j o u r n a l o f t r a f fi c a n d t r a n s p o r t a t i o n e n g i n e e r i n g ( e n g l i s h e d i t i o n ) 2 0 2 3 ; 1 0 ( 1 ) : 8 6 e1 0 3

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.keaipublishing.com/jtte

Original Research Paper

Bi-objective optimization models for mitigating traffic

congestion in urban road networks

Haritha Chellapilla a,*, R. Sivanandan a, Bhargava Rama Chilukuri a,

Chandrasekharan Rajendran b
a
Transportation Engineering Division, Department of Civil Engineering, Indian Institute of Technology Madras, Chennai
600036, India
b
Department of Management Studies, Indian Institute of Technology Madras, Chennai 600036, India

highlights

Proposed four bi-objective optimal flow distribution (ODF) models, to lower traffic congestion on urban road networks.

OFD models consider system optimality along with congestion minimization.

Constrained OFD models ensure a fair routing strategy for users.

The proposed models outperform the existing user equilibrium and system optimal flow models.

article info abstract

Article history: Traffic congestion in road transportation networks is a persistent problem in major
Received 7 November 2020 metropolitan cities around the world. In this context, this paper deals with exploiting
Received in revised form underutilized road capacities in a network to lower the congestion on overutilized links
20 August 2021 while simultaneously satisfying the system optimal flow assignment for sustainable
Accepted 13 September 2021 transportation. Four congestion mitigation strategies are identified based on deviation and
Available online 21 January 2023 relative deviation of link volume from the corresponding capacity. Consequently, four bi-
objective mathematical programming optimal flow distribution (OFD) models are pro-
Keywords: posed. The case study results demonstrate that all the proposed models improve system
Traffic congestion mitigation performance and reduce congestion on high volume links by shifting flows to low volume-
Sustainability to-capacity links compared to UE and SO models. Among the models, the system opti-
Bi-objective optimization mality with minimal sum and maximum absolute relative-deviation models (SO-SAR and
Optimal flow distribution models SO-MAR) showed superior results for different performance measures. The SO-SAR model
Urban road networks yielded 50% and 30% fewer links at higher link utilization factors than UE and SO models,
respectively. Also, it showed more than 25% improvement in path travel times compared to
UE travel time for about 100 paths and resulted in the least network congestion index of
1.04 compared to the other OFD and UE models. Conversely, the SO-MAR model yielded the
least total distance and total system travel time, resulting in lower fuel consumption and

* Corresponding author. Tel.: þ91 72 07521915.

E-mail addresses: [email protected] (H. Chellapilla), [email protected] (R. Sivanandan), [email protected] (B.R. Chilukuri),
[email protected] (C. Rajendran).
Peer review under responsibility of Periodical Offices of Chang'an University.
https://doi.org/10.1016/j.jtte.2021.09.006
2095-7564/© 2023 Periodical Offices of Chang'an University. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co.
Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 87

emissions, thus contributing to sustainability. The proposed models contribute towards

efficient transportation infrastructure management and will be of interest to trans-
portation planners and traffic managers.
© 2023 Periodical Offices of Chang'an University. Publishing services by Elsevier B.V. on
behalf of KeAi Communications Co. Ltd. This is an open access article under the CC BY-NC-
ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

2. Literature review
1. Introduction
It is a common behavioral assumption that travelers choose
Urban traffic congestion in major cities worldwide has become routes between their origins and respective destinations,
a challenge to transportation planners and traffic managers. which they perceive to minimize their travel times. The
Many measures have been taken to manage the increased resulting flow pattern in which drivers cannot reduce their
travel demand by enhancing supply through metro systems, travel times by unilaterally changing to other routes is termed
flyovers, road widening, etc. However, such measures are as an equilibrium flow pattern. In a seminal contribution,
expensive, time-consuming, and are often faced with land Wardrop (1952) stated two principles for flow distribution on a
acquisition hurdles. To improve congestion, researchers road network, namely, user equilibrium (UE) and system
developed different measures, which include optimal traffic- optimal (SO). The SO traffic assignment minimizes system
light management (Mousavi et al., 2017), road congestion travel time but sometimes leads to unacceptably longer paths
pricing (Han et al., 2019), improve bus services (Gkiotsalitis, for some vehicles. So, it is reasonable to expect that rarely
2021; Zhang and Xu, 2017), route guidance system (Jahn et al., will users rarely follow the recommended paths for the
2000), etc. Recent studies suggest that road congestion pricing benefit of the community by sacrificing their own shortest
as a beneficial alternative for efficient traffic management paths. Thus, this assignment is behaviorally unrealistic. On
and to reduce congestion. However, we believe that the the other hand, UE is the most popular and widely used
congestion pricing strategy that is based on penalizing the traffic assignment, as it represents the natural behavior of
public for using certain facilities may not be attractive for the users to opt for their shortest paths to reach the desired
implementation in most cities. Furthermore, any congestion destination (Ben-Akiva and Lerman, 1985). Subsequently, to
improvement measure would not be effective without a solve this optimization problem and its extensions, efficient
proper route guidance strategy. Hence, traffic routing is an algorithms were developed (e.g., Beckmann et al. (1956) and
effective congestion mitigation measure, that improves the Sheffi (1985)). However, some authors unveiled that the UE
whole network performance rather than focusing on the based traffic assignment is inefficient. Braess (1968) initially
travel cost or travel time of individuals (Isa et al., 2015). discovered the paradox phenomenon (Braess paradox), where
Optimal utilization of existing road network supply is an adding a new road to a network with fixed demands
attractive proposition in this context. One of the strategies is increases the total travel time of the updated UE.
to examine the potential for re-routing the origin-destination Subsequently, many studies were conducted about this
(O-D) vehicular flows to capitalize on underutilized road links/ paradox behavior (Braess et al., 2005; Wang et al., 2016).
paths, aiming to reduce traffic congestion. An aerial snapshot There on, some studies attempted to measure the
of the road network at any instant, will show the level of traffic degradation of UE performance using criteria such as the
density on the links of the network and can reveal highly worst-case ratio (Koutsoupias and Papadimitriou, 1999) and
congested to low-flow links. In this context, one of the traffic price of anarchy (Correa et al., 2004; Papadimitriou, 2001).
management questions that arises is whether the underutil- The shortcomings of UE and SO principles have led to the
ized capacities of links in a road network can be exploited to need for integrated algorithms that pay attention to the sys-
mitigate congestion. With this motivation, congestion miti- tem-wide performance as well as user needs (Henry et al.,
gation strategies that would spread the volume among the 1991; Kaysi et al., 1995). However, many studies suggested
links by diverting flows from high volume to low volume links using an optimization strategy for routing the vehicles that
are worth examining. We propose four bi-objective Optimal minimizes global and community criteria with individual
Flow Distribution (OFD) models in this paper by coupling each needs treated as constraints (Beccaria and Bolelli, 1992). Jahn
congestion mitigation strategy with the System Optimal (SO) et al. (2000) first attempted to minimize the total
objective. Each model integrates the system perspective- experienced travel time while restricting the set of feasible
simultaneous minimization of total system travel time and paths, thus considering users' choice and later successively
congestion on links (by utilizing unused link capacities), and improved (Jahn et al., 2005) using U.S. Bureau of Public Roads
the user perspective-restricting O-D flows to K-shortest paths (BPR) (BPR, 1964) latency function. To meet global benefits,
based on Uesr-Equilibrium (UE) travel time. The proposed some authors investigated distinctive objectives, such as the
models could implicitly reduce emissions and enhance safety. maximum experienced travel time on any used path (Correa
We believe that these models have significant applicability et al., 2007) and the weighted geometric mean of the travel
and potential use in real-life road networks. times of used routes (Lujak et al., 2014). Recently, linear
88 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103

optimization models that minimize system congestion while determined based on the individual optimal values of the
reducing the inconvenience caused to the userswere objectives and the unfairness ratio (defined as the ratio of
proposed (Angelelli et al., 2016). Consequently, a linear experienced travel time to that of UE travel time). The former
programming-based approach was developed by adopting a measure ensures elimination of low values of l that lead to
piecewise approximation of the non-linear BPR travel time high congestion and the latter measure guarantees that the
function (Angelelli et al., 2018, 2021). link travel times obtained by the OFD model only marginally
The concept of the normal length of an arc, which is an a deviate from the corresponding UE travel times. We present
priori estimate of the arc travel time, is introduced to deter- the computational results on a real-world network (Sioux Falls
mine the set of feasible paths that are within an acceptable network) and analyze the performance of the proposed OFD
level of fairness for the users. Different normal lengths of arcs models (for recommended lÞ against the existing traffic
were employed in the models such as geographical distance assignment models (UE and SO). Subsequently, evaluation of
(Jahn et al., 2000), free-flow travel time (Angelelli et al., 2021; the OFD models based on different criteria reveals the supe-
Lujak et al., 2014), UE travel time (Jahn et al., 2005) and rior OFD model that offers significant potential in reducing
fastest path experienced travel time (Angelelli et al., 2018). congestion on links.
The efficiency and fairness studies of the constrained The remainder of this paper is organized as follows. The
system optimal model (Schulz and Stier-Moses, 2006) next section describes the existing optimization models and
showed that UE travel times as the normal length of the arc the proposed bi-objective OFD Models. In Section 4, different
promise greater fairness. It assigns users of the same O-D solution approaches for solving the proposed bi-objective
pair to paths with travel times that do not differ much. problem are discussed. Characteristics of the case study
Based on the literature, it is evident that while there has network are presented in Section 5. In Section 6, the
been considerable focus on system optimal, user equilibrium, methodology and evaluation of OFD models are discussed.
and via media solutions for traffic assignment problems, there The performance analysis and comparison of the proposed
has not been sufficient emphasis on congestion mitigation. OFD models against the existing models are given in
The latter is essential because it is often found that the flows Section 7. The final section presents the conclusions drawn
and delays on the transportation networks are poorly from the study.
balanced from a congestion perspective, with some links
excessively congested, while others are underutilized, leading
to inefficient use of transportation infrastructure. Also, the
congestion mitigation problem may not be addressed in 3. Optimization models
isolation, it needs to be coupled with travel time minimization
since the route advisory will be better followed if the proposed In this section, we present the mathematical formulation that
solution is only a slight deviation from the user equilibrium. forms the basis of this study. Consider a directed graph G ¼ (N,
Some of the issues that need attention in this regard include: A), where N represents a set of all nodes, and A⊆N  N rep-
suitable characterization of congestion-application specific resents a set of arcs. Each arc {u, v} e A denotes the road links
definition, acceptable deviation from the user equilibrium for on which vehicles travel, where u and v represent the up-
a compromise between competing objectives, weightage stream and downstream nodes, respectively; u, v e N and u s
given to congestion mitigation vs. travel time minimization, v. Traffic flow on a link a e A represents the number of vehicles
etc. Towards addressing these aspects, the paper proposes a entering a link a per unit time, denoted as Va . The capacity of
bi-objective approach that maximizes system benefits while each link a is denoted by Ca and represents the maximum
emphasizing congestion mitigation on links by exploiting number of vehicles that can traverse a link a per unit time. The
underutilized link capacities. To exploit the underutilized link link traverse time is defined for each link a e A by a latency
capacities in a network, we intend to minimize the gap be- function ta ðVa Þ; which depends on the arrival rate, Va . Each
tween the volume and capacity of the links in the network. link a is associated with a parameter, free-flow travel time,
This strategy would eventually reduce the flow on high vol- ttFF
a , which represents the time taken to traverse a link when
ume links and re-direct to low volume links. Towards this, we no other vehicle is present on the link and is defined by
a ¼ ta ð0Þ; ca e A. Also, we consider an origin-destination
ttFF
identify four congestion mitigation strategies for the utiliza-
tion of underused link capacities and thus propose four bi- pair by set OD. An O-D pair {i, j} e OD where i e O is the origin,
objective optimal flow distribution (OFD) models. Each OFD and the corresponding destination is j e D, O, D⊆N and i s j. A
model includes SO assignment as the first objective and a set of paths Kij is associated with each O-D pair. The time
unique congestion mitigation strategy as the second objective. taken to traverse a path p e Kij for each c{i, j} e OD is repre-
From the literature, it is also known that to ensure user p
sented by ttij and is defined by the summation of travel time
adherence to the system optimal model, considering UE travel
over all the links in the path. The number of trips (demand)
time as the normal length of the arc provides greater fairness
from a node i to node j is denoted by Tij , c{i, j} e OD. The traffic
for the users. In this regard, we restrict the feasible paths for
p
each O-D pair to K-shortest paths (l) based on UE travel times. flow on a path p e Kij for each O-D pair is represented by Tij , c{i,
P p pa
The suitable value of l for the bi-objective OFD models is j} e OD and is defined as Tij ¼ Tij . The indicator variable gij
p
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 89

takes a value one if link a e A belongs to path p between O-D this concept, we formulate the bi-objective optimization
pair ij and zero otherwise. models.

3.2. Proposed models

3.1. User equilibrium (UE) and system optimal (SO)
models In our models, along with SO traffic assignment, another
objective that mitigates the congestion on high volume links
3.1.1. UE model by utilizing the underutilized capacities of the low volume
Beckmann et al. (1956) formulated a mathematical model links in a network is incorporated. Towards this, the identified
(shown below) to compute link volumes based on the UE congestion mitigation strategies are described in Section 3.2.1.
principle. The user needs are addressed in the model by restricting the
Minimize flow between each O-D pair to a limited number of paths, as
described in Section 3.2.2.
XZ
Va

Z¼ ta Va dV (1)
a 3.2.1. Congestion mitigation strategies
0
To minimize congestion on links, we exploit underutilized
subject to capacities of links to redistribute volumes from overutilized
X X p pa links. The overutilized and underutilized links are defined as
Va ¼ Tij gij ca e A (2)
ði;jÞ e ODpeKij
follows.

X p Case 1. Ca > Va
Tij ¼ Tij cfi; jg e OD (3)
p e Kij 
Ca  V a overutilized  d
Ca  Va or ¼
Ca underutilized o=w
p
X pa
ttij ¼ ta Va gij cp e Kij andcfi; jg e OD (4)
a
Case 2. Va > Ca
p
Tij  0 cp e Kij andcfi; jg e OD (5) 
V a  Ca underutilized  d
Va  Ca or ¼
Ca overutilized o=w
Va  0 ca e A (6)
The motivation here is to spread the congestion among all
Constraint set (2) imposes that the total flow on a link a e A the links and not concentrate on a few (overutilized) links. In
is the summation of flows over all paths p e Kij passing through this regard, we aim to minimize the gap between the volume
it; in other words, they express the link flows in terms of path and capacity of the links in a network. Towards this, we pro-
flows. Constraint (3) represents the set of flow conservation pose the following four functions.
constraint, it guarantees that the summation of flow on all the
paths p e Kij connecting each O-D pair {i, j} e OD is equal to trip P
1. Sum of absolute deviation (SA): jCa  Va j
rate Tij . Constraint (4) enforces that the total travel time in a a
2. Maximum of absolute deviation (MA): maxjCa  Va j
path p e Kij equals the sum of travel time over all the links in a
P
the path. Constraints (5) and (6) are the non-negativity con- 3. Sum of absolute-relative deviation (SAR): jCaCV
a
aj

a
straints that define the domain of the decision variables and 4. Maximum of absolute-relative deviation (MAR): maxjCaCV
a
aj
a
the auxiliary variables, respectively.
In a large urban network with a wide range of link char-
3.1.2. SO model acteristics, the four congestion mitigation strategies have
The system optimal model is formulated mathematically as a their own advantages. For example, if j Ca Va j is in the order
minimization of total travel time experienced by the users of hundreds, then it is preferable if Ca and Va are in the order of
(Sheffi, 1985). thousands than in hundreds. Because, for the same absolute
Minimize difference between link volume and capacity, the absolute-
X X p p relative deviation is more in the latter case. Similarly, for the
Z¼ ttij Tij (7) same value of jCaCV aj
it is better if Ca and Va are of the order of
a
ði; jÞeODpeKij
hundreds than in thousands, as in the latter case could have a
subject to Constraints: (2)e(6). larger absolute deviation between link volume and capacity.
Note that the solution obtained from the user UE model is a Thus, we investigate both the absolute deviation and the ab-
feasible solution to the SO model, and therefore, the total solute-relative deviations of link volumes from their capac-
travel time experienced in UE is more than SO. This difference ities to reduce congestion. Towards the system perspective,
in total travel times reflects how a greedy route choice (UE) we model to minimize the total absolute or absolute-relative
affects the system performance compared to a system-wide deviation of link volume from its capacity. Additionally, we
coordinated traffic assignment (SO). However, UE is a behav- also investigate minimizing the maximum gap, as this would
iorally realistic model. To strike a balance between UE and SO, lead to balanced dispersion among all the links. Hence, we
a constrained system optimal model that ensures system analyze the absolute deviation and the absolute-relative de-
optimality with user needs constraints is an option. Based on viation of all the links using both the metrics.
90 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103

3.2.2. Optimal flow distribution (OFD) models Minimize

We propose four bi-objective OFD models, SO-SA, SO-MA, SO- X X p p
SAR, and SO-MAR. The first objective of each OFD model is SO, Z11 ¼ ði;jÞeOD peKlij
ttij Tij (8)
which is employed to minimize total system travel time. The
second objective is selected from each of the four congestion X
Z12 ¼ a
Da (9)
mitigation strategies, namely SA, MA, SAR, and MAR, to
improve congestion on links. The set of paths that are likely to subject to
result in acceptable levels of fairness for the users are
considered. We adopt UE travel time as normal length of the X X p pa
Va ¼ ði;jÞeOD peKlij
Tij gij ca e A (10)
arc, and based on the UE travel times, K-successive shortest
paths are determined for each O-D pair. The eligible path set is
X
denoted as Klij c{i, j} e OD where l indicates the number of Tij ¼ Tij
p
cfi; jg e OD (11)
peKlij
shortest paths considered for each O-D. Note that the set of
eligible paths Klij is the subset of Kij . This constraint eliminates X
cp e Klij andcfi; jg e OD
p pa
the longer duration paths from being recommended to the ttij ¼ a
ta Va gij (12)
users. The time taken to traverse a path p e Klij for each c{i, j} e
cp e Klij andcfi; jg e OD
p p
OD is represented by ttij and is defined by the summation of Tij  0 (13)
travel time over all the links in the path. Table 1 shows the
notations used in this study. Va  0 ca e A (14)

1. SO-SA model Da  C a  V a ca e A (15)

While the first objective (Z11 ) aims to minimize the total Da  V a  C a ca e A (16)
travel time in the network, the second objective (Z12 ) attempts
to minimize the total sum of the absolute deviations of link
volume from the capacity of the link. 2. SO-MA model

In this model, the second objective (Z22 ) minimizes the

Table 1 e Notations used in optimal flow distribution maximum absolute deviation of the link volume from the
models. corresponding capacity.
Notation Minimize
Set
N Set of nodes in the network X X p p
Z21 ¼ ði;jÞeOD peKlij
ttij Tij (17)
A Set of links in the network
OD Set of all O-D pairs
Indice
Z22 ¼ maxðDa Þ (18)
{i, j} An O-D pair, c {i, j} e OD a

{u, v} A link with starting node u and

subject to Constraints (10) and (16).
ending node v, c {u, v} e A
Kij Set of paths for O-D pair, c {i, j}
p A path for a given O-D pair, p e Kij 3. SO-SAR model
Decision variable
p
Tij Flow on path p between the O-D The second objective ðZ32 ) aims to minimize the total sum
pair {i, j} of absolute relative deviation of the link volume from the
Auxiliary variable corresponding capacity.
Va Flow on a link a,ca e A
p Minimize
ttij Cost of flow on path p between the
O-D pair {i, j} X X p p
Z31 ¼ peKlij
ttij Tij (19)
ta ðVa Þ Travel time on link a with flow Va ði;jÞeOD

Parameter
Tij Demand/total number of trips X
Z32 ¼ RDa (20)
between origin i to destination j a

Ca Capacity of an arc a, ca e A subject to Constraints: (10)e(14), and

pa
gij 1 if an arc a belongs to path p
between O-D pair {i, j};
0 otherwise. Ca  Va
Klij Set of all eligible paths for each O-D RDa  ca e A (21)
Ca
pair for a given l (l ¼ 1 denotes that
the flow between each O-D pair is Va  Ca
RDa  ca e A (22)
restricted to the first shortest path) Ca
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 91

4. SO-MAR model functions. Table 2 describes some of the essential methods to

solve multi-objective problems.
The second objective ðZ42 ) minimizes the maximum ab- We aim to generate as many Pareto optimal or non-domi-
solute relative deviation of the link volume from the corre- nated solutions as possible so that we will have continuous so-
sponding capacity. lutions, and the users can choose the solution based on their
Minimize requirement. Every point on the Pareto front represents
X X p p
different trade-offs between the objective functions and can be
Z41 ¼ ði;jÞeOD peKlij
ttij Tij (23) adopted according to the users' choice. We systematically
generate these non-dominated solutions by using the proposed
Z42 ¼ maxðRDa Þ (24) mathematical models. The techniques like goal programming
a
and lexicographic methods vary according to decision-makers'
subject to Constraints (10)e(14), (21) and (22). choice and have to be certain about the priority of the objectives.
Constraint set (10) imposes that the total flow on a link a e A The evolutionary algorithms are suitable for solving the multi-
is the summation of flows over all paths p e Klij passing through objective optimization problems, but they are computationally
it. Constraint (11) guarantees that the summation of flow on complex. Thus, we adopt weighted sum and e-constraint ap-
all the paths p e Klij connecting each O-D pair {i, j} e OD is equal proaches, as they are effective in generating the best compro-
to trip rate Tij . Constraint (12) ensures that the total travel time mising solutions along the Pareto front with less computational
in a path p e Klij is the sum of travel time over all the links in the effort (Chiandussi et al., 2012). In the weighted-sum approach
path. Constraints (13) and (14) are the non-negativity con- (WSA), by using different combinations of weights (w), we
straints that define the domain of the decision variables and deduce the feasible solutions. Similarly, using the e-constraint
the auxiliary variables, respectively. approach, a set of feasible solutions is obtained. Combining
the feasible solutions obtained by both the methods, the non-
dominated solutions are determined. The proposed
4. Solution approaches formulations based on WSA and e-constraint approach are
shown in sections 4.1 and 4.2, respectively. Section 4.3
There are several approaches to solve multi-objective opti- illustrates the process to determine the Pareto front from the
mization problems. Few approaches transform multiple feasible solutions. Note that depending upon the complexity of
objective problems into a single-objective problem, while the the road network and quality of solution requirement,
others simultaneously optimize the individual objective appropriate values of w and e values may be chosen.

Table 2 e Some methods to solve multi-objective problems.

Method Description
Weighted-sum approach (WSA) There are different techniques in this approach, such as equal weights rank
order centroid weights, and rank sum weights etc. By assigning weights to
each objective function, the multi-objective problems are transformed into
single objective models (e.g., Zadeh (1963)). The solutions obtained by
solving the single-objective problem with a different combination of
weights would give the Pareto optimal solutions for the multi-objective
problem. This technique is adopted by Rassafi et al. (2013) to solve the
proposed multi-objective transportation problem.
e-constraint method One of the objectives is optimized, while the other objectives are treated as
constraints bounded by a pre-defined parameter e (e.g., Marglin (1967)). By
solving adequate number of single-objective problems, the Pareto front for
the multi-objective problem is obtained. This is one of the widely used
methods and can identify an efficient Pareto front independent of its shape
(Chiandussi et al., 2012).
Goal programming The decision-maker sets a priori targets/goals for each objective. All
objectives and constraints are treated as constraints (e.g., Charnes et al.
(1955)). The weighted sum of deviations concerning the pre-specified
targets of each objective is minimized.
Lexicographic method The objectives are ranked in the order of preference (e.g., Islam et al. (1997)).
The optimization is stated with the most preferred objective function in
such a way that when the ithobjective function is optimized, the optimal
amounts of i1 objectives are treated as constraints.
Multi-level programming techniques The non-dominated or the Pareto optimal solutions are generated by
simultaneous optimization of individual objective functions (for instance,
using multi-objective evolutionary algorithms (MOEA)). The first MOEAs
were developed in the 1990s. Chira et al. (2015) adopted MOEA to solve the
traffic assignment problem.
92 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103

4.1. Formulations based on the weighted-sum approach

(WSA)     Z  Z* 
Zs1  Z*s1 s2
w1 þ 1  w1 s2
(27)
Zs1 jZ*s2  Z*s1 Zs2 jZ*s1  Z*s2
The WSA incorporates multiple objective functions into the
scalar form. In this method, normalization of objectives is subject to Constraints: (10) and (16) for s ¼ 1, 2
necessary to provide a sense of fairness between the objec- Constraints: (10)e(14), (21) and (22) for s ¼ 3, 4
tives so that they are treated equally in the model. Each
objective of the OFD model is normalized as follows. 4.2. Proposed e-constraint approach
Zs1  min Zs1
Objective 1 ¼ (25)
max Zs1  min Zs1 The e-constraint method is an intuitively appealing approach
to solve multi-objective optimization problems. The e values
Zs2  min Zs2 for each OFD model are calculated using Eqs. (28) and (29).
Objective 2 ¼ (26)
max Zs2  min Zs2 Zs1 jZ*s2  Z*s1 
2s1 ¼ c s e 1; 2; 3; 4 (28)
where the subscript s varies from 1 to 4, corresponding to M
each OFD model. Zs1 and Zs2 are the two objective functions of
OFD models. Min Zs1 and min Zs2 are the minimum values Zs2 jZ*s1  Z*s2 
2s2 ¼ c s e 1; 2; 3; 4 (29)
(optimal values) of each objective. max Zs1 and max Zs1 are M
the maximum values of each objective. We obtain the min- where M denotes the pre-determined number of intervals,
imum/optimal values for each objective by minimizing each chosen in such a way that a sufficient number of solutions are
objective independently without considering the effect of obtained.
another objective. The corresponding value of the other We generate feasible solutions by employing e-constraint
objective is regarded as its maximum value, as detailed technique in two phases, as illustrated below.
below. Phase 1: Minimize Zs1 while Zs2 is incremented by 2s2 and
The optimal value of the first objective of each model is treated as a constraint.
determined as follows. Step 1. Initialization: set iteration counter m ¼ 0
Minimize Zs1 Minimize Zs1
subject to Constraints: (10) and (16) for s ¼ 1, 2 subject to Zs2 ¼ Z*s2 and
Constraints: (10)e(14), (21) and (22) for s ¼ 3, 4 Constraints: (10) and (16) for s ¼ 1, 2
We denote the value obtained by minimizing Zs1 and the Constraints: (10)e(14), (21) and (22) for s ¼ 3, 4

corresponding Zs2 value as ½Z*s1 ; Zs2 Z*s1 , where, Z*s1 is the Set ðZs1 Þ0 ¼ Zs1 and ðZs2 Þ0 ¼ Z*s2 .

optimal value or minimum value of Zs1 and Zs2 Z*s1 is the value Note that ‘0’ denotes the iteration counter.
that Zs2 takes when Zs1 is optimal. Since both objectives are Step 2. Do the following for 0 <m  M
formulated as minimization functions, if we deviate from the {
optimal values of one objective, the other objective improves Step 2.1 Minimize Zs1
and moves towards its optimal solution. Note that any value subject to Zs2  ðZs2 Þm1 þ 22s , and
 Constraints: (10) and (16) for s ¼ 1, 2
of Zs2 greater than Zs2 Z*s1 can be ignored as it cannot improve
 Constraints: (10)e(14), (21) and (22) for s ¼ 3, 4
Zs1 any further. Thus, Zs2 Z*s1 is regarded as the maximum
Step 2.2 Set ðZs1 Þ0 ¼ Zs1 (obtained from Step 2.1)
value or the upper limit of Zs2 in our analysis.
Step 2.3 Minimize Zs2
The optimal value of the second objective of each model is
subject to Zs1 ¼ ðZs1 Þ0 , and
determined as follows.
Constraints: (10) and (16) for s ¼ 1, 2
Minimize Zs2
Constraints: (10)e(14), (21) and (22) for s ¼ 3, 4
subject to Constraints: (10) and (16) for s ¼ 1, 2
Step 2.4 Set ðZs1 Þm ¼ Zs1 (obtained from Step 2.1) and,
Constraints: (10)e(14), (21) and (22) for s ¼ 3, 4
ðZs2 Þm ¼ Zs2 (obtained from step 2.3)
Similarly, we denote the solution obtained by minimizing
}
Zs2 and the corresponding Zs1 value as ½Zs1 jZ*s2 ; Z*s2 , where, Z*s2 Note that after the execution of step 2, ðZs1 ÞM ¼ Z*s1 .
indicates the optimal value of Zs2 and Zs1 jZ*s2 is the value that Phase 2: The above steps 1 and 2 are repeated, considering
Zs1 takes when Zs2 is optimal and is considered as an upper the following.
limit of Zs1 in our analysis. Minimize Zs2 while Zs1 is incremented by 2s1 and treated
The weighted sum of the normalized forms of Zs1 and Zs2 as a constraint instead of minimizing Zs1 while treating Zs2 as
for each OFD model is formulated as given in Eq. (27). The constraint.
weights for each objective are denoted by w1 and w2 ,
respectively, where w2 ¼ 1  w1 . By varying w1 (between 4.3. Pareto front
0 and 1), the feasible solutions are generated for each OFD
model. The set of feasible solutions consist of dominated and non-
Minimize dominated solutions. However, the non-dominated solutions
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 93

solutions, and we represent the decision vector as ui ¼ [Zis1 , Zis2 ],

where ie {1, 2, $$$, l}. For a minimization problem, the
dominated and non-dominated solutions are defined as
follows.

1. A solution u1 e F is said to dominate u2 e F, if u1 is no worse

than u2 in both the two objectives, with strictly better than
u2 in at least one objective. Then, u2 is a dominated
solution.
2. A solution u e F is said to be non-dominated if and only if
there is no u0 e F for which u0 dominates u.

The feasible solutions obtained from weighted-sum and

e-constraint approaches are combined, and we determine the
Pareto front of non-dominated solutions for the four OFD
Fig. 1 e Non-dominated solutions, dominated solutions, models.
and Pareto front.

5. A case study
are the points of interest as they contain only the solutions
that cannot be improved in any of the objectives without A mid-sized network of Sioux Falls (SF), South Dakota, USA
sacrificing the other objective. The boundary defined by the set (Fig. 2), commonly used in the literature is considered for this
of all points mapped from the non-dominated solutions is case study. This network consists of 24 nodes and 76 links
called the Pareto front (Fig. 1). Let F be the feasible space with l forming 528 O-D pairs (each node is considered as origin and

Fig. 2 e Sioux Falls network (Zhou et al., 2012).

94 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103

destination). The total trips between various O-D pairs are

360,600 veh/h, besides 24 nodes and 76 links. Detailed
network parameters, O-D demand, and link attributes are
sourced and adopted from the transportation network
repository (Transportation Networks for Research Core
Team, 2019).
The impedances of arcs at different congestion levels are
measured using the link performance function or the latency
function. This function captures the relationship between
travel time per unit distance and volume per unit time on the
links and is generally considered to be differentiable, non-
decreasing, and function of flow on the link (Branston, 1976;
Fig. 3 e Percentage increment in optimal values of
Cohen, 1991; Sheffi, 1985). The most popular latency function
objective functions with the increase in l.
that put forward by U.S. Bureau of Public Roads (1964),
 b
Va
a 1 þ a Ca
ta ðVa Þ ¼ ttFF is used in our work with parameter

values a ¼ 0.15 and b ¼ 4 (LeBlanc et al., 1975). improvement in individual objective functions could be ach-
ieved by increasing l > 4.
The choice of the appropriate value of l for each OFD model
6. Computational study should be investigated based on the behavior of both the ob-
jectives of the OFD models. However, the above results show
We employ UE traffic assignment to get the prior estimates of that for l less than 3, neither of the two objectives would yield
link travel times for the given trip demands. Subsequently, we desired solutions as the optimal values are high, indicating
generate K-shortest paths for each O-D pair using Yen's algo- that the eligible number of paths for each O-D pair is insuffi-
rithm (Yen, 1971). The set of all eligible paths for each O-D pair cient. Towards this, we solve the four OFD models by
is denoted by Klij , where l is an indicator of the number of restricting the eligible path for each O-D pair (l) to 3 and 4. We
shortest paths considered for each O-D {i, j} pair. For example, investigate the better value of l that satisfies both our objec-
K1ij indicates that the demand for each O-D pair is restricted to tives of minimal total system travel time and optimal capacity
only one path, the one with the least travel time. Initially, we utilization of links in the subsequent sections.
analyze the variation in optimal values of each objective, Z*s1
and Z*s2 cs e {1, 2, 3, 4} with the increase in l, to determine
the maximum number of eligible paths (l) to be generated for 6.1. Feasible solutions and Pareto front
each O-D pair. The percentage improvement in values of Z*s1
and Z*s2 corresponding to the increase in the number of We make use of general algebraic modeling system (GAMS)
eligible paths from Klij to Klþ1 ij is calculated using the Eqs. 30 optimization software to solve the four bi-objective OFD
and 31, respectively. Since the optimal value of an objective is models by restricting the eligible paths for each O-D pair to 3
independent of other objectives and since the first objective and 4 shortest paths. In WSA, the value of the weighted
of all the OFD models is SO, the optimal value of SO objective parameter w1is initially set to 0 and increased by 0.002 at each
(Z*s1 cs e f1; 2; 3; 4g) is the same for all OFD models for a given l iteration until its value reaches 1. Thus, we generate 501
. Fig. 3 summarizes the results. feasible solutions by varying weighted parameter w1. Conse-
It can be noted that with the increase in l from 1 to 2, every quently, in the e-constraint approach, the value of M (pre-
objective improves by more than 20%. Similarly, with the in- determined number of intervals) is taken as 100, and from
crease in l from 2 to 3, we can observe the minimum and Phase 1 and Phase 2, a total of 200 feasible solutions are ob-
maximum improvements of 0.6% and 46.3% for Z*s1 c s e {1, 2, tained. Consequently, by combing the solutions obtained from
3, 4} and Z*22 , respectively. But, with the increase in l from 3 to both the models, we get a total of 701 feasible solutions. A
4, there is no improvement in the objective value of Z*42 and
p
vector of the decision variable Tij and the auxiliary variables
p
the maximum improvement of only 5% (Z*12 ) is observed. Va , ttij and ta ðVa Þ corresponding to the feasible solutions are
Hence, we can reasonably assume that no significant noted.

   
Z*s1 Klij  Z*s1 Klþ1
ij
Percentage improvement in Z*s1 ¼   c s e f1; 2; 3; 4g (30)
Z*s1 Klij

   
Z*s2 Klij  Z*s2 Klþ1
ij
Percentage improvement in Z*s2 ¼   c s e f1; 2; 3; 4g (31)
Z*s2 Klij
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 95

Fig. 4 e Non-dominated solutions obtained for the four OFD models with l ¼ 3. (a) SO-SA. (b) SO-MA. (c) SO-SAR. (d) SO-MAR.
Note: LE: left extreme, LD: least distance, and RE: right extreme.

The Pareto set of sub-optimal solutions (non-dominated where Zls1 , Zls2 are the coordinates of a point on the Pareto
solutions) is determined from the 701 feasible solutions. Figs. front.
4 and 5 show the Pareto fronts obtained for the four OFD It can be observed that lower increment in system optimal
models with l equals 3 and 4, respectively. Three coordinate travel time is observed when passing from LE to LD than when
points on the Pareto front are highlighted in these figures, passing from LD to RE. This means that passing from LE to LD,
namely, left extreme (LE), right extreme (RE) and least distance a small compromise on system minimal travel time would

(LD) points. LE and RE correspond to the points ½Z*s1 ; Zs2 Z*s1  and produce greater improvement in link congestions, as the
* *
½Zs1 jZs2 ;Zs2 , respectively, and the LD is the closest point to the second objective of each OFD model is improving. Table 3
origin. Note that for multi-objective optimization problems, a reports the values of the two objective functions Zs1 and Zs2
point close to Utopia point (an infeasible solution that mini- c s e {1, 2, 3, 4} at LE, RE and LD points.
mizes both the objectives individually [Z*s1 ; Z*s2 ]) is the best As mentioned in Section 6, for a given l; the optimal Z*s1 is
compromise solution satisfying both objectives (Ozcelebi the same across different models. However, by virtue of the
et al., 2007) and is considered as optimal value (Gunantara, definition of the second objective of each OFD model, the

2018). To find this least distance (LD) point, the objective other values (Zs2 Z*s1 , Zs1 jZ*s2 , Z*s2 , Zs1 and Zs2 ) vary and lead to
values are rescaled to an interval [0, 1], and the point with different solutions. The values of almost all the objectives
the shortest Euclidean distance from the origin is are improving with the increase in l from 3 to 4. However,
determined. We calculate the Euclidean distance of point l the values of Z11 jZ*12 , Z31 jZ*32 , Z11 and Z31 are better when l is
from the origin using the following Eq. (32). Literature 3 than l is 4. The lower objective function values when l
suggested the use of LD point for evaluating the equals 3 than when it is 4, show that, in the case of bi-
performance of the models (Rassafi et al., 2013). Henceforth, objective models, the system optimal objective may
we select the LD solution to evaluate the suitable value of l sometimes worsen for higher values of l.
for each OFD model and then analyze the results of the OFD
models.
6.2. Choice of eligible paths (l) for each O-D pair
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
!2 !2ffi
u
  u Z l
 Z *
Z l
 Z *
Distance Zls1 ; Zls2 ¼ t We adopt the unfairness concept (Jahn et al., 2005) to arrive at
2
s1 s1
þ s2 s2
(32)
Zs1 jZ*s2  Z*s1 Zs2 jZ*s1  Z*s2 an appropriate number of eligible paths (l) for each OFD
model. When no route guidance system exists, users tend to
96 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103

Fig. 5 e Non-dominated solutions obtained for the four OFD models with l ¼ 4. (a) SO-SA. (b) SO-MA. (c) SO-SAR. (d) SO-MAR.
Note: LE: left extreme, LD: least distance, and RE: right extreme.

opt for their shortest travel time path between O-D pairs, and respectively. Similarly, for the SO-MAR model with l equals 3
the UE traffic assignment captures this behavior. To ensure and 4, the KS statistic and p-value are obtained as 0.05 and
user acceptance of the proposed models, the travel time on 0.99, respectively. The p-values >0.05 (and KS statistic values
the links given by the OFD models should not be much are low) in both the cases indicate that we cannot reject the
worse than that of the UE model. Hence theunfairness null hypothesis that the distributions of unfairness factors for
factor, which is the ratio of experienced travel time of each the two models at l equals 3 and 4 are the same. Since the
link obtained based on the OFD model (for given l) to that of models result in similar unfairness values for l equals 3 and 4,
UE travel time, is calculated. indicating that the SO-MA and SO-MAR models may not be
Fig. 6 shows the cumulative percentile frequency of effectively using the new routes (the fourth route for each O-D
unfairness factors for each OFD model (at respective LD pair). Hence restricting the eligible paths to 3 is also ideal for
points) for l equals 3 and 4 and the corresponding 99th SO-MA and SO-MAR models. Accordingly, for all the four OFD
percentile unfairness factor values. The lower values of 99th models (for SF network), we recommend the value of 3 for l.
percentile unfairness factors occurred at l ¼ 3 for SO-SA, The performance of each model is analyzed at l ¼ 3 in the
and SO-SAR models, and at l ¼ 4 for SO-MA and SO-MAR following section.
models. Additionally, an interesting detail that is observed is
the cumulative distribution curves of SO-MA, and SO-MAR
models for l equals 3 and 4 overlap with each other, 7. Performance analysis of OFD models
indicating similar distributions.
We conduct the Kolmogorov-Smirnov (KS) test to investi- The performance of each OFD at l ¼ 3 considering respective
gate whether these distributions are identical or not. The KS LD points is evaluated using different criteria, and the results
statistic and p-value of unfairness factor values for the SO-MA are compared with the existing user equilibrium (UE) and
model for l equals 3 and 4 are found to be 0.12 and 0.80, system optimal (SO) models in the following sections.
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 97

Table 3 e Objective function values at left extreme (LE), right extreme (RE), and least distance (LD) points obtained for each
OFD model.
OFD model LE point RE point LD point

Z*s1 (veh-min.) Zs2 Z*s1 Zs1 jZ*s2 (veh-min.) Z*s2 Zs1 (veh-min.) Zs2
# # #
l¼3 SO-SA 7,226,848 382,200 8,223,626 327,228 7,444,702 339,954
SO-MA 7,226,848 18,999 8,995,381 9121 7,715,436 12,282
SO-SAR 7,226,848 45.25 9,705,449 39.52 7,750,124 40.69
SO-MAR 7,226,848 1.58 7,840,495 0.91 7,391,819 1.13
l¼4 SO-SA 7,209,352 379,112 9,074,361 310,128 7,563,736 324,615
SO-MA 7,209,352 18,849 8,728,340 8921 7,623,415 12,141
SO-SAR 7,209,352 44.99 11,175,380 38.46 7,957,594 39.57
SO-MAR 7,209,352 1.57 7,791,665 0.91 7,367,536 1.13

Note: The subscripts s ¼ 1, s ¼ 2, s ¼ 3, and s ¼ 4, correspond to SO-SA, SO-MA, SO-SAR, and SO-MAR models, respectively. # For SO-SA and SO-
 
MA models, units of Zs2 Z* , Z* , and Zs2 is veh/h and for SO-SAR and SO-MAR models, the objective functions Zs2 Z* , Z* , and Zs2 are unitless, as
s1 s2 s1 s2
Va  Ca
they are ratios of .
Ca

Fig. 6 e Cumulative percentile frequency vs. utilization factor for l equals 3 and 4. (a) SO-SA model. (b) SO-MA model. (c) SO-
SAR model. (d) SO-MAR model.
98 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103

even if the minimum value of Xa obtained by SO-MAR is less

Table 4 e Comparison of descriptive statistics of Xa of
than the SO model, it reported the least maximum value of
links across different models.
Xa (followed by the SO-SAR model) compared to other OFD,
Descriptive OFD model UE SO
UE, and SO models. The maximum values of Xa obtained by
statistics
SO-SA SO-MA SO-SAR SO-MAR SO-SA and SO-MA are higher than UE and SO models, but
Mean 1.47 1.49 1.43 1.45 1.47 1.46 the difference is less. This indicates that to distribute the
Median 1.52 1.57 1.47 1.56 1.60 1.53 flow among the links, OFD models are not resulting in Xa
Std. 0.48 0.50 0.50 0.51 0.58 0.50 values that are much worse than UE and SO models. In
Min. 0.39 0.53 0.31 0.24 0.17 0.29 conclusion, all the models outperformed the UE model.
Max. 2.59 2.60 2.53 2.13 2.56 2.57
Whereas SO-SA, SO-SAR, and SO-MAR models showed better
results compared to the SO.
7.1. Link utilization factor
7.1.1. Congestion levels across links
The link utilization factor (Xa ) Transportation Research Board The present study focusses on minimizing congestion on links
(1994) is defined as the ratio of link volume to the by spreading the flows among all the links. Consequently, an
corresponding capacity. OFD model should result in a higher number of links in lower
congestion levels and a lesser number of links in higher
Va
Xa ¼ (33) congestion levels compared to UE and SO traffic assignments.
Ca
We group the Xa values into five congestion levels. The fre-
where Xa is the link utilization factor for a link a. Va is the quency distribution plots in Fig. 7 summarizes the number of
volume on the link a, and Ca is the capacity of link a. The link links in each congestion level.
utilization factor values for the four OFD models (l ¼ 3 at LD In Xa  1.5 congestion level, SO-SA, SO-MA, and SO-SAR
Point), UE, and SO models are determined. Table 4 show 35, 36, and 39 links respectively, which are higher
summarizes the descriptive statistics of Xa values across numbers 30 and 34 links obtained from UE and SO models,
different models. respectively. However, the SO-MAR model yielded only
The mean and median values describe the data with a 27 links with Xa  1.5. On the other hand, the number of links
single value. In the present context, lower the values of mean with Xa > 2 is less for SO-SA, SO-SAR and SO-MAR models
and median of Xa , better is the model. From Table 4, it is (8, 7, and 7 links respectively), compared to 14 and 10 links
evident that, compared to UE and SO models, the SO-SAR obtained from UE and SO models, respectively. Additionally,
model yields low mean and median values. And, either the SO-MAR model projects no link with Xa > 2.5 while all the
mean or median is less in the case of SO-SA and SO-MAR other OFD, UE, and SO models result in 2 links each in this
models compared to UE and SO models. The standard congestion range. The SO-MA model performs poorly on this
deviation of Xa obtained from OFD models is much less than network as it results in the highest number of links (11 links)
the UE model. The lower deviation of Xa values among the with Xa >2 compared to other OFD, UE, and SO models. Hence,
links depict that the OFD models yield better flow SO-SA, SO-SAR, and SO-MAR models achieve better results
distribution among the links, and these models are not than the traditional UE and SO models in terms of reducing
penalizing some links with high volume while others have the number of highly congested links (Xa > 2).
less volume. The minimum value of Xa is more for the OFD Further, the performance of OFD models is evaluated based
models (except for SO-MAR) compared to UE and SO models, on the experienced travel times of used paths between various
which shows that these models utilize the links with lower O-D pairs to assess the improvement/deterioration in path
volume to divert the flow of high-volume links. However, travel times of OFD models with respect to the UE model.

Fig. 7 e Comparison of the number of links in each congestion level for different models.
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 99

Fig. 8 e Number of used path in the given percentage range of UE travel time.

p
7.2. Path travel time variations models result in lesser number of paths with PDTij  25%. The
SO-SAR model, on the other hand, yields more number of
p
We analyze the performance of the OFD models based on the paths in PDTij range (25%, 50%) but also reports more number
p
path travel times between each O-D pair. We compute the of paths (101 paths) with PDTij e25%, indicating more than
percentage deviation of path travel times of an O-D pair from 25% improvement in path travel times from UE travel time for
the corresponding UE travel time as follows. 101 paths (highest among the OFD models). The higher figures
p
of the SO-SAR model in PDTij range (25%, 50%) could be
p
p
ttij  ttðUEÞij because this model utilized the highest number of paths in
PDTij ¼ *100 (34)
ttðUEÞij distributing the demand across the O-D pairs. Hence, even if
p some paths are penalized, congestion could be improved by
where PDTij is the percentage deviation in travel time of a path
p this model as the flow is distributed among a greater number
p between O-D pair ij from UE travel time, ttij is the experi-
of paths. Network congestion is qualified in the following
enced travel time of a path p between O-D pair ij, ttðUEÞij is the
section to further analyze the congestion caused by OFD
UE travel time between O-D pair ij.
models.
The percentage deviations of travel times of used paths by
the OFD models from the UE travel times are grouped into six
7.3. Network congestion index (NCI)
bins varying from e50% to >50%. Fig. 8 shows the total
number of used paths by each OFD model and the frequency
p Network congestion is computed based on the experienced
distribution of paths in the six bin ranges. The PDTij values
travel time on links obtained from OFD models, using the
less than 0% indicate that the experienced travel time of that
following expressions (Richardson and Taylor, 1978) given in
path is less than the corresponding UE travel time. The OFD
Eqs. (35) and (36). Fig. 9 shows the values of NCI obtained for
models show improvement in terms of travel time for some
OFD, UE, and SO models.
paths while other paths are penalized with more travel time
than UE travel time. This trend, which is generally observed Ta  FFTa
CIa ¼ (35)
in the SO model, is found in OFD models since one of the FFTa
objectives of the OFD models is system optimality.
P
SO-MA model underperforms compared to other OFD ðCIa La Þ
models, as it results in a higher number of paths (53 paths) NCI ¼ a P (36)
p La
with PDTij more than 50%. Conversely, SO-SA and SO-MAR a

Fig. 9 e Network congestion index obtained by different models.

100 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103

where CIa is the congestion index for a link a, NCI is the SO-SA, and SO-MAR models successively out-performed the
network congestion index, Ta is the experienced travel time existing UE and SO models in terms of congestion mitigation
for link a, FFTa is the free-flow travel time for link a, La is the by spreading the flow from high volume links to low volume
length of link a. links. The SO-SAR model exhibits superior results by
The OFD models are formulated as via-media solutions reporting the least number of links with Xa (1.5, 2) and Xa > 2
between UE and SO models to improve congestion on highly (30 and 7, respectively) and the highest number of links (39)
congested links. Thus, network travel time variation is ex- with Xa < 1.5, compared to other OFD, UE, and SO models.
p
pected to be in between UE and SO models, which is evident Additionally, from Fig. 8, the evaluation in terms of PDTij (O-D
from the network CI values shown in Fig. 9. The NCI values path travel times) reveals that the OFD models (except the
p
of the three OFD models, namely, SO-SA, SO-SAR, and SO- SO-MA model, which reports 53 paths with PDTij > 50%)
MAR, are in between the values obtained from UE and improve travel time of some paths compared to UE travel
SO models. However, among the OFD models, the SO-SAR time. SO-SAR, SO-MAR, and SO-SA models yield 101, 47, and
p
model yields the least NCI value of 1.04, which is slightly 19 paths, respectively, with PDTij <e25%, indicating that
more than the SO model (1.02) but much lesser than the these paths show more than 25% improvement in O-D path
UE model (1.13). travel times compared to the UE model. Finally, the network
congestion index (Fig. 9) shows that the SO-SAR model results
7.4. Sustainability related objectives in the least network-wide congestion (NCI value of 1.04),
compared to other OFD models. Additionally, SO-SA and SO-
Fuel consumption in a transportation network is an important MAR models also improve network congestion (NCI value of
factor for sustainability since it is directly related to the 1.07), except the SO-MA model, which results in the same NCI
quantity of emissions. One of the popular models to estimate value as the UE model (1.13). Consequently, from Table 3
gasoline consumption in urban conditions was proposed by a (l ¼ 3 at LD points), the total system travel time per hour
group of researchers from General Motors (Evans and Herman, (TSTT) obtained by the SO-MAR model is 73,91,819 veh-min
1978) which was used in several transportation applications (Z41 value), which is the least among all the OFD models,
(e.g., SIDRA Intersection software (Solutions, 2012)). It relates followed by the SO-SA model with a TSTT of 74,44,702 veh-
gasoline consumption as a linear function of distance min. Note that TSTT obtained for the traditional UE model is
traveled and the travel time, given as 74,80,225 veh-min (for SF network with the same link
attributes). Hence, at the best compromise solution point (LD
E ¼ k1 L þ k2 T (37) point), SO-MAR and SO-SA models are no worse than the UE
where E is the gasoline consumption (ml), L is the distance model. Additionally, since the performance of SO-SAR and
traveled (km), T is the travel time (s), and k1 and k2 are vehicle SO-MAR are comparable with regard to congestion mitigation,
related parameters. This model is shown to perform well for they are evaluated for fuel consumption. It is found that the
speeds under 40 mile/h (Daganzo and Newell, 1995) which SO-MAR model performs better in terms of sustainability.
matches well with the speeds observed in the network. Table 6 shows the rank of OFD models under various criteria.
Since the performance of SO-SAR and SO-MAR are Thus, from this study on the SF network, the SO-SAR model
comparable with regard to congestion mitigation, they are gives superior results compared to the other OFD models
evaluated on the total distance traveled and total travel time (considering the respective LD points) in terms of Xa values, O-
given by Eqs. 38 and 39. D path travel times, and NCI value. The better performance of
X the SO-SAR model is because the second objective (SAR) of the
Total distance travelled ðLÞ ¼ a
Va La (38) SO-SAR model is a relative measure of central tendency as it
considers the minimization of the sum of absolute relative
X PjCa Va j
Total system travel time ðTSTTÞ ¼ a
Va Ta (39) deviation of Va from the corresponding Ca : Ca or
a
P 

where Va is the volume on the link a, Ta is the travel time on 1  VCaa . Hence, by virtue of the definition of this model, it
a
link a, and La is the length of link a. The SO-SAR (SO-MAR)
Va
resulted in L and TSTT of 36,71,346 veh-km (35,57,953 veh-km) gives minimum Ca ratio, i.e., Xa . Similarly, the network
and 77,50,125 veh-min (73,91,820 veh-min), respectively. congestion index, which is evaluated in terms of link travel
Given that fuel consumption is a weighted function of dis-
tance and travel time, SO-MAR results in lower fuel con-
Table 6 e Ranking of OFD models in each evaluation
sumption and emissions, and hence superior in terms of
criterion.
sustainability objectives related to economic and environ-
mental benefits. Model Evaluation criterion
Total system Xa O-D path NCI
7.5. Discussion travel time travel time
SO-SA 2 2 3 2
In the earlier sections, the performance of the OFD models SO-MA 3 4 4 4
(considering the respective LD points) is evaluated based on SO-SAR 4 1 1 1
SO-MAR 1 3 2 2
different criteria. The V/C analysis (Fig. 7) shows that SO-SAR,
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 101

 4 model is estimated based on the individual optimal values of

Va
times, is a function of Ca ratio (ta ¼ ttFF Va
a 1 þ 0:15 Ca ), thus,
the objectives and the unfairness ratio. The former measure
the NCI value is also the least for this model. One of the rea- ensures that the low values l that lead to high congestion are
sons for the superior performance of the SO-SAR model in excluded, and the latter measure ensures that the link travel
terms of O-D path travel times could be the use of more paths times obtained by the OFD model only marginally deviate
compared to other OFD models allowing a better spread of from the corresponding UE travel times. We found that for
demand between the O-D pairs. Thus, in the SO-SAR model, the test network, l ¼ 3 performs better for all the OFD models.
both objectives complement each other for mitigating A comparative study against the traditional models (UE
congestion, and hence, its LD point emerges to be superior to and SO) shows that the three OFD models, SO-SA, SO-SAR, and
the corresponding LD points of other OFD models. SO-MAR (considering the respective LD points) perform better
On the other hand, the SO-MAR model gives the best per- in terms of volume distribution among the links with little
formance with respect to total system travel time and total compromise in travel times. However, in view of congestion
distance travelled. The gap between the value of Z*s1 (LE) and mitigation the SO-SAR model shows significantly superior
Zs1 jZ*s2 (RE) (Table 3) corresponding to the SO-MAR model is the results, which is evident from the Xa values, O-D path travel
least, compared to the gaps between the LE and RE solutions of times, and NCI value reported by this model. It exhibits greater
the other models. Hence, the LD solution given by the SO-MAR potential in minimizing the highly congested links by
model provides the least TSTT than the LD solutions from exploiting the underutilized capacities of low volume links
other models. Thus, the SO-MAR model is superior in terms compared to other OFD, UE, and SO models. Additionally, it
of sustainability related objectives, such as fuel consump- also offers better distribution in terms of percentage deviation
p
tion (a weighted function of distance and travel time). in travel times from UE travel time (i.e., PDTij ). The SO-SAR
p
Therefore, for operational metrics involving congestion model yields more paths with PDTij <e25%, compared to other
mitigation, one may choose the SO-SAR model and the SO- OFD models, indicates that around 100 paths (for different O-D
MAR model for sustainability benefits. pairs) would experience better travel time than that of the UE
travel time. Even if there are a greater number of paths with
p
PDTij > 25%, this model shows improvement in network-wide
8. Conclusions congestion, evident from the least NCI value of 1.04 among the
OFD models. However, in terms of sustainability, the SO-MAR
This study develops novel bi-objective mathematical pro- model performs better than the SO-SAR model. It results in
graming models that ensure system optimality along with total system travel time of 73,91,820 veh-min which is less
congestion minimization on overutilized links with user than SO-SAR (77,50,125 veh-min) and UE model (74,80,225)
needs as constraints. In this regard, four congestion mitiga- veh-min. Furthermore this model yields lower values in terms
tion strategies are identified, and consequently, four bi- of total distance travelled. Additionally, the fuel consumption,
objective optimization models are formulated. The proposed which is given by the weighted sum of distance and travel
models (SO-SA, SO-MA, SO-SAR, and SO-MAR) overcome the time is expected to be less for the SO-MAR model. Hence, this
limitation of the existing models identified in the literature by model leads to economic and environmental benefits.
adopting multiple criteria in modeling network flows and Thus, both SO-SAR and SO-MAR are competing models.
exploiting underutilized link capacities to divert flow from One may choose the SO-SAR model for operational metrics
overutilized links. An additional constraint on the number of involving congestion mitigation, and the SO-MAR model for
eligible paths between O-D pairs (l) is imposed to eliminate sustainability benefits. It can be noted that the preferred l
long-duration paths for the users. The models are tested and values and the superior OFD model need to be calibrated for
analyzed on the Sioux Falls network (SF). other networks as well as for demand conditions. Overall, the
This paper explores two methods (weighted sum and formulations developed in this paper will be of value in better
e-constraint approaches) to obtain a continuous point on the distribution of flows in urban networks and could also lead to
Pareto front for this bi-objective problem. Every point on the reduced emissions and enhanced safety. The formulation can
Pareto front represents different trade-offs between the be extended by considering variable K-shortest paths for each
objective functions and can be adopted according to the O-D pair, which can be obtained by using a heuristic approach.
users' choice. We systematically generate these non-domi- Also, the formulation can be used to identify the quantum of
nated solutions by using the proposed mathematical models. capacity addition on highly congested links in the network.
From such Pareto optimal or non-dominated solutions, users Stochasticity in OD travel demands and trip travel times can
can choose the solution based on the requirements (for also be incorporated in to the models to enhance realism.
example, the least distance (LD) point). We consider the LD Such models can be of practical relevance to traffic managers
point from the origin as it is regarded as the best compromise in alleviating traffic congestion in urban areas during peak
solution for bi-objective problems. The four bi-objective periods.
models are solved by varying l. As the l value increases, the
travel time of the considered path increases from the
shortest UE travel time path, which is not preferred the users Authors contribution
do not prefer (lower values of l are more acceptable for the
users). However, a low value of l leads to fewer alternatives The author(s) contribution to the paper is as follows: H.
for flow distribution and may lead to higher congestion on Chellapilla: study, design, methodology, analysis, interpreta-
some links. Thus, an appropriate value of l for each OFD tion of results, and draft manuscript presentation. R.
102 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103

Sivanandan: conception, design, methodology, and interpre- Chira, C., Bazzan, A.L.C., Rossetti, R.J.F., 2015. Multi-objective
tation of results. B. Chilukuri: design, methodology, and evolutionary traffic assignment. In: IEEE 18th International
interpretation of results, C. Rajendran: conception, design, Conference on Intelligent Transportation Systems, Canary, 2015.
Cohen, S., 1991. Flow Variables. Concise Encyclopedia of Traffic &
methodology, and interpretation of results. All authors
Transportation System 1991, 139e143.
reviewed the results and approved the final version of the Correa, J.R., Schulz, A.S., Stier-Moses, N.E., 2007. Fast, fair, and
manuscript. efficient flows in networks. Operations Research 55 (2), 215e225.
Correa, J.R., Schulz, A.S., Stier-Moses, N.E., 2004. Selfish routing in
capacitated networks. Mathematics of Operations Research 29
(4), 961e976.
Conflict of interest Daganzo, C., Newell, G., 1995. Methods of Analysis for
Transportation Operations. Institute of Transportation
The authors do not have any conflict of interest with other Studies, Berkeley.
entities or researchers. Evans, L., Herman, R., 1978. Automobile fuel economy on fixed
urban driving schedules. Transportation Science 12 (2), 137e152.
Gkiotsalitis, K., 2021. Improving service regularity for high-
frequency bus services with rescheduling and bus holding.
Journal of Traffic and Transportation Engineering (English
Acknowledgments
Edition) 8 (5), 778e794.
Gunantara, N., 2018. A review of multi-objective optimization:
The authors appreciate the thorough and meticulous review methods and its applications. Cogent Engineering 5, 1502242.
of the paper by the reviewers. Their thoughtful and insightful Han, L., Zhu, C., Wang, D., et al., 2019. Discrete-time dynamic road
comments have been very valuable in refining the paper. The congestion pricing under stochastic user optimal principle.
authors are thankful to the reviewers. Transportation Research Part E: Logistics and Transportation
Review 131, 24e36.
Henry, J., Charbonnier, C., Farges, J., 1991. Route Guidance,
Individual. Concise Encyclopedia of Traffic & Transportation
references Systems 1991, 417e422.
Isa, N., Yusoff, M., Mohamed, A., 2015. A review on recent traffic
congestion relief approaches. In: The 4th International
Angelelli, E., Arsik, I., Morandi, V., et al., 2016. Proactive route Conference on Artificial Intelligence with Applications in
guidance to avoid congestion. Transportation Research Part Engineering and Technology, San Francisco, 2015.
B: Methodological 94, 1e21. Islam, R., Biswal, M.P., Alam, S.S., 1997. Preference programming
Angelelli, E., Morandi, V., Savelsbergh, M., et al., 2021. System and inconsistent interval judgments. European Journal of
optimal routing of traffic flows with user constraints using Operational Research 97 (1), 53e62.
linear programming. European Journal of Operational Jahn, O., Mohring, R.H., Schulz, A.S., 2000. Optimal routing of
Research 293 (3), 863e879. traffic flows with length restrictions in networks with
Angelelli, E., Morandi, V., Speranza, M.G., 2018. Congestion congestion. In: Operations Research Proceedings, Berlin, 2000.
avoiding heuristic path generation for the proactive route Jahn, O., Mohring, R.H., Schulz, A.S., et al., 2005. System-optimal
guidance. Computers & Operations Research 99, 234e248. routing of traffic flows with user constraints in networks
Bureau of Public Roads (BPR), 1964. Traffic Assignment Manual: with congestion. Operations Research 53 (4), 600e616.
Technical Report. BPR, Washington DC. Kaysi, I., Ben-Akiva, M., de Palma, A., 1995. Design aspects of advanced
Beccaria, G., Bolelli, A., 1992. Modelling and assessment of traveler information systems. In: Gartner, N.H., Improta, G., Urban
dynamic route guidance: the MARGOT project. In: The 3rd Traffic Networks, Springer, Berlin, pp. 59e81.
Vehicle Navigation and Information Systems Conference, Koutsoupias, E., Papadimitriou, C., 1999. Worst-case equilibria. In:
Vnis, 1992. Annual Symposium on Theoretical Aspects of Computer
Beckmann, M., McGuire, C.B., Winsten, C.B., 1956. Studies in the Science, Berlin, 1999.
Economics of Transportation. Yale University, New Haven. LeBlanc, L.J., Morlok, E.K., Pierskalla, W.P., 1975. An efficient
Ben-Akiva, M.E., Lerman, S.R., 1985. Discrete Choice Analysis: approach to solving the road network equilibrium traffic
Theory and Application to Predict Travel Demand. MIT assignment problem. Transportation Research 9 (5), 309e318.
Press, Cambridge. Lujak, M., Giordani, S., Ossowski, S., 2014. Fair route guidance:
Braess, D., 1968. ÜbereinParadoxonaus der verkehrsplanung. bridging system and user optimization. In: 17th IEEE
Unternehmensforschung 12, 258e268. International Conference on Intelligent Transportation
Braess, D., Nagurney, A., Wakolbinger, T., 2005. On a paradox of Systems, Qingdao, 2014.
traffic planning. Transportation Science 39 (4), 446e450. Marglin, S.A., 1967. Public Investment Criteria. MIT Press,
Branston, D., 1976. Link capacity functions: a review. Cambridge.
Transportation Research 10 (4), 223e236. Mousavi, S.S., Schukat, M., Howley, E., 2017. Traffic light control
Charnes, A., Cooper, W.W., Ferguson, R.O., 1955. Optimal using deep policy-gradient and value-function-based
estimation of executive compensation by linear reinforcement learning. In: IET Intelligent Transport Systems,
programming. Management Science 1 (2), 138e151. London, 2017.
Chiandussi, G., Codegone, M., Ferrero, S., et al., 2012. Comparison Ozcelebi, T., Sunay, M.O., Tekalp, A.M., et al., 2007. Cross-layer
of multi-objective optimization methodologies for engineering optimized rate adaptation and scheduling for multiple-user
applications. Computers & Mathematics with Applications 63 wireless video streaming. IEEE Journal on Selected Areas in
(5), 912e942. Communications 25 (4), 760e769.
 J. Traffic Transp. Eng. (Engl. Ed.) 2023; 10 (1): 86e103 103

Papadimitriou, C., 2001. Algorithms, games, and the internet. Dr. R. Sivanandan currently serves as pro-
Proceedings of the Thirty-Third Annual ACM Symposium on fessor of civil engineering at IIT Madras in
Theory of Computing 2001, 749e753. Chennai, India. He obtained his PhD degree
Rassafi, A.A., Jamour, D., Mirzahossein, H., 2013. Different in civil engineering, specializing in trans-
network performance measures using a multi-objective portation engineering, from Virginia Tech in
traffic assignment problem. International Journal of USA. His research interests are in the areas of
Transportation Engineering 1, 61278999. optimization applications to transportation
Richardson, A.J., Taylor, M.A.P., 1978. Travel time variability on engineering, intelligent transportation sys-
commuter journeys. High Speed Ground Transportation tems (ITS), traffic engineering and opera-
Journal 12, 77e79. tions, and congestion mitigation. He has had
Schulz, A.S., Stier-Moses, N.E., 2006. Efficiency and fairness of leadership roles in several sponsored
system-optimal routing with user constraints. Networks 48 research projects. He has served as a member in many expert/
(4), 223e234. professional committees related to transportation at local and na-
Sheffi, Y., 1985. Urban Transportation Networks. Prentice-Hall, tional levels. He is also a member of several professional societies.
Engelwood.
Solutions, S., 2012. Sidra Intersection User Guide. Akcelik Assoc.
Pvt. Ltd, Greythorn Vectoria. Dr. Bhargava Rama Chilukuri is an assistant
Transportation Research Board (TRB), 1994. Highway Capacity professor at the Transportation Division of
Manual (HCM). TRB, Washington DC. the Department of Civil Engineering in In-
Transportation Networks for Research Core Team, Accessed May dian Institute of Technology Madras, Chen-
8, 2019. Transportation Networks for Research. nai, India. He holds a PhD from Georgia
Wang, W., Wang, D., Sun, H., et al., 2016. Braess paradox of traffic Institute of Technology, Atlanta, Georgia,
networks with mixed equilibrium behaviors. Transportation USA. His research interests include traffic
Research Part E: Logistics and Transportation Review 93, flow theory of homogenous and heteroge-
95e114. neous traffic, traffic operations, numerical
Wardrop, J.G., 1952. Some theoretical aspects of road traffic methods, and simulation.
research. Proceedings of the Institution of Civil Engineers 1
(3), 325e362.
Yen, J., 1971. Finding the K- shortest loopless paths in a network.
Management Science 17 (11), 712e716. Dr. Chandrasekharan Rajendran took his BE
Zadeh, L.A., 1963. Optimality and non-scalar-valued performance (Honours) degree in mechanical engineering
criteria. IEEE transactions on Automatic Control 8, 59e60. from College of Engineering Guindy, Uni-
Zhang, W., Xu, W., 2017. Simulation-based robust optimization versity of Madras, ME in industrial engi-
for the schedule of single-direction bus transit route: the neering from College of Engineering Guindy,
design of experiment. Transportation Research Part E: Anna University, and PhD in industrial en-
Logistics and Transportation Review 106, 203e230. gineering and management from IIT
Zhou, Z., Chen, A., Bekhor, S., 2012. C-logit stochastic user Madras. He has been on the faculty of IIT
equilibrium model: formulations and solution algorithm. Madras since 1987. His areas of interest
Transportmetrica 8, 17e41. include production and operations man-
agement, inventory and logistics manage-
ment in supply chains, quality management and applied
Mrs. Haritha Chellapilla is currently a PhD analytics. He is a recipient of the prestigious Alexander von
research scholar at Transportation Engi- Humboldt Research Fellowship of Germany, and he is an elected
neering Division, Department of Civil Engi- Fellow of Indian National Academy of Engineering (FNAE). He has
neering, IIT Madras, India. She received her been awarded Dr.rer.pol.h.c. (Honorary Doctorate) from Univer-
master's degree in transportation engineer- sity of Passau, Germany (ranked within the top 30 young univer-
ing from National Institute of Technology sities as per the Times Higher Education Ranking). He has been
(NIT) Rourkela, Odisha, India in 2013. Her highly ranked among researchers, globally, in operations man-
research activity focuses on traffic opera- agement, consistently by leading international journals since
tions, network flow modelling and conges- more than a decade. He occupies the RAGS Family Foundation
tion management. Institute Chair Professorship in IIT Madras since 2016.