Professional Documents
Culture Documents
The Robust Urban Transportation Network Design Problem
The Robust Urban Transportation Network Design Problem
by
Alireza Naimi
A Dissertation
Doctor of Philosophy
Major: Engineering
May 2016
Copyright © Alireza Naimi
ii
Acknowledgements
I would like to take this opportunity to express my gratitude to all those who
helped me during the various stages of my life to understand the potential of scientific
First and foremost, I would like to thank my advisors, Dr. Mihalis M. Golias (the
University of Memphis) and Dr. Sabya Mishra (the University of Memphis), for their
wonderful knowledge, guidance, patience and support throughout this research. I have
benefited greatly from their advice on many research and projects over the years.
M. Golias (committee chair), Dr. Sabya Mishra (committee member), Dr. Charles Camp
(committee member), Dr. Kyriacos Mouskos (committee member), Dr. Dincer Konur
(committee member), and Dr. Bryan Higgs. Their comments and advice were very
Last but not the least, I reserve my deepest gratitude for my family, who always
iii
Abstract
Naimi, Alireza. The University of Memphis. May 2016. The Robust Urban
Transportation Network Design Problem (UTNDP). Major Professor: Dr. Mihalis M.
Golias.
important in urban transportation networks, due to heavy use and road segments that are
close to each other. Small disturbances on an urban transportation network segment can
have a huge impact on its accessibility. Intelligent adversaries may take advantage of
these vulnerable parts of the network and disrupt transportation operations, increasing the
Often, the decision about improving the networks in transportation planning and
considering the factor of vulnerability in their decision, planners could prevent or limit
This dissertation proposes two models for designing robust networks against
intelligent attackers. In both models, three stakeholders are considered: i) the network
manager/designer, ii) the adversary (intelligent attacker), and iii) the network users. The
frameworks of both models and some other possible models are presented in this
dissertation.
The first framework is a bi-objective designer model. The designer in this model
has two objectives at the top level: to reduce the total system cost and to reduce the
iv
vulnerability of the network. The Sioux Falls network consists of 24 nodes and 76 links
was chosen for to evaluate this framework. The decision of the designer and attacker was
improving or destroying the links. Metaheuristic algorithm was used to solve the designer
and attacker problems. For the user equilibrium problem, the Frank-Wolfe algorithm was
implemented. The objective of the designer of the network in the first model, consist of
two goals. The two goals may conflict on the amount of amount of limited available
between these two objectives may forms. The results proved that the proposed multi-level
model is able to find the Pareto front solutions for the two objectives of the designer. The
second framework is a three-level zero-sum game model. In this framework, the payoffs
from the designer are assumed to have the same value to the adversary entity. Therefore,
the goal of this framework is to minimize the maximum gain that the adversary can
achieve. An example network with 6 nodes and 16 links was used to examine this
framework. The results showed that the model could be a valuable tool to reduce the
Both frameworks were tested using a medium size network with applications to larger
v
Table of Contents
Chapter Page
1 INTRODUCTION ........................................................................................................ 1
Contributions .................................................................................................... 2
Introduction ...................................................................................................... 5
Reliability ................................................................................................ 16
Resiliency ................................................................................................ 18
Vulnerability............................................................................................ 19
Robust network design ................................................................................... 30
3 METHODOLOGY ..................................................................................................... 37
Introduction .................................................................................................... 37
Players ............................................................................................................ 41
Users ........................................................................................................ 42
Adversary ................................................................................................ 44
Designer .................................................................................................. 45
Frameworks for vulnerability/robustness ....................................................... 50
vi
Hierarchy of Decision Flows................................................................... 52
Multi-period plan for NDP ...................................................................... 57
Model 1: Bi-objective Designer Model ................................................... 61
Model 2: A Zero-Sum Model .................................................................. 65
4 SOLUTION METHODOLOGIES ............................................................................. 68
Introduction .................................................................................................... 68
vii
List of Tables
Table Page
8. The trip rates for the Sioux Falls network (1000 veh/time unit)......................... 100
9. The local optimum solution for the first scenario on the 16 link network, Bz = 1
............................................................................................................................. 135
10. The local optimum solution for the second scenario on the 16 link network, Bz = 2
............................................................................................................................. 135
11. The local optimum solution for the first scenario on the Sioux Falls network, Bz=1
............................................................................................................................. 137
12. The local optimum solution for the first scenario on the Sioux Falls network, Bz=2
............................................................................................................................. 138
viii
List of Figures
Figure Page
3. Vulnerability versus reliability. The thick line is the “risk curve” of Kaplan et al.
(1981) .......................................................................................................................... 20
ix
16. Variation of reaching the solution by different size of bit-string in binary chromosome
representation. ............................................................................................................. 86
17. Elitist selection with different size, for population size 100 ....................................... 87
23. Total System Travel Time for Test Network 2 - (16-Link) ........................................ 96
25. Links Included in Expansion (links with orange color) for the Sioux Falls network 101
26. Improvement of the two objectives at the designer level by generations ................. 103
27. Individuals of the two objectives at the designer level by generations..................... 105
28. Decision of the designer (Number of lanes to be added to the network). ................. 106
30. Improvement of the capacity-expanded network compare to the initial conditions . 109
31. Flows in the Initial and Improved network, before and after the disruptions (veh/day)
................................................................................................................................... 110
32. Travel times in the Initial and Improved network, before and after the disruptions
(min).......................................................................................................................... 112
x
33. Individuals of the two objectives at the designer level by generations..................... 114
34. Decision of the designer (Number of lanes to be added to the network). ................. 115
35. The optimal decisions of the attacker for the initial and improved networks ........... 116
36. Flows in the Initial and Improved network, before and after the disruptions (veh/day)
................................................................................................................................... 117
37. Travel times in the Initial and Improved network, before and after the disruptions
(min).......................................................................................................................... 119
38. The optimal decisions of the designer and the attacker ............................................ 123
39. Flow on the links after the disruptions (veh/day) ..................................................... 127
41. Travel system travel times by 𝐵𝑧 in the Initial and Improved network after the
disruptions ................................................................................................................. 132
xi
Abbreviations
Table 1
Notations - Sets and Indices
𝒜 Set of links
𝒩 Set of Nodes
ℛ Set of origin nodes; ∀ℛ ∈ 𝒩
𝒮 Set of destination nodes; ∀𝒮 ∈ 𝒩
The complete set of available paths connecting (O/D) pairs 𝑟 − 𝑠, ∀𝑟 ∈ ℛ, ∀𝑠 ∈ 𝒮 in
𝒦𝑟𝑠
the network
𝑞𝑟𝑠 Demand between each Origin-Destination (O/D) pair 𝑟 − 𝑠, ∀𝑟 ∈ ℛ, ∀𝑠 ∈ 𝒮
𝜆𝑟𝑠 Shortest path for O/D pair 𝑟𝑠
𝑛𝑟𝑠 Number of O/D pairs in the network
Table 2
Notations – Parameters
xii
Table 3
Notations – Variables
xiii
1 INTRODUCTION
transportation system for all users (both passenger and freight). However, a transportation
network is usually not designed from scratch. The network design problem (NDP) aims
(based on various and often conflicting objectives). It proves to be one of the most
challenging problems for researchers in the field of transportation. There are various
uncertain input parameters (e.g., demand and supply) and disruptions (natural or man-
made). The latter (i.e., disruptions) may reduce the supply of the network, change
demand patterns, and may even completely interrupt the operations of a set of network
elements. The research presented herein aims to develop mathematical models and
considering an intelligent adversary/enemy entity. This research aims to fill the latter gap
in the literature and propose game theory-based frameworks to study the strategic robust
network design against intelligent attackers. Two models are proposed in this dissertation
for designing robust network. The proposed models consider the following three
stakeholders:
1
2) Users of transportation network; and
based on the order of play, the information available to each player and structure of
formulation can be defined. In the frameworks provided in this dissertation, the traffic
entity moves. On the other hand, the adversary (or evil entity) is assumed to maximize
the disruption to the network. The users respond to the adjusted network by the
transportation agency and the evil entity. The proposed model can be customized and
applied to other similar network designs, like telecommunications and biology networks.
discrete optimization problems for the designer and attacker are combinatorial and NP-
hard to solve optimally (Feremans & Laporte, 2003). Hence, no efficient exact or
heuristic methods are available to solve these problems in reasonable computational time.
Therefore, metaheuristic approaches were used to solve these problems. On the other
hand, traffic flow at the user level was modeled using Nash equilibrium concepts. The
user equilibrium problem is convex and can be efficiently solved using methods like
Contributions
who attempts to exploit the vulnerabilities of the network to the maximum of his or her
2
(to the author’s knowledge). Therefore, the goal and achievements of the intelligent entity
also need to be defined. One of the appropriate approaches to analyze and model the
(due to the stochastic events, for instance, natural disasters) have been studied extensively
in the literature, this research aims to provide additional insights by considering the
network elements that are vulnerable to the intelligent adversary entity. Two frameworks
intelligent attacks;
discrete problems, and a convex combination method to solve the user equilibrium
problem; and
networks.
literature review of the related studies. The general design of transportation networks and
Chapter 3 briefly explains the game theory and its application in the proposed
frameworks. The mathematical formulation for each model and the role of each player in
3
modeling the robust network design problems is presented. Chapter 4 discusses the
algorithms to solve each player optimization model. The Genetic Algorithms parameter
settings are also discussed. Numerical experiments were conducted in chapter 5 to test the
performance of the proposed models. Lastly, chapter 6 concludes the dissertation and
4
2 LITERATURE REVIEW
Introduction
foundation to understand the models and algorithms proposed in this dissertation. Since
three decision makers are considered in this research, the possible models to account for
network design problem (NDP). The focus of this research is on NDP considering
design of robust networks has captured the attention of many researchers. As it is often
the case with popular terms, there is not a generally adopted notion for the vulnerability
network that is vulnerable is not robust and vice versa (Snelder, 2010).
objective(s). Therefore, the following objectives for each player can be defined: From the
designer side, it might have various objectives to improve the performance of a network.
The performance of a network can be represented as the total system cost, robustness
investments. On the other hand, from a user’s perspective, they look for their optimal
route choice, mode, and destination. From an adversary viewpoint, the objective is to
degrade the performance of the network to the maximum of his capabilities. Hence, a
5
robust network design model must consider the alleviation of the potential disruptions. In
the next section, first a general summary of three common methodologies to model traffic
flow is described, followed by a brief introduction to network design problem and its
methodology.
modeling/simulation of traffic flow and the considered level of details in the models. In
model complex transportation systems. Various types of models have been published
during the last decades. They can be categorized based on the level of detail and their
6
Feet Miles Scale
Complexity
Microscopic
Mesoscopic
Macroscopic
Time
ms seconds minutes hours days months years
Unit
Level of details
Figure 1. Complexity versus level of details in traffic flow modeling. Adopted from
Washington (2008).
unit (or other types of flowing items, for example, vessels, airplanes, packages).
operational (e.g., operational lane changing models) and tactical (e.g., tactical lane
changing models and tactical overtaking models) systems (Michon, 1985; Moridpour,
Sarvi, & Rose, 2006). In the case of modeling driver-vehicle, some behaviors (like lane
changing and overtaking) requires large amounts of information, and modeling the
decisions drivers are making based on these data, is difficult (Chamieh & El-kouatly,
n.d.; Kano, Shiraishi, & Kuwahara, 2007; Suzuki & Mori, n.d.; Wheeler & Llc, n.d.). In
7
addition, defining, simulating and validating rich cognitive driver behavior models,
combining the driver-vehicle units into groups of driver-vehicle (or other transport flow
units). The Cellular Automata (CA) models usually model the transport units in groups
that are moving from one cell to the others by advancing in simulation steps. Thus, these
Macroscopic traffic flow models try to formulate the relationships between traffic
(Washington, 2008). These types of models originated under a theory that traffic flows,
as a whole, are similar to fluid streams systems. The characteristics of traffic flow in the
network are typically considered homogeneous in a specific time unit (which usually
The natural population growth, and other factors such as the increase in income
and employment, will result in the increase in travel demand on transportation networks.
This may lead to problems such as congestion and safety in the system. The
problems. This will require new infrastructures for serving the new transportation
networks or improve the existing system. The planning, design, and managing these
issues are traditionally addressed in network design problem (NDP). NDP is usually used
for determining the optimal sub-network, which will result in improvement of the whole
network. Various definitions of NDP are provided in the literature. For example, Friesz
8
(1985) defined it as: “network design problem is to determine the optimal locations of
1995). Hence, in practice, these problems are typically decomposed into a sequence of
independent sub-problems are the classical four-step planning process, network design
The network design problem can be described using graph theory. Likewise, a
a set of links 𝑳 and a set of nodes 𝑵. Each link connecting two nodes, and can be directed
or undirected. Attributes like weight/cost (𝑪) can be assigned to each nodes and link
(Figure 2). NDP transforms an existing network (graph 𝐺 = {𝑁, 𝐿, 𝐶}) into a new
improved network (graph 𝐹 = {𝑁’, 𝐿’, 𝐶’}). In road transportation network, distances
between end points of links or travel time are well known attributes of links.
9
Finding the optimal road design has been the subject of transportation studies for
a long while, and is known to be one of the most complicated problems in transportation.
A large number of methodologies and solution algorithms have been presented over the
last 50 years to provide solutions to these complex mathematical problems (S.-W. Chiou,
2005a; Leblanc, 1973; Murray, Davis, Stimson, & Ferreira, 1998; Suwansirikul, Friesz,
the networks in new aspects, and in more analytical details. Among the possible
level programming problem, with two level. In problems with conflicting objectives
within a hierarchical structure based on the sequential order of two decision makers, bi-
level optimization is an effective solution approach. It originated from the fields of game
theory and it can describe a number of problems in transportation planning and modeling.
Its hierarchal framework involves two separate optimization problems at different levels.
In case of Stackelberg competition, the first problem - called the upper-level or leader
problem - has a feasible solution set. The solution set is determined by the optimization
problem at the second level. The second problem is the lower-level problem or the
follower problem. This concept can be expanded to define multi-level programs with any
number of levels (Vicente & Calamai, 1994a). The bi-level program is an NP-hard
problem; hence, it is difficult to solve using exact algorithms. Ben-Ayed (1993) and
10
Ayala (2013) investigated on bi-level problems and concluded that even a simple bi-level
problem with both linear upper-level and lower-level problems is also NP-hard. One
reason is that bi-level model for NDP is non-convex (Gangi, Pianificazione, & Luongo,
2005). Luo, Pang, and Ralph (1996) also mentioned that even if both problems at upper-
level and lower-level is convex, the convexity of the bi-level problem is not guaranteed.
Multi-level programming, which has received significant attention during the last
Stackelberg games. The network design problem can be cast into such a framework.
Marcotte (1986) presented a formal description of the problem and developed various
(usually) non-convex and are thus difficult to solve using exact optimization algorithms
The very first studies in bi-level NDP were investigated by Leblanc (1973),
Bruynooghe (1972), and Ochoa-Rosso (1969). They used the branch and bound
techniques for solving the NDP. Moreover, Poorzahedy and Turnquist (1982) studied a
typical heuristic algorithm to find the solution using integer programming model.
Further research has been done to find more efficient heuristic algorithms, which
may give near optimal solutions or local optimum solutions (Allsop, 1974; Steenbrink,
which are computationally efficient but result in suboptimal solutions and not suitable for
large real networks problems. Gershwin and Tan (1979) formulated the continuous
11
network design problem (CNDP) as a constrained optimization problem in which the
constrained set was expressed in terms of the path flows. Patrice Marcotte & Marquis
(1992) presented heuristics for CNDP on the basis of system optimal approach and
obtained good numerical results. However, these heuristics have not been extensively
annealing) drew the attention of researchers in mid-90s and 2000s. The benefit of using
(Mathew & Sharma, 2006). For an example, Friesz (1985) and Meng (2009) utilized
simulated annealing (SA) method to solve the upper-level problem. Despite the faster
runs of SA, especially for the larger problems, the solution quality of Genetic Algorithm
(GA) was found to be better than SA and other metaheuristic algorithms (Adewole,
2012). Mathew and Sharma (2006) performed a study on using GA in CNDP. They
applied their model to the small to large size problems. Mouskos (1991) utilized the Tabu
search to solve the single class bi-level UTNDP with a Budget constraint where the
decision variable was to improve (or not) each roadway link by one lane using the static
traffic assignment as the lower level. Furthermore, Zeng (1998) utilized a hybrid SA-
Tabu search method to solve the two-class (automobiles plus trucks) to solve the bi-level
expected costs and investments. From this point of view, Meng and Yang (2002) solved
the bi-level benefit distribution for network design problem using the ratio of the benefits
12
gained from the capacity expansion for each link. Their model was non-convex, non-
differentiable, and continuous, so they chose simulated annealing method to solve their
optimization problem. Their multi-objectives trying to maximize the total benefit among
the links, while also trying to minimize the differential between beneficial gained by each
link. In the next section, a brief introduction to resiliency, reliability, and vulnerability is
presented.
The majority of planner’s decisions dealing with project selection involve single
initial costs while benefits could spread over many years in the future. Brown (1980)
taking into consideration not only present costs but also the benefits that accrue over
several years into the future. Moreover, Baskan performed a study utilizing bi-level
optimization that took into consideration the increasing future congestion and limited
budget constraints (Baskan, 2013). Optimal link capacity expansion values were found by
minimizing the total system travel time as well as the associated link investment costs
problem, Falk and Liu (1995) investigated theoretic analysis for general non-linear bi-
level optimization problem and proposed a descent approach in terms of the bundle
method to solve the non-linear bi-level problem where the gradient of the objective
function can be obtained when the subgradient information of the lower level is available.
Chiou (2005) explored a mixed search procedure to solve an area traffic control
optimization problem confined to equilibrium network flows, where good local optima
13
Several attempts were made in the last few years to find the global optimum
solution for network design problems. Wang, Meng, and Yang (2010) partitioned the
feasible space of nonlinear travel time function into several regions and provided a path
based MILP. Each region represents a piecewise linear function that can approximate the
original nonlinear travel time function. The model required a heavy computational time
and required using a large memory storage. Paramet Lauthep (2011) mentioned that
Wang’s approach is inapplicable to the case of DNDP and MNDP, because the paths in
their network structure are generated in advance, where it should change during the
design process. He modified the model to a linked-based and provided an efficient mixed
integer linear program. Li, Yang, Zhu, and Meng (2012) presented a model to convert bi-
level CNDP into a sequence of single level concave programs, based on the concept of
gap and penalty function. Furthermore, Wang, Meng, and Yang (2013) presented a global
optimization method for DNDP. The presented model was not computationally efficient
Szeto, and Rashidi (2013). They classified the available models from different aspects: by
algorithms. A list of the possible future roadmaps was also provided. The decision
variables used in the previous studies were categorized into (1) strategic, (2) tactical, and
(3) operational. Strategic decisions are about adding new links and expanding capacities.
The two later are about to maintain the current network. In the next section, a review of
14
Performance measures
set of optimal solutions, for certain decision variables by optimizing different system
(Pei, Fischer, & Amekudzi, 2010). They are the main factor in determining whether a
roadway network is viable for the future. Some of the important performance measures in
the Total System Travel Time (TSTT). The level of congestion directly affects travel
times. When a part of a network becomes overly congested, travel times will increase and
level of services (LOS) will decrease. The effects of congestion can then spill into other
portions of the network and increase the system-wide travel times. Another important
transport network for letting travelers go from the point that they enter the network to the
point that they exit the network in a reasonable amount of time (Murray, 1998).
Similarly, mobility has attributes like having access to the point of interest, maintaining
networks, benefiting from travel to social contacts and potential travel (Alsnih &
the safety of a network is to reduce the annual number of crashes to a fraction of the
current levels (Dijkstra, 2013). Resiliency, reliability, and vulnerability are three other
15
Reliability
The reliability of the transportation network refers to the probability that a system
can perform its expected function to an acceptable level of performance for a given
period of time (Bell, 2000). Berdica (2002) defined the reliability of a network as the
possibility of moving freight or passengers from one place to another successfully. Yim,
Wong, Chen, Wong, and Lam (2011) further defined reliability as the ability of the
network and its elements to operate under capacity. Reliability gained more attention
during the 90’s when the natural disasters like earthquakes damaged or completely lost
the connectivity of some of the major roadways around the world (Yim, Wong, Chen,
Wong, & Lam, 2011). Following the development of transportation networks, the
reliability studies focus on alleviating the damage effects on the network and investigate
the unpredictable variations caused by the uncertainties (Nicholson, Schmöcker, Bell, &
Iida, 2003). Some of the main measure of the reliability of transportation networks are
connectivity reliability, travel time reliability, and capacity reliability (Chen, Yang, Lo, &
following types:
2. Buffer time methods: The extra time a user has to add to the average travel time
16
The importance of the measuring reliability can lie on the fact that most of the
users get a resiliency of the cost (Travel Time) over the time, and a sudden change on it
can have a big effect on the network. This can be seen usually during the need for fast
evacuation.
highly congested links that are considered as critical links. The problem with this
may not enable planners to identify the most critical highway segments or corridors in
standard deviation of travel time was examined by Fosgerau and Karlström (2010). They
found out that the maximum expected utility has linear mean and standard relationship
correspondingly to the travel time. Markov Chain has been used by several researchers to
replace the conventional transportation planning and predicting the future pattern of flow
(Antoniou, Koutsopoulos, Yannis, & Model-based, 2007). Indrei (2006) tried to model
the traffic flow system using Markov Chain theorem. However, his work only limits to a
unit car. Iyer, Nakayama, and Gerbessiotis (2009) also used a continuous-time Markov
chain (CTMC) model for predicting the reliability of a system by evaluating cascading
failure procedure. They distinguished all the possible cascading failures of different sets
17
of elements that lead to breakdown the whole system, and based on the Markov
Resiliency
different elements work together to recover after a disruption (such as flood, hurricane,
tornado, etc.) happens. Resilience can be viewed as the opposite of brittleness, which
describes a system that cannot tolerate disruptions, and loses the capacity (or
functionality or other words that describe the productivity of a system). In this context,
Resiliency gained more attention during the last decade, and various studies have
been performed on this topic. Some of the studies focused on the resiliency of the
maritime systems. In order to examine the resilience of ports, Kamal Achuthan (2012)
interactions between different elements in a port and saw how statically they can
incorporate in the resiliency of the port due to a disruption. He also considered the
includes resilience matrices for before, during, and after disruption, the number of ships
served by each resource and also queues and delays. Some of the important vulnerability
indicators in literature are described in the rest of this section. To manage the resilience
18
analysis methodology. They mainly focused on the costs (probable disruptions,
investments in resilience strategies, losses, and gaining from using resilience strategies).
Therefore, the study can be considered as a business work with a monetary focus.
Ouyang (2012). The first stage was defined as a consistent mode, which can be used as a
representation of disasters. The second stage defined as damage propagation, and the last
two stages defined as a situation which the authorities trying to stop the damage
propagation and recover it. He and his co-researchers chosen power grid model as a case
study, and then they consider several disasters (in their case, random hazards and
hurricane hazards) and different approaches to recovering the damages. They figured out
that the annual resilience mainly happens due to its higher frequency of occurrence
compare to hurricane hazards. In addition, they found out that the type of recovery
sequences is important.
Vulnerability
aspects. From a transport side, the vulnerability can be defined as how vulnerable the
systems (Erath, Birdsall, Axhausen, & Hajdin, 2009). In another word, it defined as
sensitivity to attack or injury. According to Jenelius (2010) the technological and social
aspect of transportation networks can be distinguished from their perspective: From the
19
combination of probability and the importance of failure. From the social side, exposure
is defined as the equivalent of importance which shows the failure impact to an individual
probability of failure.
differences between these two concepts is their focus on the magnitude and the
probability of the adverse consequences. Figure 3 shows the “risk curve” of (Kaplan &
Garrick, 1981) in probability format. The probability of occurrence of a scenario and its
level of damage can be found by looking at this curve. The frequencies of occurrence of
Level of Damage
Figure 3. Vulnerability versus reliability. The thick line is the “risk curve” of Kaplan et
al. (1981).
20
The interrelationships between infrastructures, impact of risks within the system,
and consequence of events has not been studied well in the literature. A failure of a
network component could also cause the breakdown of other critical infrastructures in a
disruption event. For example, a disruption in a fuel transport network for a period time
of several days to several weeks could have a sequence of further disruptions in other
network could affect the transportation system for foods (Murray & Grubesic, 2007). It
should be noted that different transport materials also do not have the same importance in
term of overcoming the critical situation. For instance, the transports of medicine and
foods usually have a more crucial impact than the farm products in severe events. In
terms of time, the interruptions in the service of an infrastructure may last for a short
period (e.g., few hours), or longer periods (e.g., several days or weeks), or in extreme
The concept of vulnerability can be classified in the following ways: static which
evaluates the vulnerability based on a physical property of a network, and does not
depend on traffic flow; and dynamic that directly refers to the robustness of a network.
Most of the works were focused on graph theory and their property correspond the
realistic model considers traffic flows, as they are the main concern of the designer of
network if the impact for a single user under a specific scenario is to be evaluated, this
may call for exposure of the user to that scenario Jenelius, Petersen, & Mattsson, 2006).
Kröger and Zio (2011) also categorized different approaches for assessing the
21
evaluation focuses on three main elements: degree of loss, degree of exposure, and
degree of resilience.
The availability and quality of alternative routes are a very important indicator of
vulnerability. The availability of spare capacity (capacity minus the flow) also could be
an important indicator of vulnerability. Other examples could be v/c ratio, the number of
OD-pairs that use a link, number of vehicles affected by spill back (the spare capacity
can be used to bypass an incident), extra vehicle kilometers traveled as a result of link
Crucitti, Latora, and Marchiori (2004) and Latora and Marchiori (2001) provided
a measure for the performance of a network, called ‘network efficiency’. The network
network. Their performance indicator is based on shortest path between nodes and
number of nodes. Efficiency of the network defined based on the number of possible
edges (higher number of edges increase the efficiency and reduce the disruptions in
network), and shortest path between all the nodes (smaller shortest paths means higher
efficiency).
of link, nodes, and whole network. In his model traffic flow is considered as the source of
vulnerability indicator, and was based on changes in the cost of travel and unsatisfied
on the cost of travel and unsatisfied demands of elements covered in grids (Jenelius &
Mattsson, 2012). In denser network areas, grids can be defined smaller to provide better
22
Equity of impacts if network degradation among all the users is considered as a
key in analyzing and design network. Jenelius (2010) presented a methodology for link
times. The degradation is measured using the total changes in travel time.
transportation problems in which demands act as upper-level decision variables (H. Yang
& H. Bell, 1998). In this algorithm, an influencing factor for each link is a ratio between
its usage and its capacity. In this case, the link that is used to its capacity or over is likely
to receive an improvement. This algorithm is applied to ramp (H. Yang & H. Bell, 1998),
zone reserve capacity (H. A. Yang, 1997) and O-D matrix estimation (Jin & Yang, 2014).
her model, the links that cross a line perpendicular to the target link 𝑎 are considered
alternatives for the link 𝑎, if they meet the following requirements: The absolute angle
between the link 𝑎 and the alternative link must be smaller than 60 degrees. The
vulnerability index in her model is based on: (1) the ratio of capacity of link 𝑎, over the
summation of capacities of alternative links for link 𝑎, (2) a function of shortest path
between link 𝑎 and its alternative links, (3) and a parameter for the importance of the
distance.
Reniers and Dullaert (2013) used a scoring system in GIS to evaluate the
each type of materials by the mode type and give each route segments a score that
23
represents the vulnerability on its transportation. They also used a score factor for the
distribution, mode, and assignment in the model. Taylor and D’Este (2007) defined a
simulator called VISTA and Bayesian Network to model the accident potential in links of
the network. They proposed an index called ARI (accident risk index) which was the
result of the Bayesian Networks (BN) output. The topography of BN comprised from
different variables such as the pavement quality and link attributes; and includes two new
parameters from VISTA: flow and speed. Having these two parameters, they claimed it
would improve the prediction power of BN. The validation process shows about 81
percent prediction validity, which the authors mentioned it can improve by improving
different input variable to BN. Berdica and Mattsson (2007) also developed a simulation-
based method to evaluated vulnerability at the link or network level. They predefined
twelve scenarios defined (i.e., Lane/link closure, change the BPR function element) to
models to study the possible disruptions in a graph network. Furthermore, the bi-
24
objective converted into a single objective using a weighted combination of the two
objectives. The objective is to min/max the bandwidth of the network and the impacted
O/D pairs. The model was based on selected number of nodes to be interdicted, the
optimization model seeks to find worst (best) nodes to be interdicted, such a way that the
total O/D disconnected would be maximum (or minimum) (Murray et al., 2007). Bell
(2000) and Bell, Kanturska, Schmocker, and Fonzone (2008) also developed a game
multiple regions could be defined by evaluating the total loss. The vulnerable
infrastructure then could be ranked to improve its robustness against disruptions. The
Network Robustness Index (NRI) was presented in Scott, Novak, and Guo (2005) and
Scott, Novak, Aultman-Hall, and Guo (2006). This index provides a performance
measure to assess the vulnerability of the link or the whole network. The NRI value is
obtained by comparing the total changes before a link removal, to the state before
disruptions. Therefore, the alternative route and the additional cost would be considered
in the model. The Scott et al. (2005) model was further developed by Sullivan, Aultman-
Hall, and Novak (2009). The new robustness index (NRI-m) is similar to the original NRI
index, and the only difference is in the partial capacity reduction of the elements.
involving the flow of traffic. They proposed a new unified model to present the network
25
performance measure. Their performance indicator provides importance identification
and the ranking of network components. The model is based on equilibrium demand and
presented in Table 4.
Table 4
Summary of Vulnerability Indicators
________________________________________________________________________
Static / Evaluate
Reference dynamic level Model formulation and Notations Notes
Disadvantages:
Does not consider flow
26
Static / Evaluate
Reference dynamic level Model formulation and Notations Notes
Advantages:
Combined travel-destination-mode-
route model
Considering modes of transfer
Disadvantages:
𝑻𝟎𝒋 : free flow path travel time for path j
Analysis is computationally expensive
𝝆: maximum service rate of path
𝒉𝒋: bottleneck link of path j
Advantages:
Fast
Advantages:
𝟏, 𝒊𝒇 𝒍𝒊𝒏𝒌 𝒂 𝒊𝒔 𝒏𝒐𝒕 𝒕𝒉𝒆 𝒍𝒊𝒏𝒌 𝒓𝒆𝒎𝒐𝒗𝒆𝒅
𝜹𝒂 = { Measures the effects to overall users
𝟎, 𝒐𝒕𝒉𝒆𝒓𝒘𝒊𝒔𝒆
Disadvantages:
𝒒𝒂 : the value of the nri for link a Computationally expensive
𝒙𝒂 : flow on each link a
𝒕𝒂 : travel time on each link a
𝒕𝒂 : travel time on each link a
𝒄𝒂 : the system-wide, travel-time cost of removing the
link a
𝒄𝒂 : total system, travel-time
Advantages:
Fast
Direct dynamic vulnerability measure
27
Static / Evaluate
Reference dynamic level Model formulation and Notations Notes
Disadvantages:
Quality of weighting approach
Erik Jenelius et al. Dynamic Node, link, (𝒆) (𝒆) (𝟎) Properties:
∆𝒄𝒊𝒋 = 𝒄𝒊𝒋 − 𝒄𝒊𝒋
(2006) network (𝒆) Flow
(𝒆)
𝒙𝒊𝒋 𝒊𝒇 𝒄𝒊𝒋 = ∞
𝒖𝒊𝒋 = { (𝒆)
𝟎 𝒊𝒇 𝒄𝒊𝒋 < ∞ Advantages:
(𝒌) (𝟎) High-quality analysis
∑𝒊 ∑𝒋≠𝒊(𝒄𝒊𝒋 − 𝒄𝒊𝒋 )
𝑰𝒎𝒑𝒐𝒓𝒕𝒂𝒏𝒄𝒆𝒏𝒆𝒕 (𝒌) =
∑𝒊 ∑𝒋≠𝒊(𝒘𝒊𝒋) Disadvantages:
(𝒌)
∑𝒊 ∑𝒋≠𝒊(𝒖𝒊𝒋 ) Computationally expensive
𝑰𝒎𝒑𝒐𝒓𝒕𝒂𝒏𝒄𝒆𝒖𝒏𝒔
𝒏𝒆𝒕 (𝒌) =
∑𝒊 ∑𝒋≠𝒊(𝒙𝒊𝒋 )
(𝒆)
𝒄𝒊𝒋 : the cost of travel from demand
Node i to demand node j when element 𝒆 has failed
(𝟎)
𝒄𝒊𝒋 : the cost of the initial, undam- aged network
(𝒆)
𝒖𝒊𝒋 : unsatisfied demand
𝒘𝒊𝒋: weight assigned to each od pair that reflects its
significance in relation to the other pairs
Advantages:
Quality
Adjustable grid size for faster/slower
and lower/higher details of analysis
Disadvantages:
Lower quality on denser grid, while
having redundant grids on the same
link on rural are
Advantages:
Ease and flexibility in defining
various scenarios
28
Static / Evaluate
Reference dynamic level Model formulation and Notations Notes
Disadvantages:
Needs to perform simulations
𝒒: flow
𝑪: capacity
𝑰𝟏 : shows he influence of the flow.
𝒒: flow
𝑻𝒃 : the time it take before the tail of a queue reaches
the upstream junction. The higher 𝑻𝒃 is the lower
will be the impact of an blockage
𝑰𝟐 : indicator showing the impact of an blockage
𝒒: flow
𝑰𝟑 : similar to 𝑰𝟏 however, limited to links with a
capacity of 2500 pcu/hour
𝒒: flow
𝑰𝟓 : similar to 𝑰𝟒 , capturing both effects and incident
probability. However, i5 also takes the possible
effect of blocking back into account
𝒒: flow
𝑰𝟔 : same as 𝑰𝟓 , however restricted to lower-capacity
links. (good for e.g. Risk-prone off ramps just
downstream) of a junction.
29
Static / Evaluate
Reference dynamic level Model formulation and Notations Notes
𝟕
𝑰 : sum of the effects (estimated by i1) on all
upstream links j of link i, which might be blocked
due to spillback of congestion of a blocking on links
i. This shows the links that cause large problems in
blocking back effects
𝒒: flow
𝑪: capacity
𝑽: captures the links that have a large volume
compared to their capacity. This usually is an
indication that the link is heavily used, and that if an
blockade happens, the queue will grow quickly.
In the previous sections, the concept of robust network design has been discussed.
The goal is to design reliable and robust networks that are less vulnerable to disruptions.
Robust network design focus is on the reduction of the impact of disruptions in terms of
reliability, vulnerability, and resiliency (Snelder, n.d., 2011). Disruptions could occur in
the travel times, trip rates, capacity, traffic signals, and even the change in direction of a
2. Strategy-specific (same as 1, but guided, which means some arc/nodes are more
likely to be disrupted.
30
3. Structured approaches (optimization-based)
models. Several studies in literature utilized dynamic traffic assignment (DTA) in robust
network design problem. A framework was presented by (Snelder, n.d., 2011) for robust
network design problem, considering combined route choice, mode choice and trip
distribution in the lower level DTA problem. In the model, the designer has the decision
of adding capacity to link, route or buffer lanes. The vulnerability is considering using a
term in the top-level objective function. This term, comprised from the multiplication of
expected number of incidents, by total system travel time loss, and multiplying by the
that the probability of an incident is a function that depends on the number of vehicle
kilometers driven. The short-term variation in supply caused by incidents is also included
in the model. They solved the model formulation using the genetic algorithm on several
test networks.
transformed the bi-level hierarchical problem into a single level. The vector of link
demand growth for its equilibrium flow. The lower level UE problem is solved by
parametric variational inequality, and a single level minimax model was provided.
Dziubiński and Goyal (2013) studied various games between a designer and an
adversary. The designer tries to form a network consisting of n links - which are costly to
construct - and protect a set of them against disruptions. On the other hand, the adversary
31
entity is interested in damaging the network to the maximum of its capabilities. Perfect
and imperfect information in different scenarios is assumed available to the designer. The
difference is considered as the knowledge of the designer of the possible moves of the
adversary, which depends on their payoffs. Their main finding was with limited available
resources, the best defense would be in sparse networks, rather than centralized.
damage the network, using bi-level formulation. In their method, the vulnerability index
value is based on the utility of alternative routes, considering the current flow, and ratio
of flow over demand. The utility is based on the ratio of free flow travel time over
marginal travel time and the relative capacity. The utility values range from 0 to 1, where
1 indicates that the link is extremely important to the connectivity of specific O/D route.
Martin (2007) studied various types of network design against attacks and
the defender in the inner model tries to minimize the users’ costs. The proposed
framework assumes that the defender at the outer level uses limited defensive resources
to protect a system from attacks. At the middle level, the attacker uses their limited
resources to attack the unprotected components while at the inner level the defender
operates the system to minimize operating costs from damage (resulting from the
attacker).
Zhang, Xu, Hong, Wang, & Fei, (2012) tried to utilize the unified performance
indicator defined by Qiang and Nagurney (2007) in order to provide a robust network
32
design. The model is formulated using bi-level optimization. In the upper level, the
constraint by his available budget/resources. The lower level problem is user equilibrium.
Chen, Zhou, Chootinan, Ryu, Yang, and Wong (2011) presented a bi-objective
model that optimized capacity reliability and travel time reliability. These performance
measures give the supply and demand of a roadway network’s reliability. The
minimization of total system travel time is a key objective when using bi-level
topic (Ben-Ayed et al., 1988; Gao, Wu, & Sun, 2005; Yang & Bell, 1998).
find the optimal set of projects which maximize the total reduction in the expected
networks. They considered two players in the game, designer, and adversary. The
designer forms link between a set of defined nodes. The adversary attacks on the nodes
based on his resources. They found out that the best shape of the network in terms of
affordability and reliability, is sparse and heterogeneous, and either fully or centrally
33
Table 5
summary of methods for robust transportation network design
Users Intelligent
Designer/Defender note
Adversary
Decisions Goal Decisions Goal Decisions Goal
Single/Multiple optimization
Solution method LL
Destination Choice
Max Robustness
Max Damage
Mode Choice
Variable(s)
Variable(s)
Author(s)
C/D/MI1
Route
link / route / Micro
Snelder (2010) S
buffer cap
D ● ● S GA ● ● ● ● DTA S - -
Simulation
Murray-Tuite and Mahmassani T/ Dis.
(2004) O
Rerouting C ● S ● SO I ● BF2
link
link / route / Converted to
S. Chiou (2015) S
buffer cap
C ● S GB ● SUE S - - single level
Dis.
Dziubiński and Goyal (2013) S adding links D ● ● S BF ● - I ● BF Topological
link
PS
Wu, Guo, Sun, and Wang (2014) S link-cap C ● S O ● DUE N - -
1
Continues, Discrete and Mixed Integer
2
Brute force - Enumeration
34
Summary and Future Research Needs
The works studied in this chapter provide a basis for the research in this
within a network, and design the robust networks against intelligent disruptions.
existing network under the budget constraint such a way that the social welfare and the
network robustness is maximized while accounting for the equilibrium of the route choice
of the network users. Improving a transportation network, with limited resources, could
approaches to model these types of problem and the solution methods have been
reviewed and studied. Furthermore, an introduction to game theory and network design
problem based on its concepts was presented. Likewise, the vulnerability indicators in
considered in many studies focusing on the initiate of the problem base on stochastic
In this study, three decision makers are considered to form the robust network
design models. This type of optimization problem is hard to solve since sets of decision
makers with different objectives are inherently involved. There is no clear urban
Therefore, attaining models that consider these players into the game forms the basis for
35
future areas of research. In the next chapter, the frameworks to model this problem is
provided.
36
3 METHODOLOGY
Introduction
This chapter first discusses the concept of game theory, followed by introducing
the players that are used in the proposed frameworks. Furthermore, the proposed
framework and the mathematical formulation of the models are presented. To formulate
the models, the sets, parameters, and variables are defined in Table 1 through Table 3.
The notations are similar to model and graph representations in (Sheffi, 1985), and are
adopted for the proposed models. To have a better understanding of the methodology,
some essential information about game theory concepts are presented in the next section.
Game theory
Game theory provides mathematical tools for analyzing situations in which parties
- called players - make independent decisions. A game is defined as a finite game when
each player has a finite number of options, the number of players is finite, and the game
game is the optimal decisions of the players, who may have similar, conflicting, or mixed
interest and the outcomes that may result from these decisions.
Hence, a game is a set of strategies for each player that does depend on other
players’ strategy. If the solution of any player does not depend on other players’ decision,
the problem is not a game. In a game of regular network design problem, the decision of
37
Perfect information available. A player that has perfect information knows
everything about the moves in the game at all the time. They player with
perfect information may not some information on other players payoff or the
chess. If one player is aware of another one, (i.e. human be aware of computer
moves), the human can reduce the final computer score (or improve her
of poker. Each player does not know all of their opponents’ cards. The payoffs
Based on the bind between decision of the players in variable-sum games, games
or if they do not move simultaneously, the later players are unaware of the
38
Sequential games (or dynamic games) are games where later players have
knowledge.
zero sum game and non-zero sum game. A zero-sum game is a situation in which a gain
or loss in utility of each player is exactly balanced by the losses or gains of other players’
utility. In other words, if the total gains of the players are added up and the total losses
are subtracted, the summation will be zero. On the other hand, in non-zero-sum games,
the summation of losses and gains is not equal to zero. Zero sum games are strictly
competitive, which means there exist some losses associated with each gain. Non-zero
sum games can be competitive or non-competitive. One of the common approaches for
solving zero-sum games is Minimax theorem. In game theory, Minimax is a decision rule
for minimizing the worst-case scenario loss (maximum loss). For the two players finite
zero-sum games, the solution from Minimax, Maximin, and Nash equilibrium are
equivalent. Therefore, in a zero-sum game, the participant’s loss of utility is exactly equal
to the gain of the utility of the other participant, while it is not the case for the non-zero
sum games. Many conventional games are considered in this category. A list of some of
39
Table 6
Examples of adversarial games
Deterministic Stochastic
A player, who is playing against a perfect opponent, knows that his opponent is
sequences of moves to find the best next move. However, due to the time limits of most
of running most of the known algorithms, it is usually impractical to consider all the
possibilities. Considering chess game, the initial state is the current board configuration.
Therefore, the operations to move in the search space are the moves of the players on the
board. The final state of search could be won or lost. Exact minimax analysis for the
chess game is infeasible with current solution algorithms and computational power in a
reasonable time. In order to reduce the complexity of minimax search, technics like α-β
pruning are being utilized, which works by pruning portions of the search space. To
evaluate the moves in the search space, a utility function (or payoff function) needs to be
defined. The aim of the utility function is to measure how badly the opponent is beaten.
The quantitative typical values are usually +1 (win) and -1 (lose), -infinity and +infinity
40
Dealing with conflicting objectives in optimization problems has always been
programs parameterized by their remaining variables. When the other programs are pure
mathematical programs, the problem is bi-level programming problem (BPP). When the
other programs are bi-level themselves, the problem becomes three level programming.
This notion can be extended to multilevel programs with any number of levels (Vicente
& Calamai, 1994b). The simplest form of BPP in the linear form shown to be NP-Hard
(Bard, 1998; Ben-Ayed et al., 1988; Vicente & Calamai, 1994b). Programs like linear
integer problem, bilinear, quadratic, and minimax programs can be stated as a special
instance of bi-level programs. There have been studies conducted to provide a link
between bi-objective problems and BPP (Bard, 1998; Ünlü, 1987). However, they were
not succeeded in a sense that the optimal solution of a given bi-level program is Pareto
optimal or efficient for both lower and upper-level problems. Game theory approaches
are widely accepted solutions to deal with the conflicting problems. The main problem
can be decomposed into sub-problems, which are the players in the game. In the next
Players
The problem and the solutions could be viewed from the aspect of three main
players involved: the designer of the network, the users of it, and the adversary entity.
Frameworks can be defined based on one or more of these players. The usage of any of
the frameworks differs by the involved players and the order of moves.
41
Other elements that can be considered in the frameworks are the non-intelligent
disruptors. Examples of these disruptions are natural disasters and unintended human
mistakes. Natural disaster refers to natural events like earthquakes, tornados, and
component, are not optimization models. The goal and decisions of each player are
Users
Users are the travelers in the transportation network. They can be defined as
persons, driver-vehicle units or any other modes of transportation. The goal of the user
level problem is to decide upon their route, in order to minimize the individuals’ travel
cost. The decision variables are flows on paths or bushes between the origin and the
destination(s). Users may have a choice of transport modes, and the departure time
choice.
Therefore, the problem at the user level is to assign the trip matrix into the
network using the route choice algorithm. A first mathematical investigation of the
problem has been done by Wardrop (as cited in Sheffi, 1985). He developed the so-called
Wardrop’s first and second principle of equilibrium model that are based on the concept
of Nash equilibrium in game theory. His model denotes that no user can experience a
lower travel time by unilaterally changing routes. In simple terms, the equilibrium is
achieved when the travel cost on all used paths is equal. This principle is behaviorally
42
robust, computationally efficient, and possesses the unique solution. A convenient way to
model the travel time of a roadway link is the Bureau of Public Roads (BPR) travel time
function that is has been used widely to model the static traffic assignment problem. It is
noted that this function is simplistic as it cannot capture the time dimension and the
traffic control characteristics of a roadway link as well as the impact of other adjoining
roadway links on the travel time of the subject roadway link and the interactions among
various classes of vehicles. The BPR travel time function 𝑡𝑎 specific to a given link 𝑎 is
given by
𝛽𝑎
𝑥𝑎
𝑡𝑎 (𝑥𝑎 ) = 𝑡𝑜 (1 + 𝛼𝑎 ( )) (1)
𝐶𝑎
Where 𝛼𝑎 and 𝛽𝑎 are link specific constants, and 𝑡𝑜 is the free flow time on
link 𝑎. Generally, the constants 𝛼𝑎 and 𝛽𝑎 are calibrated using the observed field data.
One important feature of BPR function is its monotonically increasing convex format.
The nonlinear programming model for the user level problem is provided in
𝑥𝑎
𝑀𝑖𝑛𝑖𝑚𝑖𝑧𝑒 𝑇𝑇 = ∑ ∫ 𝑡𝑎 (𝑤𝑎 , 𝑦𝑎 , 𝑧𝑎 )𝑑𝑤 (2)
𝑎∈𝐴 0
Subject to:
43
𝑟𝑠 𝑟𝑠
𝑥𝑎 = ∑ ∑ 𝛿𝑎,𝑘 𝑓𝑘 ∀𝑎 ∈ 𝐴 (4)
𝑟−𝑠 𝑘∈𝒦
Equation (2) denotes the objective function of the UE problem. Constraint (3)
describes the demand conservation condition. That is to say, all trips should be assigned
to the network. The flow on all routes between each OD pair has to be equal to the OD
trip rate. Constraint (4) outlines the relation between flows on links and route(s), for each
𝑟𝑠
OD. The binary value of 𝛿𝑎,𝑘 is 1, when link 𝑎 is on the path 𝑘, and it is zero otherwise.
Constraint (5) and (6) satisfies the non-negativity of path flow and travel demand
Adversary
Adversary or attacker is the entity who is trying to degrade the performance of the
network. Its objective can be maximizing total system travel cost, the number of travelers
who cannot reach their destinations, and other vulnerability related factors. Examples for
manipulate VMSs/DMSs, and signal timing. The adversary entity possibly can increase
his total gain by having some information about the vulnerability of the network.
The degradation of links due to the damages from adversary can be modeled in
various approaches: 1) removing the link(s) from the graph, 2) Adding a big number as a
constant term to the cost of using the disabled link(s), 3) decreasing the capacity of
link(s), 4) increasing the free flow travel time of the link(s). The first approach is not
44
suitable for this problem since it make the graph disconnected. One tactic to deal with
this issue is eliminating the OD pairs that cannot make their trips and associating a cost to
this removal. The second approach does not consider any property of links. The third
approach would conflict with the capacity expansion procedure of the model. The last
Designer
designer (or the defender) can be defined as a single level process to improve the
vulnerability, reliability, system cost or etc. The objective of the designer is related to
economic growth and is about to improve the desired performance measures. For
instance, he may try to lower travel times, provides more reliable travel times, improve
social equity, improve environmental conditions, and improve livability. Other objectives
45
The decision variables available to the designer to achieve the mentioned
• Capacity
• New links
• Transit schedules
• Number of lanes
• Toll pricing
• Signal timing
• Maintenance
In the games with asynchronous moves like Stackelberg games, each player has
different benefits from the hierarchy structure of the optimizations. For example, in the
Stackelberg game with two players, the follower player that moves after the leader, has
the advantage of having perfect information regards the leader’s moves. On the other
hand, the first player has the benefit of implicitly control the next player’s move, such a
way that optimize his/her own goal. In this case, the evil entity has the advantage of
moving after the designer move. The decisions by the designer already are in place, and
at the time of adversary’s move, the designer cannot take any more actions. Hence,
adversary entity has a perfect information about the designer’s move, while the designer
does not. The objective can be defined in different approaches. For instance, it can be a
46
maximization of the average travel time, minimizing the connectivity, maiming the total
system travel time of a specific region or etc. However, a more realistic objective is to
method. The designer of the network does not have perfect information about the possible
moves of the adversary. Therefore, an appropriate model should search over the possible
The first move is completed by the designer of the network who has the advantage
of putting his decision in place, and observing the reaction of the other players. The
designer decision is defined by vector y. The value of 𝑦𝑎 shows the amount of expanding
the capacity of link a. In this research, 𝑦𝑎 is the number of lanes to be added to link a. In
the proposed model, it is assumed that the adversary entity finds the maximum possible
damage to the network. His decision in the model is defined as vector z. The value of
𝑧𝑎 = 1 shows the state to which link 𝑎 is damged and not available to the users;
otherwise, the link is not affected. The damage is evaluated as the increase in the total
system travel time. After the decisions of the designer and the adversary were made, the
users of the network complete the next move. The reaction of the users is modeled using
user equilibrium principles. The bi-level formulation models the relationship between the
network manipulated by designer and adversary at the upper level, and the users at the
lower level problem. These problems are described in details in the rest of this chapter.
in the process of network design is crucial to alleviate the consequences of such events.
47
Furthermore, a selection of vulnerability indicators was presented. As it was studied in
the literature, the vulnerability can be evaluated by measuring the increase in the Total
System Travel Time (TSTT). This way, the damage due to the disturbance is evaluated
over the whole system. One approach to deal with the robust design of the network would
be modeling the problem as a bi-level formulation. Therefore, the solutions of NDP can
be compared to the increase in their TSTT, after the disruption occurs; a network that its
TSTT has less variation would be considered more robust. Hence, the goal of the NDP is
defined as to reduce the maximum possible damages that the adversary entity can imply.
From the designer side, the model formulated as a min-max optimization, where the
designer tries to minimize the maximum damages which enemy entity would put on the
network.
investing the available budget/resources in the expansion of the current capacity of links,
by adding new lanes. Therefore, the designer makes his decision by adding new lanes to
the network, considering his budget as a constraint. Then, the model examines the
maximum damage that an adversary can inflict on the network, by incapacitating the
resources/budget. Hence, the goal at upper level can be modeled as equation (7).
48
where 𝐷𝑦,𝑧 represents the total payoff to the adversary. The value of the payoff
can be considered as the increase in total system travel time. Therefore, the adversary can
look for the damage which results in the maximum possible travel time of users of the
system. In this case, the objective at the upper level may be written as:
The decision of the designer and adversary are constrained by the following
limits:
∑ 𝑔𝑎 (𝑦𝑎 ) ≤ 𝐵𝑑 (9)
𝑎∈𝐴
∑ 𝑧𝑎 ≤ 𝐵𝑧 (10)
𝑎∈𝐴
𝑔𝑎 (𝑦𝑎 ) = 𝑦𝑎 . 𝑑𝑎 , ∀𝑎 ∈ 𝐴 (11)
adversary, 𝑔𝑎 is the total cost of adding 𝑦𝑎 lanes to link 𝑎, and 𝑑𝑎 is the cost of
constructing one lane for link 𝑎 (eq. (11)). Finally, constraint (12) and (13) respectively
requires non-negativity of designer’s decision, and the binary decision of the adversary
entity:
49
𝑦𝑎 ≥ 0, ∀𝑎 ∈ 𝐴 (12)
𝑧𝑎 = {0,1}: ∀𝑎 ∈ 𝐴 (13)
where 𝑧𝑎 = 1 shows that link 𝑎 is disabled, and 𝑧𝑎 = 0 indicates that link 𝑎 is not
affected. The result of the upper-level of bi-level model is available to the users. The
users’ move, is done after the first two players find their decisions, and passed it to the
user level. In the next section, the behavior of users in reaction to the decisions made at
upper-level is discussed.
In the proposed model, the infrastructure that is selected for investment by the
designer might also be selected by the adversary. The model does not prevent the
designer from investing in these type of infrastructures that are attractive to the adversary
entity. One question may arise concerning why investing in a component, which the
network of two links, connecting the same origin to a destination. There is a total flow of
100 units, which at the equilibrium, 70 units take link 1 and 30 units take link 2.If the
decision of the designer is to increase the capacity of the links, he has two options:
scenario 1 to invest on link 1, and scenario 2 to invest on link 2. If the investment goes to
link 1, it will be more attractive for the current users, which will result in diverting more
traffic to link 1. If the adversary wants to damage this network by disabling one of the
links, the worst-case scenario would be attacking link 1, since it carries the mainstream
50
flow. Hence, the potential damage to the network will be even higher than the original
2
a) Base network
1 1
2 2
On the other hand, if the designer decided to invest on link 2, the flow will be
distributed more uniformly over the network, which decreases the total possible damage
(Figure 4c). In this case, if the adversary decides to disable link 2, the maximum flow that
he can affect is 60. On the contrary, if he decides to disable link 1, he will be able to gain
more than the payoff from disabling link 2. A similar pattern could exist in networks that
are more complex. By comparing the two scenarios, it can be concluded that even if the
51
investment goes to a link that will be more vulnerable and more attractive to an
adversary, it might be the case that this decision reduces the maximum potential damage.
different levels of the hierarchy, decisions are to be made. One of the common
approaches to deal with this kind of problems is to focus on one level and consider the
decisions of the other levels as constant. The constraint domain of a multilevel problem is
52
(P1) min
1
𝑓1 (𝑥1 , 𝑥 2 , … , 𝑥 𝑘 )
𝑥
s.t. 𝑔1 (𝑥1 , 𝑥 2 , … , 𝑥 𝑘 ) ≤ 0
where 𝑥 2 solves
(P2) min
2
𝑓2 (𝑥1 , 𝑥 2 , … , 𝑥 𝑘 )
𝑥
where 𝑥 𝑘 solves
(Pk) min
𝑘
𝑓𝑘 (𝑥1 , 𝑥 2 , … , 𝑥 𝑘 )
𝑥
s.t. 𝑔𝑘 (𝑥1 , 𝑥 2 , … , 𝑥 𝑘 ) ≤ 0
The first level problem P1, represent the problem of the first player in the
controlling the decision variable 𝑥1 . This problem (P) known to be np-hard. Even a bi-
The bi-level problem of leader and follower was first proposed by Stackelberg
(1952). The general bi-level programming problem (BLPP) can be written as follows:
53
In the above formulation, 𝑋 is the feasible set of 𝑥 variables, and 𝑌 is the feasible
where 𝑦(𝑥) is a solution to the lower level problem P2, for any fixed 𝑥 variables.
MIN
MIN
a) Conventional bi-level
54
MIN
MAX
MIN
MAX
MIN
MAX
MIN
55
MO Opt.
MIN
objective bi-level problems are presented in Figure 5. Each pattern inside the circles
represents different objectives. For instance, in the bi-level problem in Figure 5a, the
leader problem at the upper-level problem ULP implicitly controls the lower level
problem LLP, by feeding it with his/her own decision. Then, the LLP solved using the
fixed decisions from ULP. The decision maker at ULP observes the reaction from LLP
and further proceeds finding the optimal solution for himself/herself considering a stable
56
Designer min 𝐷 x*, y*, z*
𝑦
y1 y2 y3
Designer max 𝐴 z* z3
𝑧
z2
z1 z9
z2 z3 z8
z4 z5 z6 z7
Designer min 𝑈 x1 x9
𝑥 x1* x8
(Convex x3 x4 x6 x7
x5
optimization) x7
presented in Figure 6. The decision makers at each level try to implicitly control the
decision of his/her nested problems. In other words, the decision maker at each level is
provided with a set of feasible solutions governed by its leader(s). The flow of the
The goal of classical network design problem is to find solutions that improve the
network in a single period. Planners usually need to have the flexibility to change, delay,
or abandon the future investments. In real world problems, the decisions of investment
are usually spread over many years in the future. Therefore, models to find sequencing
network investments over time are developed. An example of a multi-period NDP for the
57
min ∑ ∑ 𝑥𝑎𝑝 𝑡𝑎 (𝑥𝑎𝑝 , 𝑦𝑎𝑝 )
𝑦
𝑝∈𝑃 𝑎 ∈𝐴
(17)
s.t.
∑ 𝑔𝑎 (𝑦𝑎𝑝 ) ≤ 𝐵 𝑝 , ∀𝑝 ∈ 𝒫 (18)
𝑎∈𝐴
s.t.
𝑝
𝑞𝑟𝑠 = ∑ 𝑓𝑘𝑟𝑠,𝑝 ∀𝑟, 𝑠, 𝑘, ∀𝑝 ∈ 𝒫 (22)
𝑘∈𝒦
𝑥𝑎𝑝 = ∑ ∑ 𝛿𝑎𝑘
𝑟𝑠,𝑝 𝑟𝑠,𝑝
𝑓𝑘 , ∀𝑎 ∈ 𝒜, ∀𝑝 ∈ 𝒫
(23)
𝑟,𝑠 𝑘∈𝒦
α
𝑥𝑎𝑝
𝑡𝑎 (𝑥𝑎𝑝 , 𝑦𝑎𝑝 ) = 𝑡𝑎𝑜 [1 + βa ( 𝑝 ) ] (24)
(𝐶𝑎 + 𝑦𝑎𝑝 )
𝑝−1
𝑟𝑠,𝑝
𝑓𝑘 ≥ 0, ∀𝑟, 𝑠, 𝑘, ∀𝑝 ∈ 𝒫
(26)
𝑝
𝑞𝑟𝑠 ≥ 0, ∀𝑟, 𝑠, 𝑘, ∀𝑝 ∈ 𝒫
where 𝐵 𝑝 is the available budget at time period 𝑝, and 𝐶𝑎𝑝 is the capacity of the
link 𝑎 at time period 𝑝. Variable 𝑦𝑎𝑝 is the capacity expansion vector on link 𝑎 at time 𝑝.
58
Figure 7 presents a numerical experiment on investing on the Braess network over a 3
year span. The investment budget is B1=50 for the first year, B2=20 for the second and
59
Bi LGoal
U
TSTT 3450
3400
3350
3300
0 5 10 15 20 25 30 35 40 45 50
Bi
x
100
Flow
50
0
0 5 10 15 20 25 30 35 40 45 50
Bi year 1
y
15
Investment
10
0
0 5 10 15 20 25 30 35 40 45 50
Bi year 2
y
8
Investment
0
0 5 10 15 20 25 30 35 40 45 50
Bi year 3
y
15
Investment
10
0
0 5 10 15 20 25 30 35 40 45 50
Iteration
60
Model 1: Bi-objective Designer Model
distract the focus of the designer from the important objectives in NDP under normal
infrastructures that are beneficial after the disruptions. However, it is possible that no
disruptions occur. Hence, there is a tradeoff between focusing on improving the system
under normal condition and after disruption events. An intellectual approach would be
simultaneously considering both aspects of reducing the system-wide cost, and the
that the social welfare and robustness of the network are maximized simultaneously. The
methodology presents a bi-objective formulation for the road network planner to model
The general formulation for this model is presented in equations (27) through
(31).
61
in which 𝑧(𝑦) is the solution to the problem 𝐴 in equations (29) and (31);
likewise, 𝑥(𝑦) and 𝑥′(𝑦, 𝑧) are the solutions to the problems 𝑈 and 𝑈′ in equations (30)
and (31) respectively. Statements (27) and (28) form a bi-objective problem, which tries
to reduce the total system cost of the network simultaneously at the normal condition
(before attack) and at the degraded condition (after attack). The model formulation is
𝑔𝑎 (𝑦𝑎 ) = 𝑦𝑎 . 𝑑𝑎 , ∀𝑎 ∈ 𝐴 (35)
𝑦𝑎 ≥ 0, ∀𝑎 ∈ 𝐴 (36)
𝑥
𝑥𝑎 = 𝑎𝑟𝑔𝑚𝑖𝑛𝑥𝑎 ∑𝑎∈𝐴 ∫0 𝑎 𝑡𝑎 (𝑤𝑎 , 𝑦𝑎 , 𝑧𝑎 )𝑑𝑤 (37)
xa α
s.t. t a (xa , ya ) = t oa [1 + βa ((y ) ] (38)
a +Ca ).ha
62
𝑥
s. t. 𝑥́ 𝑎 = 𝑎𝑟𝑔𝑚𝑖𝑛𝑥́ 𝑎 ∑𝑎∈𝐴 ∫0 𝑎 𝑡́𝑎 (𝑤𝑎 , 𝑦𝑎 , 𝑧𝑎 )𝑑𝑤 (44)
𝑥́ a α
𝑡́a (𝑥́ a , ya , 𝑧𝑎 ) = t oa [1 + βa ((y ) ] . (1 + 𝑀𝑎 . 𝑧𝑎 ) (45)
a +la ).ha
𝑞́ 𝑟𝑠 ≥ 0 ∀𝑟, 𝑠 (49)
∑𝑎∈𝐴 𝑧𝑎 ≤ 𝐵𝑧 (50)
𝑥́ a α
𝑡́a (𝑥́ a , ya , 𝑧𝑎 ) = t oa [1 + βa ((y ) ] . (1 + 𝑀𝑎 . 𝑧𝑎 ) (51)
a +la ).ha
The defender/designer of the network not only minimizes the overall cost to users,
similar to the conventional urban network design problems, but also they attempt to
reduce the potential effect of any intelligent adversary movements. The decision of the
designer implicitly affects the behavior of the users and the potential moves of the
adversary entity. Therefore, the hierarchy of decision flow starts from the designer at
level L1. From a high-level perspective, it has two objectives: minimizing a performance
measure (e.g. total system travel cost), and concurrently minimizing vulnerabilities.
There are two players in the next sequence of the hierarchy structure of decision makers
(L2): a user player for the normal condition of the network, and an adversary. The
network modified by the adversary provides input data for the last player in the hierarchy
63
sequence of decisions (L3). This user level is based on Wardrop’s first principle. His first
principle states that no user can experience a lower travel time by unilaterally changing
his/her routes. The road users individually select routes such a way that their travel costs
are minimized, while the planners look for the best network improvement but have no
control over the users’ route choices. Similar rules apply to adversary-user interactions, in
which adversary entity has implicit control over the users’ decisions.
In equations (32) through (50), two separate user equilibrium problems are
defined. The first problem is on the second level of optimization, which provides
information regarding the flow in an undamaged network. The second one is at the third
level, which feeds the adversary problem with information about the degraded network.
The variables for the second user equilibrium problem are differentiated using the
Constraint (34) represents the total available budget 𝐵𝑑 for network capacity
improvements. Similarly, Constraint (50) shows the limitation on the number of elements
that adversary entity can damage. The investment cost function at constraint (35) is
adopted in the same way as Suwansirikul et al., (1987) and Abdulaal & LeBlanc (1979).
The bi-objective problem in equations (32) and (33) can be combined and form a
64
where (𝛾) is the weight parameter. In this convex combination, the weighting
Looking at the system-wide view to the network, the decision of the intelligent
adversary could be reverse to the designer/defender, in the sense that both players try to
minimize/maximize the total cost to the users. If the objective of both of the mentioned
players is the same, the problem lies in the category of zero-sum games. In zero-sum
games, the payoffs from one player to his/her opponent have the same value. Minimax
problems are similar to the Stackelberg games, in the sense that two non-cooperative
the two players have the same objective function (to be maximized for one player, and
minimized for another one). This type of programming suitable for optimization under
uncertainty and each player decision is in reaction to the other one in most destructive
manner. Therefore, at any stage, each player is interested in a strategy which minimizes
the worst possible future contingency (Tsoukalas & Wiesemann, n.d.) (Figure 5b). An
example of minimax problem is the game of chess. In this game, each player tries to
optimize his/her objective, which is a utility function that works as a scoring operator.
The sequential movements of each player could end at the final move with the associated
score of win, lose, or draw. The score at each node of the hierarchy is calculated by
summing up the defined utility function’s scores, from leaves to that node. The optimal
65
solution for the player at the root is the worst score for the opponent at the nodes right
The general formulation of this model is presented in equations (55) through (57):
The proposed zero-sum based game for robust NDP against intelligent disruptions
𝑦𝑎 ≥ 0, ∀𝑎 ∈ 𝐴 (61)
𝑧𝑎 = {0,1}: ∀𝑎 ∈ 𝐴 (64)
𝑥
𝑥𝑎 = 𝑎𝑟𝑔𝑚𝑖𝑛𝑥𝑎 ∑𝑎∈𝐴 ∫0 𝑎 𝑡𝑎 (𝑤𝑎 , 𝑦𝑎 , 𝑧𝑎 )𝑑𝑤 (65)
xa α
s.t. t a (xa , ya , 𝑧𝑎 ) = t oa [1 + βa ((y ) ] . (1 + 𝑀𝑎 . 𝑧𝑎 ) (66)
a +Ca ).ha
66
∑𝑘∈𝒦 𝑓𝑘𝑟𝑠 = 𝑞𝑟𝑠 ∀𝑟, 𝑠 (67)
𝑟𝑠 𝑟𝑠
𝑥𝑎 = ∑𝑟−𝑠 ∑𝑘∈𝒦 𝛿𝑎,𝑘 𝑓𝑘 ∀𝑎 ∈ 𝐴 (68)
the hierarchical relationship between two independent, and possibly conflictual, decision
makers. Mathematical program in equation (58)-(64) and (65)-(70) are connected using
shared variables, specifically the decision of designer 𝑦𝑎 and adversary 𝑧𝑎 at upper level,
and flows 𝑥𝑎 at lower-level problem. Another important thing that should be noted is the
fact that the decision of the planner cannot be computed until flows are known. This issue
can be addressed by passing an initial vector of flows to the upper level, found by solving
the flow assignment for the basic network. The designer and adversary cannot control the
In the next chapter, solution algorithms to the problems defined in this chapter are
provided in details.
67
4 SOLUTION METHODOLOGIES
Introduction
In this chapter, solution methodologies to the problems of the three players in the
frameworks defined in chapter 2 are discussed in detail. First, the complexity and the
subset set of graph edges satisfying a set of constraints with minimum total weights (or
with three nodes. In the case of a network with three nodes, at least, two links are needed
to connect all nodes. The total number of different network configurations can be
68
1 𝑛(𝑛−1)
𝑛(𝑛−1)
2
Number of possible networks = ∑𝑖=𝑛−1 ( 𝑖
2 ) (71)
where 𝑛 is the number of nodes in the graph. Figure 9 shows the number of
possible undirected networks by the number of nodes. As it can be seen, the number of
10,000,000
1,000,000
Number of Possible Networks
100,000
10,000
1,000
100
10
1
0 1 2 3 4 5 6 7 8
Number of Nodes
other decisions like the number of lanes to be added, speed limit, signal timing and toll
pricing might be considered. Therefore, due to the combinatorial nature of this problem,
the domain size of the feasible solutions grows exponentially. For instance, if adding up
to 5 lanes is considered in a problem, the size of the feasible solution will be as Figure 10.
69
10,000,000-100,000,000 1,000,000-10,000,000 100,000-1,000,000
100,000,000
Feasible Region Size
10,000,000
1,000,000
100,000
10,000
1,000
100
10
1
1 5 Lanes
2 4 Lanes
3 de
d
4 3 Lanes ad
5 be
Numb 6 2 Lanes st
o
er of L 7 a ne
inks to 8 1 Lane fL
be Im 9 r o
prove 10 be
d in N
etwor um
k a xN
M
For the users of the network, the feasible region of is his/her route choice. The
assignment of flow to the links can be presented at node level (using nodes’ input/output,
specific section(s) of networks nodes, links, or a combination of nodes and links. The
damage to the network can be applied by destroying a bridge or tunnel, staging a car
accident, and stalling a vehicle. Models like (Qiang & Nagurney, 2007) and (Wu, Guo, &
algorithms (e.g., calculating a large amount of shortest paths and/or finding equilibrium
70
flows). Similar to the network design problem, modeling the problem of an adversary
entity as a modifier of network’s edges and vertexes is also a combinatorial problem. For
an instance, consider a network that consists of 200 nodes and 1000 links. Reducing the
ways. If two nodes can be selected, the total number of possible combinations is (200
2
)=
19,900. Similarly, expanding the limit to 10 nodes, make the size of solution feasible
space equal to 2.2e16. If the decision were to focus on disabling links instead of nodes,
the size of the problem would be even larger. In general, the larger discrete feasible space
will require more time and effort for search algorithms to solve the problem.
Several studies have tried to decrease the size of the problem by discretizing the
continuous feasible space into counterparts. For instance, in Jenelius’ approach, the
geographic map of the network is divided into grids (fixed or variable sizes) (Jenelius,
2011; Jenelius & Mattsson, 2012). The grid based analysis method is said to be a vital
complement to the existing studies of single link failures, since various types of events
such as storms, wildfires, floods, and snowfall can cause widespread degradations in
transportation network. In this way, the damages to the network affect the grids, which
their size of the feasible region is usually smaller than the complete network. In his study,
decreasing the size of the cell grids from 50 km to 12.5 km increased the calculation time
(the time required for calculating the importance of every cell and the exposure of every
have been used in literature to solve the NDP. For discrete NDP problems and non-
71
convex problems (e.g., combinatorial optimization and integer programming), exact
methods like enumeration and branch and bound and metaheuristics (e.g. genetic
algorithm and simulated annealing) have been used. For continuous NDP, technics like
sensitivity analysis based algorithms are common. Furthermore, other technics such as
researchers.
If the domain at design or adversary level is integer, a branch and bound method
for small/medium size problem is found to be a good solution approach. Studies like
Leblanc (1973) utilized Branch and Bound technique to solve the NDP problems. Further
improvements on this method have been done in later methods like Bard’s algorithm
(LeBlanc & Boyce 1986). However, Discrete variables make the problems NP-hard and
non-convex. Therefore, exact methods like branch and bound method cannot solve these
problems efficiently.
More efficient heuristics methods were discovered later. Heuristics are usually
developed from the insight of the problem, but there may not be convergence (Farahani et
al., 2013). Some of the heuristics approaches have been discussed in chapter 2. Many
engineering problems deal with functions and/or constraints that their derivative
Algorithms or Ant Colony approach showed a good performance to solve the NDP
72
related problems. Table 7 presents a review of the metaheuristics approaches for solving
network design problems. No studies found that utilized the Genetic Programming (GP)
approach. However, it might not be a suitable approach for these types of problems.
Table 7
Metaheuristics algorithms for NDP
Descent local search (LS) (Patriksson & Rockafellar, 2002)
Single solution based
73
Several studies tried to convert the bi-level problem into a single level problem.
Methods like Benders decomposition were used in several studies (Gao et al., 2005;
Martin 2007). Karush-Kuhn-Tucker conditions are also used to get the advantage of the
dual problem of user equilibrium, and force its associated constrained in the bi-level
problem to be feasible (Allende & Still, 2013; Dempe & Zemkoho, 2012; Wan, Mao, &
Wang, 2014). However, the single level problem is not necessarily simpler than the bi-
level problem to solve. The reason is the fact that the newly introduced constraints
increase the complexity of the solution algorithms for the leader problem.
The overall flowchart of the solution algorithms for the three decision makers in
model 2 (chapter 3) is presented in Figure 11. It should be noted that to keep the diagram
simple, the convergence criteria for user-level problems is not demonstrated in this
figure. The dashed box represents the inner bi-level problem of adversary and users. First,
at the designer level, a base feasible network enhancement vector (𝑦) is selected and
passed to the bi-level problem of adversary and users. The adversary chooses a vector of
decision to degrade the network (𝑧). The new network, which is affected by decision of
designer and adversary (𝑧 and 𝑦) will be transferred to the user level at the bi-level
problem of adversary-users. The user level problem, provides a new link flows vector (𝑥)
which is passed to the adversary level. The adversary-level objective function will be
computed to produce a new vector (𝑧), and will be delivered to the user-level. This
procedure is repeated until convergence. A second user-level problem, which shows the
behavior of user (𝑥′) to the decision of designer (𝑦) under normal condition, will be
evaluated (independent of 𝑥 and 𝑧). Then optimal solution of the bi-level problem of
74
adversary-users (𝑧) after convergence, along with users reaction under normal condition
(𝑥′) will be evaluated at designer level problem to determine the convergence. The
procedure will continue until the convergence of the designer level will be satisfactory.
Since there are multiple objective functions in designer level, a set on non-dominated
Start
No Converged
Yes
Investment Function
Convergence Criteria
OD Matrix
𝑧(𝑥, 𝑦) ,
Network Data 𝑥́ (y)
𝑥(y, z),
y 𝑥́ (y)
Alternative User Level w/o
Designer Level Disruptions
User Equilibrium
Minimize TSTT
Traffic Assignment
Minimize Vulnerabilities
𝑧(𝑥, 𝑦)
Link Capacity
Adversary Level Flows
Maximize Damage 𝑥(y, z(y))
Cost of Construction
Yes
z(y), y Converged
User Level
User Equilibrium 𝑥(y, z(y)) End
Traffic Assignment
In the following sections, a detail review of the solution algorithm for each
players is presented.
75
Algorithm for Users
Since the user level problem is a nonlinear derivable convex problem, it can be
solved using efficient heuristic algorithms such as the Frank-Wolfe algorithm (FW), bush
algorithm, gradient descent, and many other available gradient-based algorithms. The
model formulation for Frank-Wolfe algorithm (also known as reduced gradient and
convex combination algorithm) is provided in equations (2) through (6). The heuristic
search algorithm is used for convergence of the objective function to its optimal value
using the associated direction vector’s move size. The objective function is the sum of the
integrals of the link performance functions. The steps of the Frank-Wolfe algorithm are
as follows:
Step 0: Initialization
Perform all-or-nothing assignment based on ta = ta(0) ∀𝑎. A new flow vector {xa} will be
Step 1: Update
Set ta = ta(xa) ∀𝑎
Perform all-or-nothing assignment based on {ta}. A new auxiliary flow vector {x’a} will
be generated.
𝑥𝑎 +∝(x’𝑛 𝑛
𝑎 −𝑥𝑎 ) (72)
min 𝑧(𝑥) = ∑ ∫ 𝑡𝑎 (𝑤)𝑑𝑤
𝑎 0
76
The line search problem is solved using bisection algorithm (Bolzano search). The
converge criteria for the bisection method is defined as the distance between the lower
Step 4: Move
If the convergence criterion is met, stop and accept the current solution {𝑥𝑎𝑛+1 },
as the set of equilibrium link flows. If the convergence criterion is not met, set 𝑛 = 𝑛 +
√∑𝑎(𝑥𝑎𝑛+1 − 𝑥𝑎𝑛 )2
≤𝑘 (74)
∑𝑎 𝑥𝑎𝑛
convergence in the first iterations. Since the update step is “one-at-a-time” generically, it
77
Algorithm for Designer and Adversary
as discrete variables. In addition, since their objectives are nonlinear and non-convex,
they fall into the category of NP-hard problems. Therefore, efficient exact and heuristic
solutions do not exist, and approximate algorithms like meta-heuristic algorithms would
be the best approach for medium to large-scale sizes of these types of problems. Meta-
heuristics are a subfield of stochastic optimization that combines basic heuristic methods
in higher-level frameworks, and their goal is to search effectively and efficiently. They
are especially suited for problems with imperfect or incomplete information or limited
computation capacity.
computations’ (EC) algorithms, and one of the well-known and efficient technic in the
search heuristic that mimics the procedure of natural selection. The GA works by
toward better solutions. Each candidate solution has a set of properties, which is its
chromosome or also known as genotype. Its process routinely produces a new population
by mutating and altering the current population. Each population – also known as
The solution algorithm for the three-player model can be written as follows:
Step 0: Find an initial solution for the designer and user level
78
Step 2: Calculate vulnerability measures of links from attacker point of view, and
Step 2-1: Find the user equilibrium flows base on decision of designer and
The problem size can vary based on the number of the links in the network, the
number of the lanes, the number of nodes, and improvement approaches. A genetic
the next section. The problem of the upper level can be defined as a multi-objective
problem. In a multi-objective problem, more than one objective is optimized at the same
time. A good solution approach would be combining the objectives of the designer and
adversary into a monotonic objective function. This monotonic function has the set of
variables of both original objective functions, which can be solved using algorithms like
GA.
The designer’s objective function result will be a trial additional number of lanes
vector (𝑦𝑎 ). The designer’s decision variables are limited by the budget and resources
constraints. The decision of the designer and adversary modifies the network, and passes
it to the next player in the defined sequence of decision flow. The pseudo code of the
solution).
79
Step 1: Evaluate. Evaluate fitness of each organism/chromosome
– For the adversary, the fitness value is based on running traffic assignment
– For the designer, the fitness value is based on the new design of network
Step 3: Exploit. Eliminate the weak organisms and produce new organisms based
Step 4: Explore. Stochastically mutate organisms’ genes. Then repeat the process
starting at step 1.
Each individual variable, when encoded into genotype, is converted to the real
representation value of the objective function. Therefore, the output of the decoding
function makes the suitable variables for inserting into the objective function(s). A
chromosomes for the designer and adversary, using 3 bits of data, for 3 links is presented
in Figure 12.
80
0 0 0 1 1 0 1 1
y0 y1 y2 y3
0 0 0 1 1 0 1 1
y0 = 0 × 21 + 0 × 20 = 0 y1 = 0 × 21 + 1 × 20 = 1 y2 = 1 × 21 + 0 × 20 = 2 y3 = 1 × 21 + 1 × 20 = 3
algorithm process is first converted from the stream of binary values, into an integer
representation value. If the integer number was greater than the defined upper bound
value, the chromosome is not a legitimate solution, and it should be discarded. In this
case, a new chromosome should be generated and checked for validity. This process
should continue until a valid chromosome is found. This genotype representation can be
used as a general form that can be scaled to any desired boundary defined by the specific
problem. Depending on the boundaries defined by the specific problem, the integer
representation value can scale the output of a genetic algorithm to the size of decision
variables of the problem. The decision maker should be aware of the boundaries of the
decision variables of their problem. Therefore, they can be confident that the genetic
algorithm progress would not miss any solutions, because of not being able to touch all
81
the feasible space of the problem. In the above example, each variable is encoded into 2
bits of data. Therefore, the scale of the genotype of each variable is ranging from 0 to 3.
The boundaries of the phenotype variables in the actual problem are defined by the
feasible space of the problem. In this case, each non-negative variable has a maximum
value of three as their upper-level value. This value in the next step is used to scale the
genotype value range to the actual feasible space range. The constraint of the adversary
entity, which is the number of links that can be affected by its problem, can implement in
0 1 0 0 0 𝐵𝑧 = 1
z0 z1 z2 z3 z4
1 0 0 1 0 𝐵𝑧 = 2
z0 z1 z2 z3 z4
In this approach, the alleles with a binary value equal to one in the corresponding
chromosomes represent the attacked links. The upper bound of the adversary budget
82
the mutation operators should be avoided. The moves in the genotype space can be
To test the algorithm, and perform sensitivity analysis on the properties of the
genetic algorithm, bi-level network design problems were defined for two small test
networks. The network configurations are described in the next sections. The objective
based roulette is a good approach for the selection. The population size for the both
the GA operators and the sensitivity to their parameters, along two NDP test networks are
presented.
The first network is the simple Braess network (Figure 14). The Braess network
data, demand and link attributes are adopted from the Suwansirikul et al. (1987).
83
Figure 14. Test Network 1 – Braess Paradox Network (5-Link)
The upper-level objective function for this network is total system travel time
(∑ 𝑥. 𝑡). The lower level problem is based user equilibrium, defined in equations (2)
through (6), with the Bureau of Public Roads (BPR) travel time-congestion function
defined in equation (1). Table 8 presents the necessary input data that for the first test
network.
Table 8
Data for Test Network 1 (5-Link)
𝑥𝑎 4
(𝑥 )
𝑡𝑎 𝑎 , 𝑦𝑎 = 𝐴𝑎 + 𝐵𝑎 ( )
𝐶𝑎 + 𝑦𝑎
Arc a Aa Ba Ca da
1 4 0.60 40 2
2 6 0.90 40 2
3 2 0.30 60 1
4 5 0.75 40 2
5 3 0.45 40 2
84
Example Network 2 – 16 links network
The second test network consists of 6 nodes and 16 links (Figure 15). Two OD
pairs from nodes 1 to 6 and nodes 6 to 1 provide the flows on the network. Three cases of
travel demand levels are used for illustration where case 1 = 2.5 - 5.0, case 2 = 5.0 - 10.0,
and case 3 = 10.0 - 20.0. The travel time and investment cost functions used in problems
9 and 10 are adopted from Suwansirikul et al., (1987) together with the details of data
As it was described and shown in Figure 12, a good approach to encoding and
decoding the phenotype space can be using a one-dimensional series of binary variables.
The precision of the decoding and encoding operators directly depends on the number of
85
y y y
12 10 8
10 8
6
8
Expansion Unit
Expansion Unit
Expansion Unit
6
6 4
4
4
2
2 2
0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Iteration Iteration Iteration
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
8 8 8
6 6 6
Expansion Unit
Expansion Unit
Expansion Unit
4 4 4
2 2 2
0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Iteration Iteration Iteration
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
16 bits 32 bits 64 bits
Figure 16. Variation of reaching the solution by different size of bit-string in binary
chromosome representation.
Figure 16 presents the convergence of the solutions in the Braess network, by six
different bit-string size of decoding and encoding operators: 2, 4, 8, 16, 32, and 48 bit.
The higher size of bit-string as a binary representation (for values in the actual problem)
in phenotype space makes the solution become more accurate and provide better
precision. If the decision maker is concern about a specific level of accuracy in the
86
optimal solution, he or she can achieve that by choosing an appropriate length of bit-
string. It should be noted that choosing a big size for bit-string would result in higher
deciding the best bit-string size for the GA decoding and encoding operators. A smartly
chosen value can satisfy both accuracy and the computational time.
Elitism
Retaining the best individuals in unaltered from current generation to the next
generation, is called elitism or elitist selection. It helps to avoid losing the best solution
found, possibly by GA operators. This strategy guarantees that the solution quality
obtained by the GA will not decrease from one generation to the next.
GA Best Obj. value by Generations GA Best Obj. value by Generations GA Best Obj. value by Generations
1204 1208
1210
1208 1206
1206
1202 1204
1204
1202
1202
1200 1200 1200
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Generation Generation Generation
Keep
Series1 4 best individuals Keep only the best
Series1 Series1 No Elitism
individual
Figure 17. Elitist selection with different size, for population size 100
Figure 17 shows a variant of the best solution found in the ULP using Genetic
Algorithm, with three different size for Elitism: no elitism, keep one best solution, and
87
keep the best four solutions found. As it can be seen, the no elitism strategy, in some of
the generations, losses the best solutions found, which resulted in increasing the
minimization objective value. Retaining one best solutions in generations, avoids losing
the best solution, which resulted in monotonically decreasing the objective value.
Furthermore, increasing the size of retaining individuals to four further helped to have
faster convergence. In this example, the size of retaining individuals larger than four
percent did not significantly help the convergence of the algorithm procedure.
Crossover Operators
procedure of taking a portion of the solutions and generating offspring solutions from
them.
GA Best Obj. value by Generations GA Best Obj. value by Generations GA Best Obj. value by Generations
1212
1205.15
1204
1205.1
1210
1205.05 1202
1205
1208 1204.95 1200
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Generation Generation Generation
88
Figure 18 shows different types of crossover and their convergence rate. In this
example, the only active GA operators were the crossover operators and the selection. No
mutation operator is considered in this example, and the initial population of the solutions
Crossover (UPX) demonstrates the best results among other crossover operators. The
results are based on stochastic nature of GA, and they might not show the same
performance on different problems. However, it can give ideas to researchers that how
89
y y y
12 14 12
12 10
10
10
Expansion Unit
8
Expansion Unit
Expansion Unit 8
8
6 6
6
4 4
4
2 2 2
0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Iteration Iteration Iteration
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
8 8 10
8
6 6
Expansion Unit
Expansion Unit
Expansion Unit
6
4 4
4
2 2
2
0 0 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
Iteration Iteration Iteration
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
rate (crossover type: Uniform Crossover (UPX), mutation rate: 2%, population size 30)
Figure 19 shows the sensitivity of the model to crossover rates. The optimum
capacity expansion vector for this example was {7.1, 0, 0, 0, 3.39}. As it can be seen in
the figure, the optimum value for crossover rate was not at extremes 0 or 100 percent.
The best value for crossover rate for this specific problem was found at around 50
percent. Figure 20 presents the sensitivity of GA convergence to the mating rate. In this
test, the mutation rate was 0 percent and the initial populations of the solutions are
90
populated with random individuals. In Figure 20, the higher mating rate shows faster
1212
1208
1206
1204
1202
0 20 40 60 80 100
Generation
1216 1214
1214 1212
GA Best Obj. val
1212
1210
1210
1208
1208
1206 1206
1204 1204
0 20 40 60 80 100 0 20 40 60 80 100
Generation Generation
Series1
10% Mating rate 100%Series1
Mating rate
Figure 20. Crossover values by Mating rate (one-point Crossover)
91
Mutation Operator
Mutation is a genetic algorithm operator used to keep genetic diversity from one
similar to biological mutation. Mutation alters one or more gene values in a chromosome
92
GA Best Obj. value by Generations GA Best Obj. value by Generations
1210 1212
1208 1210
1208
1206
1206
1204
1204
1202 1202
1200 1200
0 20 40 60 80 100 0 20 40 60 80 100
Generation Generation
Series1
0% Mutation (Just Crossover) Series1 2% Mutation
1220 1230
1225
1215
GA Best Obj. val
1220
1210 1215
1210
1205
1205
1200 1200
0 20 40 60 80 100 0 20 40 60 80 100
Generation Generation
the mutation rates. As it can be seen, the mutation rate has significant effect on the speed
and quality of reaching the optimal solution. The low values of mutation may prevent
93
exploration of not-visited feasible spaces and may slow down the genetic algorithm
procedure. On the other hand, the high values of mutation may lead to a stochastic search
that does not keep track of previous results efficiently. For this specific problem, a value
Figure 22 and Figure 23 show the graphical results for the sensitivity to demand
for the following variables: the optimal capacity expansion vector, flow, and objective
values for the upper and lower levels. The sensitivity of test network 2 to demand was
tested for three different sets of O/D pairs. For this specific example, the higher demand
shows faster convergence rate. The network users’ reaction to the changes in capacity
expansion vector was evaluated by iteration. This reaction guides the direction of the
94
Expansion vectors (y)
y y y
6 10 8
8
6
Expansion Unit
Expansion Unit
Expansion Unit
4
6
4
4
2
2
2
0 0 0
0 20 40 0 20 40 0 20 40
Iteration Iteration Iteration
1 5 8 11 14 1 5 8 11 14 1 5 8 11 14
2 6 9 12 15 2 6 9 12 15 2 6 9 12 15
3 7 10 13 16 3 7 10 13 16 3 7 10 13 16
4 4 4
(a) Demand 1-6=2.5, Demand 6-1=5 (b) Demand 1-6=5, Demand 6-1=10 (c) Demand 1-6=10, Demand 6-1=20
flow:
flow flow flow
5 10 6
5
4 8
4
3 6
3
2 4
2
1 2
1
0 0 0
0 20 40 0 20 40 0 20 40
Iteration Iteration Iteration
(a) Demand 1-6=2.5, Demand 6-1=5 (b) Demand 1-6=5, Demand 6-1=10 (c) Demand 1-6=10, Demand 6-1=20
Converge
Convergence ofConverge
Upper Level objective value: Converge
98
105
96 210
94 100
92 205
95
90
200 90
88
86 195 85
0 20 40 0 20 40 0 20 40
Iteration Iteration Iteration
(a) Demand 1-6=2.5, Demand 6-1=5 (b) Demand 1-6=5, Demand 6-1=10 (c) Demand 1-6=10, Demand 6-1=20
Convergence ofConverge
Lower Level objective value: Converge Converge
80 170 80
78 165
Lower Level (TA) Obj. Value
76 160 75
74
155
72 70
150
70
145
68 65
66 140
64 135 60
0 20 40 0 20 40 0 20 40
Iteration Iteration Iteration
(a) Demand 1-6=2.5, Demand 6-1=5 (b) Demand 1-6=5, Demand 6-1=10 (c) Demand 1-6=10, Demand 6-1=20
95
Converge Converge Converge
80 400
150
Obj. Value (TSTT)
50
20 100
0 0 0
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Iteration Iteration Iteration
(a) Demand 1-6 = 2.5, (b) Demand 1-6 = 5, (c) Demand 1-6 = 10,
Demand 6-1 = 5 Demand 6-1 = 10 Demand 6-1 = 20
Figure 23. Total System Travel Time for Test Network 2 - (16-Link)
In the next chapter, implementations of the models presented in chapter 3 are discussed.
96
5 NUMERICAL EXPERIMENTS
This section discusses the results obtained from the methods covered in the
previous section. Several numerical experiments were conducted in order to evaluate the
performance of the models and observe the results. To implement the solution algorithms,
the designer and adversary level algorithms were coded and solved in MATLAB. The
user level algorithm was implemented in C++. The methods were processed on a 64-bit
Numerical experiments are conducted for the Sioux Falls network in order to
evaluate the proposed methodology. The network consists of 76 links and 24 nodes,
which are also, defined as the demand origin/destinations. Links and nodes ID and the
lengths of the links are provided in Figure 24a and Figure 24b respectively. It is assumed
that all the links in the initial network have three lanes. The links attributes and OD trips
were adopted from (Suwansirikul et al., 1987). All of the links of the network are
97
3
1 2
1
2 5 4 14
8 11 15
5
3 6
6 4
9 12
13 23 16 19
21 8
17
7 35 10 31 9
7
24 20
25 26 22 47 18 54
33 27 10 48 16
55
11
12 18
36 32 29 50
51 49 52
30
17
34 40 28 43
53 58 56
44 57 60
15
14 19
37 38
41 45
42 71 46 67
70
22 59 61
23
72
63
73 76 69 65
68
74 66 62
13 20
39 24 21
75 64
98
3.4
3.4
2.3 2.3 2.1 2.1
1.8 1 1.5
1.8 1 1.5
1.9 1.1
2.4 2.4 1.5 1.5
1.9 1.1
0.6 0.6
1.1 1.1
2.5
2 2 1.1 1.1
2.5
0.9 0.9 1.2 1.2
1 1.1 1.1
1.4
0.5 0.5 0.6 0.6
1.4
1.5 1 1.3
1.5 1 1.3
99
The O/D matrix for these experiments is presented in Table 9. Similar to test
network 1, two scenarios were performed on this network using different budget
Table 9
The trip rates for the Sioux Falls network (1000 veh/time unit)
D
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
O
1 .1 .1 .5 .2 .3 .5 .8 .5 1.3 .5 .2 .5 .3 .5 .5 .4 .1 .3 .3 .1 .4 .3 .1
2 .1 .1 .2 .1 .4 .2 .4 .2 .6 .2 .1 .3 .1 .1 .4 .2 0 .1 .1 0 .1 0 0
3 .1 .1 .2 .1 .3 .1 .2 .1 .3 .3 .2 .1 .1 .1 .2 .1 0 0 0 0 .1 .1 0
4 .5 .2 .2 .5 .4 .4 .7 .7 1.2 1.4 .6 .6 .5 .5 .8 .5 .1 .2 .3 .2 .4 .5 .2
5 .2 .1 .1 .5 .2 .2 .5 .8 1 .5 .2 .2 .1 .2 .5 .2 0 .1 .1 .1 .2 .1 0
6 .3 .4 .3 .4 .2 .4 .8 .4 .8 .4 .2 .2 .1 .2 .9 .5 .1 .2 .3 .1 .2 .1 .1
7 .5 .2 .1 .4 .2 .4 1 .6 1.9 .5 .7 .4 .2 .5 1.4 1 .2 .4 .5 .2 .5 .2 .1
8 .8 .4 .2 .7 .5 .8 1 .8 1.6 .8 .6 .6 .4 .6 2.2 1.4 .3 .7 .9 .4 .5 .3 .2
9 .5 .2 .1 .7 .8 .4 .6 .8 2.8 1.4 .6 .6 .6 .9 1.4 .9 .2 .4 .6 .3 .7 .5 .2
10 1.3 .6 .3 1.2 1 .8 1.9 1.6 2.8 4 2 1.9 2.1 4 4.4 3.9 .7 1.8 2.5 1.2 2.6 1.8 .8
11 .5 .2 .3 1.5 .5 .4 .5 .8 1.4 3.9 1.4 1 1.6 1.4 1.4 1 .1 .4 .6 .4 1.1 1.3 .6
12 .2 .1 .2 .6 .2 .2 .7 .6 .6 2 1.4 1.3 .7 .7 .7 .6 .2 .3 .4 .3 .7 .7 .5
13 .5 .3 .1 .6 .2 .2 .4 .6 .6 1.9 1 1.3 .6 .7 .6 .5 .1 .3 .6 .6 1.3 .8 .8
14 .3 .1 .1 .5 .1 .1 .2 .4 .6 2.1 1.6 .7 .6 1.3 .7 .7 .1 .3 .5 .4 1.2 1.1 .4
15 .5 .1 .1 .5 .2 .2 .5 .6 1 4 1.4 .7 .7 1.3 1.2 1.5 .2 .8 1.1 .8 2.6 1 .4
16 .5 .4 .2 .8 .5 .9 1.4 2.2 1.4 4.4 1.4 .7 .6 .7 1.2 2.8 .5 1.3 1.6 .6 1.2 .5 .3
17 .4 .2 .1 .5 .2 .5 1 1.4 .9 3.9 1 .6 .5 .7 1.5 2.8 .6 1.7 1.7 .6 1.7 .6 .3
18 .1 0 0 .1 0 .1 .2 .3 .2 .7 .2 .2 .1 .1 .2 .5 .6 .3 .4 .1 .3 .1 0
19 .3 .1 0 .2 .1 .2 .4 .7 .4 1.8 .4 .3 .3 .3 .8 1.3 1.7 .3 1.2 .4 1.2 .3 .1
20 .3 .1 0 .3 .1 .3 .5 .9 .6 2.5 .6 .5 .6 .5 1.1 1.6 1.7 .4 1.2 1.2 2.4 .7 .4
21 .1 0 0 .2 .1 .1 .2 .4 .3 1.2 .4 .3 .6 .4 .8 .6 .6 .1 .4 1.2 1.8 .7 .5
22 .4 .1 .1 .4 .2 .2 .5 .5 .7 2.6 1.1 .7 1.3 1.2 2.6 1.2 1.7 .3 1.2 2.4 1.8 2.1 1.1
23 .3 0 .1 .5 .1 .1 .2 .3 .5 1.8 1.3 .7 .8 1.1 1 .5 .6 .1 .3 .7 .7 2.1 .7
24 .1 0 0 .2 0 .1 .1 .2 .2 .8 .6 .5 .7 .4 .4 .3 .3 0 .1 .4 .5 1.1 .7
Figure 25 demonstrates the 14 candidate links that are selected for improvements
(Link numbers 13, 23, 30, 51, 27, 32, 34, 40, 49, 52, 53, 58, 39, and 74).
100
Figure 25. Links Included in Expansion (links with orange color) for the Sioux Falls
network
For the first experiment, the attacker upper bound was limited to 1 link. The three
level algorithm performed to solve this problem. The designer level problem is a bi-
objective problem of minimizing the total system travel time, and the total system travel
time of the damaged network, which is obtained by solving the attacker problem in the
them. In this experiment, the budget is assumed 20 million dollars. Their decision of
101
adding lanes to the current links is also constrained by an upper bound of a maximum of
considered; the chromosomal size for and attacker problem was 20. For the designer and
the adversary problems, respectively 100 and 20 generations were populated by the
102
6
x 10
8.15
8.1
8.05
7.95
7.9
7.85
7.8
7.75
7.7
0 20 40 60 80 100 120
Generation
Best fitness value for designer objective 1 (TSTT under damaged condition)
6
x 10
6.2
6.15
Best Fitness value of objective 2
6.1
6.05
5.95
0 20 40 60 80 100 120
Generation
Best fitness value for designer objective 2 (TSTT under normal condition)
Figure 26. Improvement of the two objectives at the designer level by generations
103
Figure 26 presents the trend of improving the two objectives of the
designer/defender. As it can be seen in the figure, the designer objective under normal
condition converged to its best solution after 16 generations. The convergence for the
objective of the designer under the degraded network, which is damaged by the attacker
best decision, occurs after 21 generations. The elitism selection feature of the genetic
algorithm, prohibit the best solutions to be eliminated. Therefore, the best solutions
104
6
x 10 Designer Pareto front
7
6.9
6.8
6.7
TSTT DMG
6.6
6.4
6.3
6.2 13: 1 23:1 27:3 30: 1 32:1 34:1 39: 1 49:1 51:1 52: 2 53:1 74:3 (o1:8.3e+06 o2:6. 2e+06)
6.1
8.2 8.4 8.6 8.8 9 9.2 9.4 9.6
TSTT 6
x 10
10st generation
100st generation
Figure 27. Individuals of the two objectives at the designer level by generations
105
Figure 27 presents the movements of the individuals by generations. As it can be
seen in the figure, individuals tend to be closer to the best ones by proceeding in
generations. The red dots represent the Pareto frontier solutions. The solutions with bad
fitness values are removed during the process of the genetic algorithm. In the last
iteration, all the individuals are confined to the best solutions. The best decision for the
designer is to choose the solution from the Pareto front of the last generation. In this
example, the optimal solution found was adding one lane to link 23, 34, 39, 40, 49, 53,
58, and adding 2 lanes to links 13, 27, 32, 51, 52, 74, and adding three lanes to link 30.
Figure 28. Decision of the designer (Number of lanes to be added to the network).
106
The final decision of the designer is represented in Figure 28. The optimal
decision of the adversary entity under the initial network and improved network
conditions are provided in Figure 29. The numbers on links in the figure show the total
zz 0 zz e
1 2 1 2
3 4 5 6 3 4 5 6
9 8 7 9 8 7
12 11 10 16 18 12 11 10 16 18
17 17
14 15 19 14 15 19
23 22 23 22
13 24 21 20 13 24 21 20
0 2.00 0 2.00
a) Initial network b) Improved network
The initial total system travel time under normal conditions is reduced from
7.48×106 minute to 5.98×106 minutes in the improved network. Considering the degraded
network after the attack, the total system travel time is 7.80×106. This value is
107
comparable to the value for the initial network after the attack, which is 10.89×106
minutes. Figure 30 presents the changes in the total system travel time after the
improvement. The additional system cost due to the attacks is represented by red stacks
over the network under normal condition, which represented by blue bars. The addition
of the new lanes to the network could reduce the total system travel time by 20.1 percent.
However, the improvement by the proposed model is more significant from the
vulnerability perspective. The new improved network could reduce the imposed
additional system cost from 3.41×106 minutes to 1.82×106 minutes, which is 46.4 percent
less.
108
6
x 10
12
10
8
TSTT (min)
0
Initial Network Improved Network
conditions
The volume of traffic flows in the Initial and Improved network, and before and
after the disruptions is presented in Figure 31. The addition of the lanes to the links, make
them more attractive to the travelers. As it can be seen in Figure 31b and Figure 31c,
damages to the links make traveler reroute and use alternative way, which makes higher
109
x0 3.6 xe
4.5
4.5 3.6
8.1 8.1 6.0 6.0 7.2 7.2 5.4 5.4
5.6
9.8 8.5 5.5 6.8 4.4
7.1 7.3 5.6 5.4
13.7 16.6 9.2
14.5 23.3 7.6
17.6 20.7 8.4
10.0 13.3
13.5 15.7 8.8
16.5 11.5
8.119.3 13.2 11.8
10.5 17.0
10.5 13.1 5.8 6.1 4.7 13.4
7.6 13.3 14.3 10.0 11.7 6.3
8.6 17.2
23.6 15.8 8.8 10.3 13.0 16.7
0.015.2 10.5 8.9 21.5 17.7
7.5 15.8 14.5 21.2
8.9 21.5 11.1 18.8
8.1 19.5 11.5 16.3 11.5 29.1 9.6 17.8
12.9 11.6 11.2
8.1 10.3 15.7 14.6
8.5
10.9 14.5 19.5 0.0 11.2 11.0 16.0 20.0 21.4
Figure 31. Flows in the Initial and Improved network, before and after the disruptions
(veh/day)
110
The cost of travel in terms of travel times are provided in Figure 32. As it can be
seen in the figure, the travel times are reduced in the improved network. The model
increased the travel times of the damaged links, which penalize the use of them by the
users. The links with hotter color in Figure 32 represent the higher travel times. The
damages to the link 43 in the initial network reroute the users to its available alternative
paths. Therefore, a percentage of the traffic volume transferred to other links, which
111
6.00 t0 6.00 te
1 2 1 2
6.00 6.00
4.014.01 6.576.60 4.014.01 6.066.06
9.65
9.67 14.69
14.83 5.69
6.30 9.81
10.42
5.68
5.72 10.73
10.78 2.06
2.06 6.37
5.74 8.62
8.87 2.04
2.05
17 17
13.69 13.85 13.72 13.81 10.52 10.27 10.37 10.47
7.447.41 4.694.80
8.46
12.37 23 22
23 22
7.75 7.60 8.62
12.24 9.46 9.52
7.87
7.71
4.884.14 4.494.54
3.763.72 4.204.21
7.71
7.71
8.01 8.85 9.96
17.62 11.92 8.17 13 24 21 20
13 24 21 20 10.77
17.66 7.36 8.64
11.75 8.08
0 121.00
0 121.00
1 2 1 2
6.00 6.00
4.024.01 6.137.71 4.014.01 6.196.06
10.51
6.32 19.13
33.37 5.04
8.27 17.68
12.06
6.73
3.75 11.94
17.84 2.03
2.08 103.00
3.64 18.99
12.29 2.21
2.10
5.12
17.96 4.324.99
15.55 9.01
23 22 23 22
10.02 16.92 6.63 10.05 8.42 12.98
7.71 8.97
5.513.37 5.922.95 5.076.96 5.116.73
17.26 7.71
c)
0 Initial network after the disruption121.00 d)
0 Improved network after the disruption
121.00
Figure 32. Travel times in the Initial and Improved network, before and after the disruptions
(min)
112
The second experiment considers that two links can be damaged by the adversary entity.
The both objective of the designer are converged after 20 generations. The goals of the
individual solutions by generations are provided in Figure 33. Similar to the previous
case, the chromosomes move toward the best solutions during the process of evolutionary
optimization.
113
6
x 10 Designer Pareto front
6.75
6.7
6.65
6.6
TSTT DMG
6.55
6.45
6.4
6.35
13: 1 23:1 27:1 32: 1 34:1 39:1 40: 1 49:1 51:1 52: 1 53:3 58:1 74: 3 (o1:1. 8e+07 o2:6.3e+06)
6.3
1.75 1.8 1.85 1.9 1.95 2
TSTT 7
x 10
10st generation
100st generation
Figure 33. Individuals of the two objectives at the designer level by generations
114
The optimal solution for the designer is presented in Figure 34. The optimal
Figure 35. The optimal solution found for the adversary was changed from the links 43
y0A
1 1
1
2
1
2 3
1
2 2
3 2
2
2
0 4.00
Figure 34. Decision of the designer (Number of lanes to be added to the network).
115
zz 0 zz e
1 2 1 2
3 4 5 6 3 4 5 6
9 8 7 9 8 7
1
12 11 10 16 18 12 11 10 16 18
17 17
14 15 19 14 15 19
23 22 23 22
13 24 21 20 13 24 21 20
0 2.00 0 2.00
a) initial network b) improved network
Figure 35. The optimal decisions of the attacker for the initial and improved networks
Compared to the results of the previous experiment, links 34, 40, 52, 53, and 58
are collecting more investment in this test. The total system travel time for the initial
network and the improved network are 7.48×106 minutes and 5.97×106 minutes
respectively. The values for the degraded network are 13.53×106 minutes before
improvements, and 11.90×106 minutes after the improvements. Increasing the adversary
entity’s budget, results in a larger total cost to the users in terms of damaging the network
elements.
116
4.5 x0 3.6 xe
4.5 3.6
8.18.1 6.06.0 7.27.2 5.45.4
10.09.9 14.214.0
9.7 8.8
9.6 8.7 8.7 8.7 7.6 7.6
7.0 7.0
7.9 7.9 8.6 8.6 8.5 8.5 8.8 8.7
7.0 7.0
6.0 3.8
8.4 9.5 6.9 5.8 8.6 4.9 4.5 8.2
12.019.4 15.613.5
11.0 15.3
8.3 8.7 9.9 11.3 8.0 7.4
7.6 8.8
9.8 7.8 9.6 8.8 9.4 7.4 10.89.1
5.0 6.9
0 27 .50 0 27 .50
g) Initial network after the disruption h) Improved network after the disruption
Figure 36. Flows in the Initial and Improved network, before and after the disruptions
(veh/day)
117
The link volumes for initial and improved network, before and after the
disruptions are provided in Figure 36. Similarly, Figure 37 provides the travel time of the
links, under these volumes. The links with warmer color represent higher travel time.
Since the damage to the links is modeled by increasing the travel time to a higher value,
the travelers are not interested in using these links. However, there still exist low traffic
volumes that use these links in the degraded network. This is due to the degree of damage
to the network, in this test, is significant enough that a portion of users utilizes these links
despite the higher cost. In other words, the damages to the mentioned links create high
congestion on some of the alternative routes. The improved network influences drivers to
use links other than centralized links. Hence, the potential damages that the attacker can
118
6.00 t0 6.00 te
1 2 1 2
6.00 6.00
4.014.01 6.576.60 4.014.01 6.066.06
9.65
9.67 14.69
14.83 6.26
6.53 9.90
10.33
5.68
5.72 10.73
10.78 2.06
2.06 5.88
6.19 9.41
9.59 2.04
2.04
17 17
13.69 13.85 13.72 13.81 8.39 9.35 9.97 9.65
7.447.41 3.414.77
9.78
12.37 23 22
23 22
9.42 7.24 7.12
12.24 9.46 9.52
7.71
7.71
4.304.40 4.364.34
3.763.72 4.204.21
7.71
7.71
7.47 9.72 10.21
17.62 11.92 8.17 13 24 21 20
13 24 21 20 7.57
17.66 9.48 10.06
11.75 8.08
0 121.00
0 121.00
1 2 1 2
6.00 6.00
4.014.02 7.816.36 4.014.00 5.51
10.54
13.35
7.37 18.43
12.73 5.80
105.24 16.17
72.47
6.22
4.27 10.65
9.92 2.05
2.04 5.69
4.15 9.65
10.78 2.05
2.07
13.50
80.62 4.074.39
18.27 57.16
23 22 23 22
8.52 9.41 13.12 19.77 7.99 6.92
8.81 11.64
6.163.67 5.384.45 5.483.38 7.474.72
5.72 7.56
0 121.00 0 121.00
g) Initial network after the disruption h) Improved network after the disruption
Figure 37. Travel times in the Initial and Improved network, before and after the disruptions
(min)
119
Similar to the previous tests, several tests were conducted using higher available
budgets to the adversary entity. To be able to compare the outcomes, the other parameters
were kept similar to the previous tests. Since the size of the feasible space for the
good solution would require more effort. To address this issue, the number of generations
for the adversary was increased to 30 for the 𝐵𝑧 equal to 5 links and more. This change
increased the total number of user equilibrium processes from around 2 million to 3
million runs; therefore, more process time was required for these tests.
The solutions that were found in the designer and the attacker problem (before
can be seen that the pattern of the investments is changing with increasing the available
budget to the adversary. Different patterns of allocating resources slightly push the
vehicles to reroute their initial path to a new lower cost path. These marginal changes will
result in different patterns of attack by the adversary entity. However, as it will be seen in
the next section, the influence of the designer’s decision on overall system cost is
120
Decision of the adversary
Decision of the designer
𝐵𝑧 Initial network Improved network
zz 0 zz e
1 2 1 2
3 4 5 6 3 4 5 6
9 8 7 9 8 7
12 11 10 16 18
12 11 10 16 18
1
17
17
14 15 19
14 15 19
23 22
23 22
13 24 21 20
13 24 21 20
0 2.00 0 2.00
zz 0 zz e y0A
1 2 1 2
3 4 5 6 3 4 5 6
1 1 1
9 8 7 9 8 7
1 1
12 11 10 16 18 12 11 10 16 18
2
1
2 1
17 17
2 2
1
2 3
3 2
14 15 19 14 15 19
23 22 23 22
2
13 24 21 20 13 24 21 20
z 0
z
z e
z
y0A
1 2 1 2
3 4 5 6 3 4 5 6
1 1 1 1 1
9 8 7 9 8 7
1 1 2
12 11 10 16 18 12 11 10 16 18
3 1
1
2 3
17 17
1 1
1
14 15 19 14 15 19
23 22 23 22
1
13 24 21 20 13 24 21 20
0 4.00
0 2.00 0 2.00
121
Decision of the adversary
Decision of the designer
𝐵𝑧 Initial network Improved network
z 0
z
z e
z
y0A
1 2 1 2
3 4 5 6 3 4 5 6
1 1 1
1 1
9 8 7
9 8 7
1 1
12 11 10 16 18
12 11 10 16 18 1
1
4 17
2
2
3 2
17
1 3
1
1
14 15 19
14 15 19
1 1
23 22
23 22
13 24 21 20
13 24 21 20
2
0 2.00 0 2.00
0 4.00
z 0
z
z e
z
y0A
1 2 1 2
3 4 5 6 3 4 5 6
1 1 1
9 8 7 9 8 7
1 1
12 11 10 16 18 12 11 10 16 18
3
2
2 1
5 1
17 17
1 3
1
1 1
2 2
14 15 19 14 15 19
23 22 23 22
1
1
13 24 21 20 13 24 21 20
1
1
0 4.00
0 2.00 0 2.00
z 0
z
z e
z
y0A
1 2 1 2
3 4 5 6 3 4 5 6
1 2 1
1 9 8 7 9 8 7
12 11
1
10 16 18 12 11 10 16 18
1
1 1 1
1
6 17
1
17
1
3 2
1 1 1 1 3
14 15 19
1 14 15 19
1
1
23 22 23 22
13 24 21 20 13 24 21 20
1
1 1 1
122
Decision of the adversary
Decision of the designer
𝐵𝑧 Initial network Improved network
z 0
z
z e
z
y0A
1 2 1 2
3 4 5 6 3 4 5 6
1 3 1
1 9 8 7 9 8 7
1 2
12 11 10 16 18
12 11 10 16 18
1 1
1
2
7 1
3 1
1
17
1 1
17
2 1
14 15 19
1 14 15 19
1
1
23 22
23 22
1
13 24 21 20
13 24 21 20
1
1 1
0 4.00
0 2.00 0 2.00
Figure 38. The optimal decisions of the designer and the attacker
Figure 39 presents the flows on the links after the network is exploited by the
adversary entity, under different budget available to him. Taking into account the links
that are attacked in Figure 41, the higher costs of the damaged links push the vehicles to
change their route to the available alternative routes. Providing the adversary with a
higher available budget, makes the damages to the network more severe; in which, due to
the new damages, the available alternative paths will be more restricted.
123
𝐵𝑧 Initial network Improved network
6.9 x0A
4.2 xeA
5.6
9.8 8.5 5.5 6.8 4.4
7.1 7.3 5.6 5.4
13.7 16.6 9.2
14.5 23.3 7.6
17.6 20.7 8.4
10.0 13.3
13.5 15.7 8.8
16.5 11.5
8.119.3 13.2 11.8
10.5 17.0
10.5 13.1 5.8 6.1 4.7 13.4
7.6 13.3 14.3 10.0 11.7 6.3
8.6 17.2
23.6 15.8 8.8 10.3 13.0 16.7
0.015.2 10.5 8.9 21.5 17.7
7.5 15.8 14.5 21.2
8.9 21.5 11.1 18.8
8.1 19.5 11.5 16.3 11.5 29.1 9.6 17.8
12.9 11.6 11.2
1 8.1 10.3 15.7 14.6
8.5
10.9 14.5 19.5 0.0 11.2 11.0 16.0 20.0 21.4
8.7 13.0 18.6
10.7 11.4
10.6 14.3 21.8 7.5 20.8
13.3 16.5 6.6 22.2 17.0 15.4
30.3 8.5 19.9
7.7 7.4 19.5 11.7 7.8 8.5 16.9 19.0
10.5 8.5 8.2 9.8
8.9 10.8 7.2 8.9
7.0 7.7
9.4 7.4 9.9 7.0 9.1 10.2 9.4 10.4
10.2 7.0
14.7 11.1 6.2 14.3 9.3 7.3
11.5 9.8 7.9 16.0 9.8 6.8
0 33 .00 0 33 .00
6.0 3.8
8.4 9.5 6.9 5.8 8.6 4.9 4.5 8.2
12.019.4 15.613.5
0 27 .50 0 27 .50
124
𝐵𝑧 Initial network Improved network
3.7 4.4
10.96.6 4.48.7 6.79.4 6.33.6
10.010.1 11.89.7
0 28.60 0 28.60
9.0 x0A
4.5 x0
5.5
4.5 11.98.4 5.79.2
8.18.1 6.06.0
12 .6 18 .6 10 .8
14 .0 18 .0 8.8
16 .2 20 .4 8.5
14 .0 18 .0 8.8
14. 8 10.7 14 .4 20 .2
1515
.8 .8 1212
.5 .5
3.8 13 .3
6.9 12 .1 12 .1 12 .0 8.0 6.3
10 .0 10 .0 5.2 5.3 9.7 13 .5
6.8 12 .0
20 .0 21.9 8.8 9.0 15 .3 15.1
2121
.7 .8 8.48.4 1515
.8 .9
7.6 11.3 10 .1 14 .4
8.4 17 .7 11.1 15 .3
8.9 14 .8 13 .2 22 .2
8.4 17 .6 11.0 15 .3
8.1 7.2 10 .6 15.5
11.7
11.7
8.3
4 8.1
8.9 9.5 19 .8 22 .5
9.8 9.8 23 .1 23 .2 27 .3
19 .1
6.4 10 .2
10 .09 .9
19 .1 5.8 10 .6
9.1 19 .0
12 .3 12 .4 14 .1 15 .5 7.4 20 .3
9.0 19 .1 19 .0
8.7 10 .9 4.6 15 .2
8.48.4 1818
.4 .4 13 .0 7.1
9.7 13 .0
9.6 8.7 8.7 11.5
15 .4
7.0 7.4
7.97.9 8.68.6 8.1 4.1 9.4 8.5
7.0
11.1 10 .3 6.3 12 .3 14 .0 14 .3
11.1 10 .3 6.2 11.1 12 .0 7.0
0 30.80 0 30.80
125
𝐵𝑧 Initial network Improved network
8.7 2.6
7.610 .7 8.65.5 12 .69.7 0.02.9
15 .9 19 .7 8.5 13 .9 13 .9 4.1
16 .3 20 .4 10 .2 20 .7 21 .4 11.5
1415
.8 .8 1813
.7 .9 1313
.5 .5 5.8
10 .3
8.3 14 .4 7.0 7.5
11.0 14 .4 4.9 5.3 8.8 12 .7 6.5 7.3
12 .8 21 .0 8.5 12 .6
1622
.4 .0 11.58.9 2216
.7 .1 2021
.1 .7 8.57.6 1510
.8 .7
8.7 17 .2 12 .2 15 .1 9.5 12 .9 10 .8 13 .3
7.6 17 .7 16 .9 17 .2 9.4 19 .1 12 .3 20 .6
7.2
5 8.1
1610
.0 .7
9.7
1317
.0 .9
4.7
11.59.6 9.712 .2
9.8 21 .7 17 .4 15 .5 28 .1
30 .0
12 .5 15 .0 15 .6 19 .5 9.2 22 .0
7.7 14 .0 21 .2 15 .5
8.08.0 1718
.3 .1 10 .6
9.6 1.8
15 .6
10 .5 12 .2
8.2 7.2 13 .0 11.2 11.4 7.4
13 .6 15 .9
8.36.0 12 .89.0 9.07.0 4.39.9
6.9 6.7
0 33.00 0 33.00
3.7 7.5
10 .17.7 4.67.0 6.19.1 8.05.0
18 .1 18 .7 9.1 14 .3 18 .6 8.5
14 .2 16 .6 6.6 4.8 11.6 10 .0
1717
.6 .1 6.9
11.8 4.1
12 .6 1814
.6 .0
5.2 11.0 9.2 12 .6
17 .711.4 4.0 5.9 9.6 3.1 8.6 11.2
6.5 9.0 3.8 11.9
2223
.3 .2 6.68.2 1214
.1 .1 1721
.7 .0 8.48.5 1516
.4 .1
9.1 17 .7 10 .1 8.2 8.7 16 .3 12 .6 15 .6
8.5 18 .9 19 .2 19 .4 14 .7 11.4 10 .0 14 .1
8.8 13 .2
11.415 .2 1213
.4 .8
6.7 5.8
6 6.4 10 .1 17 .5 24 .2 13 .4 5.2 17 .9 15 .9
12 .3
18 .2 12 .1
20 .9
9.1 19 .2 15 .4 9.2 13 .3 15 .1
22 .5 15 .7 9.7 19 .7 17 .8 17 .4 13 .4 21 .3
6.1 17 .1
6.2
10 .5 1522
.6 .1 9.75.7 1719
.6 .3
9.5 10 .9
10 .4 6.1 11.5 13 .4 8.1 8.8
7.9 6.9
6.3
11.5 8.5
11.2 10 .38.7 9.89.9
5.0 7.8
12 .4 9.8 7.0 16 .3 11.9 6.9
19 .3 11.4 6.0 16 .8 14 .0 8.7
0 26.40 0 26.40
126
𝐵𝑧 Initial network Improved network
4.9 9.1
10.77.4 5.7 9.1 6.311.0 9.0 4.3
19.912.2 25.211.9
0 27.50 0 27.50
Under higher values of 𝐵𝑧 , still the main target area for the adversary are in the
more centralized links. Therefore, the damages limit the number of paths available
through the center of the network. This new condition, force the vehicles to use the more
paths with less centralized links, and the surrounding areas attract the excess flows.
127
𝐵𝑧 Initial network Improved network
1 2 1 2
6.00 6.00
4.024.01 6.137.71 4.014.01 6.196.06
10.51
6.32 19.13
33.37 5.04
8.27 17.68
12.06
6.73
3.75 11.94
17.84 2.03
2.08 103.00
3.64 18.99
12.29 2.21
2.10
17 17
18.83 50.46 9.92 106.00 9.19 8.88 7.75 10.33
5.12
17.96 4.324.99
15.55 9.01
23 22 23 22
10.02 16.92 6.63 10.05 8.42 12.98
7.71 8.97
5.513.37 5.922.95 5.076.96 5.116.73
17.26 7.71
0 121.00 0 121.00
1 2 1 2
6.00 6.00
4.014.02 7.816.36 4.014.00 5.51
10.54
13.35
7.37 18.43
12.73 5.80
105.24 16.17
72.47
6.22
4.27 10.65
9.92 2.05
2.04 5.69
4.15 9.65
10.78 2.05
2.07
17 17
21.01 75.99 11.69 111.35 5.85 19.45 17.12 8.60
13.50
80.62 4.074.39
18.27 57.16
23 22 23 22
8.52 9.41 13.12 19.77 7.99 6.92
8.81 11.64
6.163.67 5.384.45 5.483.38 7.474.72
5.72 7.56
0 121.00 0 121.00
128
𝐵𝑧 Initial network Improved network
6.01 t A 6.00 t A
0 e
1 2 1 2
6.00 6.00
4.034.00 5.47
12.10 4.004.02 6.915.21
8.70
114.07 19.45
87.44 6.13
114.73 11.28
123.48
5.70
4.85 10.77
12.87 2.05
2.08 5.33
4.52 10.26
6.38 2.10
2.03
7.457.73 12.65
6.86
5.96
28.96 16.51
7.40 7.13
10.89 10.59
7.43
66.26 56.78
23 22 23 22
12.06 14.82
2.573.07 9.264.73 6.413.22 6.234.74
7.71 7.28
0 132.00 0 132.00
6.01 t A 6.00 t A
0 e
1 2 1 2
6.00 6.00
4.044.01 6.26
14.06 4.014.02 7.675.25
8.56
125.82 24.23
89.37 6.09
126.25 13.17
135.64
4.91
5.76 11.87
12.54 2.05
2.05 5.39
5.81 11.62
8.08 2.08
2.01
4 13.09
17.16
7.06
25.20
9.92
15.05
7.13
9.70
17 17
10.53 12.58 10.19 12.92 5.39 24.41 10.66 10.77
2.958.01 2.906.86
9.73
18.21 103.80
5.82 14.99
12.64 103.00
6.26
31.75 20.29
23 22 23 22
20.64 31.09 6.44 18.44 17.49 7.90
68.02 79.67
3.383.92 2.125.19 4.003.23 2.035.20
10.90 7.28
0 154.00 0 154.00
129
𝐵𝑧 Initial network Improved network
6.00 t A 6.00 t A
0 e
1 2 1 2
6.01 6.00
4.014.03 11.73
6.14 4.054.02 105.00
5.09
8.57
9.68 65.24
21.52 7.50
121.42 7.96
132.63
3.87
5.80 25.30
12.37 2.27
2.07 4.97
5.67 11.13
8.86 2.06
2.01
11.74
6.77 3.565.88
8.24
8.18 7.79
8.63 16.91
12.66 103.02
6.15
15.47 25.10
23 22 23 22
44.03 77.38
4.132.60 12.82
4.64 4.913.08 2.145.85
7.56 7.28
0 176.00 0 176.00
6.00 t A 6.00 t A
0 e
1 2 1 2
6.00 6.01
4.024.01 5.568.06 4.004.01 10.08
5.77
12.18
11.45 161.57
12.19 105.06
5.61 64.38
22.21
5.97
6.49 7.14
10.23 2.02
2.04 4.19
5.32 10.65
11.18 2.06
2.07
6 19.58
160.37
8.79
23.46
26.57
10.18
5.01
6.56
17 17
151.01 15.23 151.0015.24 14.92 106.69 8.79 136.59
14.69
62.93 5.75
106.83
5.49
16.25 6.14
15.60 13.02
5.06 8.11
10.40
11.79 17.73
23 22 23 22
15.11 5.29 20.70 34.50 8.04 9.83
9.47 7.56
2.719.77 4.148.34 7.024.55 5.655.92
5.71 9.15
0 176.00 0 176.00
130
𝐵𝑧 Initial network Improved network
6.01 t A 6.00 t A
0 e
1 2 1 2
6.00 6.01
4.034.01 6.31
13.36 4.004.03 13.14
5.42
15.31
118.25 28.45
98.00 5.82
6.19 79.70
13.70
6.35
4.28 12.35
13.30 2.06
2.07 3.84
7.27 35.18
10.47 2.27
2.03
88.66
14.32 72.72
13.30
19.96
8.16 10.24
8.10 13.16
9.41 15.03
7.54
20.46 10.88
23 22 23 22
7.98 5.99
8.55 7.87
0 231.00 0 231.00
Figure 40 presents the travel time of using the links of the initial network and the
network after exploitation by the adversary. The travel times are presented by different
budget available to the adversary entity. It should be noted that a higher value for travel
time for a specified link means that fewer vehicles are motivated to use that link. The
links with very high cost do not attract traffic flows. Therefore, they do not directly
change the performance of the network, in terms of the total system travel time; and they
have an indirect influence on the total cost of the system, by pushing the vehicles to use
the alternative routes. The improvements of the network, reliefs the additional pressures
131
due to the damages, especially for the cases with lower values for 𝐵𝑧 . However, as the
budget limit for the adversary increases, the magnitude of effects of the improvements
25
MILLIONS
20
15
TSTT (MIN)
10
0
1 2 3 4 5 6 7
𝐵𝑧
Figure 41. Travel system travel times by 𝐵𝑧 in the Initial and Improved network after
the disruptions
Figure 41 presents the total system travel time in the test networks, by various
values for the adversary’s budget constraint 𝐵𝑧 . From the figure, it can be argued that the
damage to the network is increasing almost linearly by increasing the budget limit for the
adversary. The influence of the methodology to reduce the effects of the adversary’s
attacks is declining with increasing the adversary’s budget. However, large values for the
132
adversary budget should be infrequent in reality. In addition, the test network size is not
large, which makes it more vulnerable to the damages. In larger scale networks, more
alternative routes are available to the users, which alleviate the pressure of the damages
to the network. In adversary’s budget equal to one, the total system travel time after
7 (𝑙𝑖𝑛𝑘) were found 12.1%, 6.4%, 6.8%, 4.8%, 6.1% and 2.7%. The most benefits for the
The computational time required for each run with a different budget limit for the
adversary was increased from around 20 hour for 𝐵𝑧 ≤ 5, to 28 hours for 𝐵𝑧 > 5 (link).
The additional computational time was due to the incensement in the number of
expensive user equilibrium process. In the next section, a zero-sum based game model
will be discussed.
Due to the nature of this study, a very basic network may not be an appropriate
sample. The method requires damaging a portion of the network, and small networks
consist of a few links, usually lose its connectivity. Hence, it may not represent the real
world. The goal of this research is to improve the robustness of larger networks. Thus for
the first example, to keep the simplicity, a network consists of 6 nodes and 16 links was
selected (Figure 42). Two OD pairs from nodes 1 to 6 and nodes 6 to 1 are considered.
The trips demand from node 1 to 6 is 2.5 units of flow (for example, thousands vehicle
133
per hour (vph)), and from node 6 to 1 is 5.0 units of flow. The travel time and investment
cost functions used in problems 9 and 10 are adopted from (Suwansirikul and Friesz,
1987) together with the details of data input for each link. The number of existing lanes
Two tests were conducted on this network. First, considering the 𝐵𝑑 = 100
budget unit (e.g. 1000 vph.mile.lane, or in million dollars using a monetary conversion
factor) for the designer, and the maximum number of links that the adversary could
disable considered as 𝐵𝑧 = 1 (link). The best solution for the designer is shows on Table
10.
134
Table 10
The local optimum solution found for the first scenario on the 16 link network, 𝐵𝑧 =
1 (𝑙𝑖𝑛𝑘)
Link No. 6 7 9 11 12 15 TSTTinit, norm TSTTinit, dmg TSTTimp, norm TSTTimp, dmg
The TSTT for the best found solution after the disruption was 96.76 time units
(e.g. thousands vehicle hour during the analysis period). This is 18.4 percent
improvement compare to the base network. One interesting finding was the adversary in
most of the cases, only attacked link 14. Therefore, there exist varieties of solutions for
the designer, with a relatively similar objective value of the upper level. This would give
the designer larger available good solutions, and he could consider other factors such as
The second test was conducted similar to test one, with the adversary
constraint 𝐵𝑧 = 2. The best solution for the designer is shows on Table 10 (the time units
Table 11
The local optimum solution found for the second scenario on the 16 link network and
for 𝐵𝑧 = 2 (𝑙𝑖𝑛𝑘)
Link No. Solution No. 1 3 5 6 11 12 15 TSTTinit, norm TSTTinit, dmg TSTTimp, norm TSTTimp, dmg
Number of
1 2 1 1 2 3 2 3 84.38 129.03 78.29 109.74
lanes to be
2 0 3 1 3 3 1 3 84.38 129.03 78.93 113.07
added
The TSTT in the degraded network for the best-found solutions in the second test
was 109.74 time units for the solution number 1, and 113.07 time units for the solution
135
number 2. This shows a 14.9 percent improvement compare to the base network, for the
solution number 1. In the first solution, the adversary selected links 9 and 16 to be
disabled, and in the second solution, he selected links 9 and 14. It should be noted that the
behavior of response of adversary to the decision of the designer is changed in the second
test. A small variation in the designer’s decision would result in changing the decision of
the adversary. Hence, one can say that the higher resources available to the adversary
make it more sensitive to the decision of the designer. This finding needs to be further
investigated.
For the first scenarios on test network 2, the available budget to the designer
considered as 𝐵𝑦 = 300 budget unit. The first test was done using the one element that
can be affected by adversary entity. A stochastic search was performed in order to find
the best possible action for designer, against the attacks from enemy. The analysis found
link 56 as the link that would be attacked under the optimal design solution of the
designer of the network. The best solution according to the available budget for the
136
Table 12
The local optimum solution for the first scenario on the Sioux Falls network, 𝐵𝑧 = 1
(link)
Link # Number of lane to be added
16 3
19 2
26 4
39 4
74 5
Under this design, the most vulnerable link would be link number 56. The TSTT
for the best solution of the first scenario was 8.83e6 time units. It should be noted that the
rest of the solutions that were examined, showed more damages after the attack of the
enemy entity.
The second test was done considering two links as the enemy entity budget, and
for the designer considered as 𝐵𝑦 = 100 budget unit. The budget for designer selected
lower compare to first scenario, because with the current equation for the designer budget
constraint, the results shows a large number of links that would be selected to be
improved on this scenario. The feasible space and computational time is much higher
than the previous scenario. Selecting a set of potential links for the designer and
adversary could significantly reduce the size of the problem. The total 2000+ solutions
were examined in the analysis of the second scenario. It required running about half a
million-traffic assignment to obtain the local optimal design. The total time required for
137
this analysis was 20 hours. The results found in the second scenario on Sioux Falls
Table 13
The local optimum solution for the first scenario on the Sioux Falls network, 𝐵𝑧 = 2
(link)
Link # Number of lanes to be added
19 2
29 1
39 1
48 1
74 1
The TSTT in the best solution found for the second scenario was 26.45e6 time
units. This value increased significantly, due to the higher extent of damages and less
available budget for improvements. The most vulnerable links in this analysis were link
number 7 and 74. Link 19 was found to receive the most planner’s budget to have a more
robust network.
138
6 CONCLUSIONS AND FUTURE RESEARCH
agencies. Various objectives could be defined in order to improve the different indicators
of the performance of the transportation network, for example maximizing the social
welfare. One of the important objectives that the transportation agencies should consider
network against vulnerabilities. From this point of view, damages to the networks could
adversary may look for vulnerabilities in the network to degrade its performance. At the
the intelligent adversary may not help reduce the vulnerabilities and their effects, or
To address this issue, this research presented two models for designing robust
is presented. The planner of the network is assumed to look for investing the assigned
budget on the links/projects of the network while the enemy was assumed to damages the
links. The results of the optimization of these players are passed to the lower level. The
lower level problem provides flow vectors based on user equilibrium principles. The
interaction between these three players forms a complex game with multi-levels
optimization.
139
In order to deal with the proposed programming models, solution algorithms for
each level of optimization (each player) were presented. Since the defined programming
models for the designer and attacker are discrete optimizations, they are combinatorial
proven to solve these problems efficiently. In this study, a genetic algorithm - which is an
artificial intelligence search metaheuristic – was utilized to solve the attacker and planner
level programming problems. Since the designer in the first framework has two
objectives, it forms a bi-objective problem. The methodology for solving this problem
can identify the best acceptable tradeoffs solutions, based upon multi-objective Pareto
optimization. The last player in the frameworks is the user-level problem, which is a
convex problem based on Nash equilibrium theory. The Frank-Wolfe algorithm was used
to solve the defined programming model for the behavior of the users of the network.
Several test networks were examined to evaluate the performance of the two
models and their algorithms. For the first model, Sioux Falls network, and for the second
model, a 16 links network was used. The results showed that the proposed model could
search over the possible results for the designer and choose an optimal solution for the
bi-objective model for robust network design problem. As an example, in the first model
with attacker budget equal to one, considering the vulnerabilities of the intelligent
adversary in the modeling could increase the total benefits (reduction in the imposed
140
additional TSTT from the intelligent adversary entity) from 20.1% to 46.4%. The highest
influence of the designer’s decision to the overall improvement of the system travel time
after a disruption event was found in low values for adversary’s budget. The most
benefits were gained at attacker’s budget equal to 1 and 2 links, with respectively 28 and
12% reduction in total system travel time after the improvement in a disruption event.
Lower benefits gained at the higher budget for the attacker and the total benefits obtained
at adversary’s budget equal to 7 was 2.7%. In the second model, similar results were
found, with 18.4% improvement in TSTT compares to the base network for the adversary
budget equal to 1 link, and 14.9% for the adversary budget equal to 2 links.
goals for the designer. Other objectives can include but are not limited to reducing
pollution emission, safety and improving accessibility. The goal of the adversary and user
models.
In this research, resources are allocated without considering the time factor of
budget is allocated during multi-year spans. The expected results would be the amount of
budget that should be allocated for each year of the analysis period. In addition, risk
analysis could be performed due to the uncertainties in demand, travel times, discount
rate, and other parameters. Since a series of simulations would be incorporated into this
141
testing the real-world applicability and the computational performance. Studying other
metaheuristics search algorithm for the planner and adversary could make the models
142
References
Abdulaal, M., & LeBlanc, L. J. (1979). Continuous equilibrium network design models.
Transportation Research Part B: Methodological, 13(1), 19–32.
http://doi.org/10.1016/0191-2615(79)90004-3
Adewole, A. P. (2012). A Comparative Study of Simulated Annealing and Genetic
Algorithm for Solving the Travelling Salesman Problem, 4(4), 6–12.
Allende, G. B., & Still, G. (2013). Solving bilevel programs with the KKT-approach.
Mathematical Programming, 138, 309–332. http://doi.org/10.1007/s10107-012-
0535-x
Allsop, R. E. (1974). Some possibilities for using traffic control to influence trip
distribution and route choice. In Transportation and Traffic Theory, Proceedings
(Vol. 6).
Alsnih, R., & Hensher, D. A. (2003). The mobility and accessibility expectations of
seniors in an aging population. Transportation Research Part A: Policy and
Practice, 37(10), 903–916. http://doi.org/10.1016/S0965-8564(03)00073-9
Antoniou, C., Koutsopoulos, H. N., Yannis, G., & Model-based, A. (2007). Traffic state
prediction using Markov chain models, 5000, 2428–2435.
Ayala H., L. F., & Leong, C. Y. (2013). A robust linear-pressure analog for the analysis
of natural gas transportation networks. Journal of Natural Gas Science and
Engineering, 14, 174–184. http://doi.org/10.1016/j.jngse.2013.06.008
Bard, J. F. (1998). Practical bilevel optimization : algorithms and applications .
Nonconvex Optimization and Its Applications ; v. 30, v. 30, xii, 473 p.
http://doi.org/10.1007/978-1-4757-2836-1
Bar-gera, H. (1999). Origin-based algorithms for Transportation Network Modeling,
(103).
Baskan, O. (2013). Determining Optimal Link Capacity Expansions in Road Networks
Using Cuckoo Search Algorithm with Lévy Flights. Journal of Applied
Mathematics, 2013, 1–11. http://doi.org/10.1155/2013/718015
Beimborn, E. A. E. (1995). A transportation modeling primer. Inside the Blackbox:
Making Transportation Models Work for Livable Communities, 2006(May 1995), 1–
21. Retrieved from
http://www.jamescitycountyva.gov/jccplans/2009_comprehensive_plan/pdf/steering
committee/10_14_08models.pdf
Bell, M. G. . (2000). A game theory approach to measuring the performance reliability of
transport networks. Transportation Research Part B: Methodological, 34(6), 533–
545. http://doi.org/10.1016/S0191-2615(99)00042-9
Bell, M. G. H., Kanturska, U., Schmocker, J. D., & Fonzone, A. (2008). Attacker-
143
defender models and road network vulnerability. Philosophical Transactions of the
Royal Society A: Mathematical, Physical and Engineering Sciences, 366(1872),
1893–1906. http://doi.org/10.1098/rsta.2008.0019
Ben-Ayed, O. (1993). Bilevel linear programming. Computers & Operations Research,
20(5), 485–501.
Ben-Ayed, O., Boyce, D. E., & Blair III, C. E. (1988). A general bilevel linear
programming formulation of the network design problem. Transportation Research
Part B: Methodological, 22(4), 311–318.
Berdica, K. (2002). An introduction to road vulnerability: what has been done, is done
and should be done. Transport Policy, 9(2), 117–127. http://doi.org/10.1016/S0967-
070X(02)00011-2
Berdica, K., & Mattsson, L.-G. (2007). Vulnerability: a model-based case study of the
road network in Stockholm. In Critical Infrastructure (pp. 81–106). Springer.
Retrieved from files/3368/chp:10.1007/978-3-540-68056-7_5.pdf
Bhushan, M., Narasimhan, S., & Rengaswamy, R. (2008). Robust sensor network design
for fault diagnosis. Computers and Chemical Engineering, 32, 1067–1084.
http://doi.org/10.1016/j.compchemeng.2007.06.020
Brown, D. B. (1980). The use of dynamic programming in optimally allocating funds for
highway safety improvements. Transportation Planning and Technology, 6(2), 131–
138. http://doi.org/10.1080/03081068008717183
Brown, G., Carlyle, M., Salmerón, J., & Wood, K. (2006). Defending critical
infrastructure. Interfaces, 36(6), 530–544. http://doi.org/10.1287/inte.1060.0252
Bruynooghe, M. (1972). An optimal method of choice of investments in a transport
network. PTRC Proceedings.
Cao, J. X., Li, X. X., Wang, Z. Y., & Wu, J. (2013). The Robust Model of Continuous
Transportation Network Design Problem with Demand and Cost Uncertainties.
Procedia - Social and Behavioral Sciences, 96(Cictp), 970–980.
http://doi.org/10.1016/j.sbspro.2013.08.111
Chamieh, M., & El-kouatly, R. (n.d.). Impact of IDM Lane-Changing on the Performance
of AODV on VANETs, 36–41. http://doi.org/10.5013/IJSSST.a.13.06.06
Chen, A., Kongsomsaksakul, S., Zhou, Z., Lee, M., & Recker, W. W. (2007). Assessing
network vulnerability of degradable transportation systems: an accessibility based
approach.
Chen, A., Yang, C., Kongsomsaksakul, S., & Lee, M. (2007). Network-based
accessibility measures for vulnerability analysis of degradable transportation
networks. Networks and Spatial Economics, 7, 241–256.
Chen, A., Yang, H., Lo, H. K., & Tang, W. H. (2002). Capacity relaibility of a road
144
network: an assessment methodology and numerical results. Transportation
Research Part B: Methodological. http://doi.org/10.1016/S0191-2615(00)00048-5
Chen, A., Zhou, Z., Chootinan, P., Ryu, S., Yang, C., & Wong, S. C. (2011). Transport
Network Design Problem under Uncertainty: A Review and New Developments.
Transport Reviews, 31(6), 743–768. http://doi.org/10.1080/01441647.2011.589539
Chiou, S. (2015). A bi-level decision support system for uncertain network design with
equilibrium fl ow. Decision Support Systems, 69, 50–58.
http://doi.org/10.1016/j.dss.2014.12.004
Chiou, S.-W. (2005a). Bilevel programming for the continuous transport network design
problem. Transportation Research Part B: Methodological, 39(4), 361–383.
http://doi.org/10.1016/j.trb.2004.05.001
Chiou, S.-W. (2005b). Bilevel programming for the continuous transport network design
problem. Transportation Research Part B: Methodological, 39(4), 361–383.
http://doi.org/10.1016/j.trb.2004.05.001
Crucitti, P., Latora, V., & Marchiori, M. (2004). Model for cascading failures in complex
networks. Physical Review E, 69(4), 045104.
http://doi.org/10.1103/PhysRevE.69.045104
de Araújo, D. R. B., Bastos-Filho, C. J. a., & Martins-Filho, J. F. (2015). An evolutionary
approach with surrogate models and network science concepts to design optical
networks. Engineering Applications of Artificial Intelligence, 43, 67–80.
http://doi.org/10.1016/j.engappai.2015.04.004
Dempe, S., & Zemkoho, a. B. (2012). On the KarushKuhnTucker reformulation of the
bilevel optimization problem. Nonlinear Analysis, Theory, Methods and
Applications, 75(3), 1202–1218. http://doi.org/10.1016/j.na.2011.05.097
Dijkstra, A. (2013). Assessing the safety of routes in a regional network. Transportation
Research Part C: Emerging Technologies, 32, 103–115.
http://doi.org/10.1016/j.trc.2012.10.008
Dimitriou, L., & Stathopoulos, A. (2009). Optimal co-evolutionary strategies for the
competitive maritime network design problem. Lecture Notes in Computer Science
(Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics), 5484 LNCS, 818–827. http://doi.org/10.1007/978-3-642-01129-
0_92
Dziubiński, M., & Goyal, S. (2013a). Network design and defence. Games and Economic
Behavior, 79, 30–43. http://doi.org/10.1016/j.geb.2012.12.007
Dziubiński, M., & Goyal, S. (2013b). Network design and defence. Games and Economic
Behavior, 79, 30–43. http://doi.org/10.1016/j.geb.2012.12.007
Erath, A., Birdsall, J., Axhausen, K. W., & Hajdin, R. (2009). Vulnerability assessment
145
methodology for Swiss road network. Transportation Research Record: Journal of
the Transportation Research Board, 2137, 118–126.
Falk, J. E., & Liu, J. (1995). On bilevel programming , Part I : general, 70, 70–71.
Farahani, R. Z., Miandoabchi, E., Szeto, W. Y., & Rashidi, H. (2013). A review of urban
transportation network design problems. European Journal of Operational
Research, 229(2), 281–302. http://doi.org/10.1016/j.ejor.2013.01.001
Feremans, C., Labbé, M., & Laporte, G. (2003). Generalized network design problems.
European Journal of Operational Research, 148(1), 1–13.
http://doi.org/10.1016/S0377-2217(02)00404-6
Flisberg, P., Lidén, B., & Rönnqvist, M. (2009). A hybrid method based on linear
programming and tabu search for routing of logging trucks. Computers &
Operations Research, 36(4), 1122–1144. http://doi.org/10.1016/j.cor.2007.12.012
Fosgerau, M., & Karlström, A. (2010). The value of reliability. Transportation Research
Part B: Methodological, 44(1), 38–49. http://doi.org/10.1016/j.trb.2009.05.002
Friesz, T. L. (1985). Transportation network equilibrium, design and aggregation: Key
developments and research opportunities. Transportation Research Part A: General,
19(5-6), 413–427. http://doi.org/10.1016/0191-2607(85)90041-X
Gallo, M., Acierno, L. D., & Montella, B. (n.d.). Compendium of Papers A meta-
heuristic algorithm for solving the road network design problem in regional
contexts, (1984).
Gallo, M., D’Acierno, L., & Montella, B. (2012). A Meta-heuristic Algorithm for Solving
the Road Network Design Problem in Regional Contexts. Procedia - Social and
Behavioral Sciences, 54, 84–95. http://doi.org/10.1016/j.sbspro.2012.09.728
Gangi, M. Di, Pianificazione, A., & Luongo, A. (2005). © Association for European
Transport and contributors 2005, (1999).
Gao, Z., Wu, J., & Sun, H. (2005). Solution algorithm for the bi-level discrete network
design problem. Transportation Research Part B: Methodological, 39(6), 479–495.
http://doi.org/10.1016/j.trb.2004.06.004
Gershwin, S. B., Tan, H.-N., & others. (1979). Hybrid Optimization: control of traffic
networks in equilibrium.
Gregoriades, A., & Mouskos, K. C. (2013). Black spots identification through a Bayesian
Networks quantification of accident risk index. Transportation Research Part C:
Emerging Technologies, 28, 28–43. http://doi.org/10.1016/j.trc.2012.12.008
Grubesic, T. H., Murray, A. T., & Mefford, J. N. (2007). 10 Continuity in Critical
Network Infrastructures : Accounting for Nodal Disruptions.
Indrei, E. (2006). Markov Chains and Traffic Analysis, 2(1), 1–15.
146
Iyer, S. M. S. M., Nakayama, M. K. M. K., & Gerbessiotis, A. V. a. V. (2009). A
Markovian Dependability Model with Cascading Failures. IEEE Transactions on
Computers, 58(9), 1238–1249. http://doi.org/10.1109/TC.2009.31
Jenelius, E. (2010a). Large-Scale Road Network Vulnerability Analysis, (October).
Jenelius, E. (2010b). User inequity implications of road network vulnerability, 2, 57–73.
Jenelius, E. (2011). Road net vul a grid based.
Jenelius, E., & Mattsson, L. G. (2012). Road network vulnerability analysis of area-
covering disruptions: A grid-based approach with case study. Transportation
Research Part a-Policy and Practice, 46(5), 746–760.
http://doi.org/10.1016/j.tra.2012.02.003
Jenelius, E., & Mattsson, L.-G. (2006). Developing a methodology for road network
vulnerability analysis. Nectar Cluster, 1, 1–9. Retrieved from files/3349/Jenelius
and Mattsson - 2006 - Developing a methodology for road network vulnerab.pdf
Jenelius, E., & Mattsson, L.-G. (2012). Road network vulnerability analysis of area-
covering disruptions: A grid-based approach with case study. Transportation
Research Part A: Policy and Practice, 46(5), 746–760.
http://doi.org/10.1016/j.tra.2012.02.003
Jenelius, E., Petersen, T., & Mattsson, L.-G. (2006). Importance and exposure in road
network vulnerability analysis. Transportation Research Part A: Policy and
Practice, 40(7), 537–560. http://doi.org/10.1016/j.tra.2005.11.003
Jin, X. Z., & Yang, G. H. (2014). Robust synchronization control for complex networks
with disturbed sampling couplings. Communications in Nonlinear Science and
Numerical Simulation, 19(6), 1985–1995.
http://doi.org/10.1016/j.cnsns.2013.10.030
Kano, M., Shiraishi, T., & Kuwahara, M. (2007). Driver Behavior Model for ITS
Evaluation System, 5(1), 17–25.
Kaplan, S., & Garrick, B. J. (1981). On The Quantitative Definition of Risk. Risk
Analysis, 1(1), 11–27. http://doi.org/10.1111/j.1539-6924.1981.tb01350.x
Karami, A., & Guerrero-zapata, M. (2015). Neurocomputing A hybrid multiobjective
RBF-PSO method for mitigating DoS attacks in Named Data Networking.
Neurocomputing, 151, 1262–1282. http://doi.org/10.1016/j.neucom.2014.11.003
Konur, D., Golias, M. M., & Darks, B. (2013). A mathematical modeling approach to
resource allocation for railroad-highway crossing safety upgrades. Accident;
Analysis and Prevention, 51, 192–201. http://doi.org/10.1016/j.aap.2012.11.011
Kröger, W., & Zio, E. (2011). Vulnerable Systems. London: Springer London.
http://doi.org/10.1007/978-0-85729-655-9
147
Latora, V., & Marchiori, M. (2001). Efficient Behavior of Small-World Networks.
Physical Review Letters, 87(19), 198701.
http://doi.org/10.1103/PhysRevLett.87.198701
Latora, V., & Marchiori, M. (2004). Vulnerability and Protection of Critical
Infrastructures, 11–14. Retrieved from http://arxiv.org/abs/cond-mat/0407491v1
Leblanc, L. J. (1973). Mathematical programming algorithms for large scale network
equilibrium and network design problems.
Leblanc, L. J. (2014). An Algorithm for the Discrete Network Design Problem, 9(3),
183–199.
Li, C., Yang, H., Zhu, D., & Meng, Q. (2012). A global optimization method for
continuous network design problems. Transportation Research Part B:
Methodological, 46(9), 1144–1158. http://doi.org/10.1016/j.trb.2012.05.003
Li, M. (2006). Combining DTA Approaches for Studying Road Network Robustness.
Civil Engineering, (November).
Luathep, P., Sumalee, A., Lam, W. H. K. K., Li, Z.-C., & Lo, H. K. (2011). Global
optimization method for mixed transportation network design problem: A mixed-
integer linear programming approach. Transportation Research Part B:
Methodological, 45(5), 808–827. http://doi.org/10.1016/j.trb.2011.02.002
Luo, Z.-Q., Pang, J.-S., & Ralph, D. (1996). Mathematical programs with equilibrium
constraints. Cambridge University Press.
Mansouri, M., & Mostashari, A. (2009). A Decision Analysis Framework for Resilience
Strategies in Maritime Systems.
Marcotte, P. (1986). Network design problem with congestion effects: A case of bilevel
programming. Mathematical Programming, 34(2), 142–162.
Marcotte, P., & Marquis, G. (1992). Efficient implementation of heuristics for the
continuous network design problem. Annals of Operations Research, 34(1), 163–
176.
Martin, P. A. S., & Thesis. (2007). TRI-LEVEL OPTIMIZATION MODELS TO
DEFEND CRITICAL INFRASTRUCTURE, (September).
Mathew, T. V., & Sharma, S. (2006). Continuous Network Design with Emission Pricing
as a Bi-Level Optimization Problem. Applications of Advanced Technology in
Transportation, 804–809. http://doi.org/10.1061/40799(213)129
Melachrinoudis, E., & Kozanidis, G. (2002). A mixed integer knapsack model for
allocating funds to highway safety improvements. Transportation Research Part A:
Policy and Practice, 36(9), 789–803. http://doi.org/10.1016/S0965-8564(01)00040-
4
148
Meng, Q., & Yang, H. (2002). Bene ® t distribution and equity in road network design,
36, 19–35.
Miandoabchi, E., Daneshzand, F., Szeto, W. Y., & Zanjirani Farahani, R. (2013). Multi-
objective discrete urban road network design. Computers & Operations Research,
40(10), 2429–2449. http://doi.org/10.1016/j.cor.2013.03.016
Michon, J. A. (1985). A CRITICAL VIEW OF DRIVER BEHAVIOR MODELS :
WHAT DO WE KNOW , WHAT SHOULD WE DO ?, 485–520.
Migdalas, A. (1998). Multilevel Optimization: Algorithms and Applications. (A.
Migdalas, P. M. Pardalos, & P. Värbrand, Eds.) (Vol. 20). Boston, MA: Springer
US. http://doi.org/10.1007/978-1-4613-0307-7
Moridpour, S., Sarvi, M., & Rose, G. (2006). Modelling Lane Changing Behaviour of
Heavy Commercial Vehicles, 1–15.
Mouskos, K. C. (1991). A Tabu-based Heuristic Search Strategy to Solve a Discrete
Transportation Equilibrium Network Design Problem. University of Texas at
Austin. Retrieved from https://books.google.com/books?id=UVhlnQEACAAJ
Murray, A. T., Davis, R., Stimson, R. J., & Ferreira, L. (1998). Public Transportation
Access. Transportation Research Part D: Transport and Environment, 3(5), 319–
328. http://doi.org/10.1016/S1361-9209(98)00010-8
Murray, A. T., & Grubesic, T. H. (2007). Critical Infrastructure. (A. T. Murray & T. H.
Grubesic, Eds.). Berlin, Heidelberg: Springer Berlin Heidelberg.
http://doi.org/10.1007/978-3-540-68056-7
Murray, A. T., Matisziw, T. C., & Grubesic, T. H. (2007). Critical network infrastructure
analysis: interdiction and system flow. Journal of Geographical Systems, 9(2), 103–
117. http://doi.org/10.1007/s10109-006-0039-4
Murray-Tuite, P., & Mahmassani, H. (2004). Methodology for Determining Vulnerable
Links in a Transportation Network. Transportation Research Record, 1882(1), 88–
96. http://doi.org/10.3141/1882-11
Nicholson, A., Schmöcker, J. D., Bell, M. G. H., & Iida, Y. (2003). Assessing transport
reliability: malevolence and user knowledge.
Numrich, S., & Tolk, A. (2010). Challenges for Human, Social, Cultural, and Behavioral
Modeling. SCS M&S Magazine, 01, 1–9. Retrieved from
http://www.scs.org/magazines/2010-01/index_file/Files/article_Numrich-Tolk.pdf
Ochoa-Rosso, F., & A, S. (1969). Optimum project addition in urban transportation
networks via descriptive traffic assignment models.
Ouyang, M., Dueñas-osorio, L., & Min, X. (2012). A three-stage resilience analysis
framework for urban infrastructure systems. Structural Safety, 36-37, 23–31.
http://doi.org/10.1016/j.strusafe.2011.12.004
149
Parvaresh, F., Husseini, S. M. M., Golpayegany, S. a H., & Karimi, B. (2014). Hub
network design problem in the presence of disruptions. Journal of Intelligent
Manufacturing, 25, 755–774. http://doi.org/10.1007/s10845-012-0717-7
Patriksson, M., & Rockafellar, R. T. (2002). A Mathematical Model and Descent
Algorithm for Bilevel Traffic Management. Transportation Science, 36(3), 271–291.
http://doi.org/10.1287/trsc.36.3.271.7826
Pei, Y. L., Fischer, J. M., & Amekudzi, A. A. (2010). Performance Measurement in State
Departments of Transportation: A Literature Review and Survey of Current Practice.
Poorzahedy, H., & Rouhani, O. M. (2007). Hybrid meta-heuristic algorithms for solving
network design problem. European Journal of Operational Research, 182, 578–596.
http://doi.org/10.1016/j.ejor.2006.07.038
Poorzahedy, H., & Turnquist, M. A. (1982). Approximate algorithms for the discrete
network design problem. Transportation Research Part B: Methodological, 16(1),
45–55.
Qiang, Q., & Nagurney, A. (2007). A unified network performance measure with
importance identification and the ranking of network components. Optimization
Letters, 2(1), 127–142. http://doi.org/10.1007/s11590-007-0049-2
Reniers, G. L. L. L. L., & Dullaert, W. (2013). A method to assess multi-modal Hazmat
transport security vulnerabilities: Hazmat transport SVA. Transport Policy, 28, 103–
113. http://doi.org/10.1016/j.tranpol.2012.05.002
Scott, D. M., Novak, D. C., Aultman-Hall, L., & Guo, F. (2006). Network Robustness
Index: A new method for identifying critical links and evaluating the performance of
transportation networks. Journal of Transport Geography, 14(3), 215–227.
http://doi.org/10.1016/j.jtrangeo.2005.10.003
Scott, D. M., Novak, D., & Guo, F. (2005). Network Robustness Index : A New Method
for Identifying Critical Links and Evaluating the Performance of Transportation
Networks, (April).
Sharma, S., & Mathew, T. V. (2011). Multiobjective network design for emission and
travel-time trade-off for a sustainable large urban transportation network.
Environment and Planning B: Planning and Design, 38(3), 520–538.
http://doi.org/10.1068/b37018
Sheffi, Y. (1985). Urban transportation networks.
Snelder, M. (2010). Designing Robust Road Networks.
Steenbrink, P. A. (1974). Optimization of transport networks. New York.
Sullivan, J. L., Aultman-hall, L., & Novak, D. C. (2009). A review of current practice in
network disruption analysis and an assessment of the ability to account for isolating
links in transportation networks. Transportation Letters: The International Journal
150
of Transportation Research, 1(4), 271–280.
http://doi.org/10.3328/TL.2009.01.04.271-280
Suwansirikul, C., Friesz, T. L., & Tobin, R. L. (1987). Equilibrium Decomposed
Optimization: A Heuristic for the Continuous Equilibrium Network Design Problem.
Transportation Science, 21(4), 254–263. http://doi.org/10.1287/trsc.21.4.254
Suzuki, T., & Mori, M. (n.d.). CALIBRATION AND EVALUATION OF LANE-
CHANGE MODEL ON HIGHWAY FROM MICROSCOPIC & MACROSCOPIC
VIEWPOINTS, 1–12.
Tampère, C. M. J., Stada, J., Immers, B., & Leuven, K. U. (n.d.). Methodology for
Identifying Vulnerable Sections in a National Road Network, 1–18.
Taylor, M. A. P., & D’Este, G. M. (2007). Transport network vulnerability: a method for
diagnosis of critical locations in transport infrastructure systems. Springer.
Retrieved from files/3398/chp:10.1007/978-3-540-68056-7_2(1).pdf
Tsoukalas, A., & Wiesemann, W. (n.d.). Global Optimisation of Pessimistic Bi-Level
Problems, 1–31.
Ünlü, G. (1987). A linear bilevel programming algorithm based on bicriteria
programming. Computers & Operations Research, 14(2), 173–179.
Vicente, L. N., & Calamai, P. H. (1994a). Bilevel and multilevel programming: A
bibliography review. Journal of Global Optimization, 5(3), 291–306.
http://doi.org/10.1007/BF01096458
Vicente, L. N., & Calamai, P. H. (1994b). Bilevel and multilevel programming: A
bibliography review. Journal of Global Optimization, 5(3), 291–306.
http://doi.org/10.1007/BF01096458
Wan, Z., Mao, L., & Wang, G. (2014). Estimation of distribution algorithm for a class of
nonlinear bilevel programming problems. Information Sciences, 256, 184–196.
http://doi.org/10.1016/j.ins.2013.09.021
Wang, D. Z. W., & Lo, H. K. (2010). Global optimum of the linearized network design
problem with equilibrium flows. Transportation Research Part B: Methodological,
44(4), 482–492. http://doi.org/10.1016/j.trb.2009.10.003
Wang, S., Meng, Q., & Yang, H. (2013). Global optimization methods for the discrete
network design problem. Transportation Research Part B: Methodological, 50, 42–
60. http://doi.org/10.1016/j.trb.2013.01.006
Washington, S. E. (2008). Traffic Analysis Toolbox Volume VIII: Work Zone Modeling
and Simulation— A Guide for Decision-Makers, VIII(August).
Wheeler, T. A., & Llc, R. B. (n.d.). A Probabilistic Framework for Microscopic Traffic
Propagation.
151
Wu, J., Guo, X., Sun, H., & Wang, B. (2014). Topological Effects and Performance
Optimization in Transportation Continuous Network Design. Mathematical
Problems in Engineering, 2014, 1–7. http://doi.org/10.1155/2014/490483
Xu, M., & Gao, Z. (2009). Chaos in a dynamic model of urban transportation network
flow based on user equilibrium states. Chaos, Solitons & Fractals, 39(2), 586–598.
http://doi.org/10.1016/j.chaos.2007.01.077
Yang, H. A. (1997). SENSITIVITY ANALYSIS FOR THE ELASTIC-DEMAND, 31(I),
55–70.
Yang, H., & H. Bell, M. G. (1998). Models and algorithms for road network design: a
review and some new developments. Transport Reviews, 18(3), 257–278.
http://doi.org/10.1080/01441649808717016
Yilmaz, L. (2009). Agent-Directed Simulation and Systems Engineering. Systems
Engineering, 897–904. http://doi.org/10.1002/9783527627783
Yim, K. K. W., Wong, S. C., Chen, A., Wong, C. K., & Lam, W. H. K. (2011). A
reliability-based land use and transportation optimization model. Transportation
Research Part C: Emerging Technologies, 19(2), 351–362.
http://doi.org/10.1016/j.trc.2010.05.019
Yun-peng, W., Yu-qin, F., Wen-xiang, L. I., Fulcher, W. C., & Zhang, L. (n.d.). Road
network vulnerability analysis based on improved ant colony algorithm, 1–11.
Zeng, Q. (1998). Zeng , Qifeng 1998 A combined simulated annealing and TABU search
strategy to solve network design problem with two classes of users.
Zhang, J., Xu, X., Hong, L., Wang, S., & Fei, Q. (2012). Attack vulnerability of self-
organizing networks. Safety Science, 50(3), 443–447.
http://doi.org/10.1016/j.ssci.2011.10.005
152