optimal integral gain for smooth PI
degree with The Bradley Department of Electrical
control,” IFAC Proc. Volumes, vol. 46, no. 11, pp. 529–533, Jan. 2013.
and Computer Engineering. She works as a Graduate
[43] D. Singh, N. Singh, B. Singh, and S. Prakash, “Optimal gain tuning of Research Assistant with the Center for Sustainable
PI current controller with parameter uncertainty in DC motor drive for Mobility, Virginia Tech Transportation Institute. Her
speed control,” in Proc. Students Conf. Eng. Syst. (SCES), Apr. 2013, research interests include traffic control, traffic mod-
pp. 1–6. eling and simulation, and intelligent transportation
[44] R. Thangaraj, T. R. Chelliah, M. Pant, A. Abraham, and C. Grosan, systems.
“Optimal gain tuning of PI speed controller in induction motor drives
using particle swarm optimization,” Log. J. IGPL, vol. 19, no. 2,
pp. 343–356, 2011. Hossam M. Abdelghaffar received the B.Sc. degree
[45] R. Anandanatarajan, M. Chidambaram, and T. Jayasingh, “Limitations (Hons.) in electronics engineering from the Fac-
of a PI controller for a first-order nonlinear process with dead time,” ulty of Engineering, Mansoura University, Egypt,
ISA Trans., vol. 45, no. 2, pp. 185–199, Apr. 2006. the M.Sc. degree in automatic control system engi-
[46] S. W. Sung and I.-B. Lee, “Limitations and countermeasures of PID neering from Mansoura University, and the Ph.D.
controllers,” Ind. Eng. Chem. Res., vol. 35, no. 8, pp. 2596–2610, degree in electrical engineering from The Bradley
Jan. 1996. Department of Electrical and Computer Engineering,
Virginia Tech. He is currently an Assistant Professor
[47] M. Elouni and H. Rakha, “Weather-tuned network perimeter control—
with the Department of Computer Engineering and
A network fundamental diagram feedback controller approach,” in Proc.
Systems, Faculty of Engineering, Mansoura Uni-
VEHITS, 2018, pp. 82–90.
versity, and the Center for Sustainable Mobility,
[48] A. Kouvelas, M. Saeedmanesh, and N. Geroliminis, “Enhancing model- Virginia Tech Transportation Institute.
based feedback perimeter control with data-driven online adaptive opti-
mization,” Transp. Res. B, Methodol., vol. 96, pp. 26–45, Feb. 2017.
[49] B. Mirkin, J. Haddad, and Y. Shtessel, “Tracking with asymptotic sliding Hesham A. Rakha (Fellow, IEEE) received the
mode and adaptive input delay effect compensation of nonlinearly B.Sc. degree (Hons.) in civil engineering from Cairo
perturbed delayed systems applied to traffic feedback control,” Int. University, Cairo, Egypt, in 1987, and the M.Sc. and
J. Control, vol. 89, no. 9, pp. 1890–1903, Jul. 2016. Ph.D. degrees in civil and environmental engineering
[50] C. Edwards and S. Spurgeon, Sliding Mode Control: Theory and from Queen’s University, Kingston, ON, Canada,
Applications. Boca Raton, FL, USA: CRC Press, 1998. in 1990 and 1993, respectively. His research focuses
[51] J.-J. E. Slotine and W. Li, Applied Nonlinear Control. on large-scale transportation system optimization,
Upper Saddle River, NJ, USA: Prentice-Hall, 1991. modeling, and assessment. He and his team have
expanded the domain of knowledge (in traveler and
[52] V. I. Utkin and A. S. Poznyak, Adaptive Sliding Mode Control (Advances
driver behavior modeling) and developed a suite
in Sliding Mode Control: Lecture Notes in Control and Information
of multimodal agent-based transportation modeling
Sciences). Berlin, Germany: Springer, 2013.
tools, including the INTEGRATION microscopic traffic simulation software.
[53] F. Dion, H. Rakha, and Y.-S. Kang, “Comparison of delay estimates at This software was used to evaluate the first dynamic route guidance system,
under-saturated and over-saturated pre-timed signalized intersections,” TravTek in Orlando, FL, USA; to model the Greater Salt Lake City area
Transp. Res. B, Methodol., vol. 38, no. 2, pp. 99–122, Feb. 2004. in preparation for the 2002 Winter Olympic Games; to model sections of
[54] H. Rakha, Y.-S. Kang, and F. Dion, “Estimating vehicle stops at Beijing in preparation for the 2008 Summer Olympic Games; to optimize and
undersaturated and oversaturated fixed-time signalized intersections,” evaluate the performance of alternative traveler incentive strategies to reduce
Transp. Res. Rec., J. Transp. Res. Board, vol. 1776, no. 1, pp. 128–137, network-wide energy consumption in the Greater Los Angeles area; and to
Jan. 2001. develop and test an Eco-Cooperative Automated Control (Eco-CAC) system.
[55] H. Rakha, H. Yue, and F. Dion, “VT-meso model framework for esti- He and his team have developed various vehicle energy and fuel consumption
mating hot-stabilized light-duty vehicle fuel consumption and emission models that are used world-wide to assess the energy and environmental
rates,” Can. J. Civil Eng., vol. 38, no. 11, pp. 1274–1286, Nov. 2011. impacts of ITS applications and emerging connected automated vehicle (CAV)
[56] H. Rakha, K. Ahn, and A. Trani, “Development of VT-micro model for systems. The models include the VT-Micro, VT-Meso, the Virginia Tech
estimating hot stabilized light duty vehicle and truck emissions,” Transp. Comprehensive Fuel consumption Model (VT-CPFM), the VT-CPEM, and
Res. D, Transp. Environ., vol. 9, no. 1, pp. 49–74, Jan. 2004. VT-CPHEM models.

Authorized licensed use limited to: University of Canberra. Downloaded on July 26,2020 at 04:44:17 UTC from IEEE Xplore. Restrictions apply.